AI Coding Tools Brief · Oct 9–22, 2026

AI Coding Tools Brief, Oct 9–22, 2026: Model Churn, New Meters and the Review Bottleneck

Contents

1. Key takeaways from the two weeks
2. GitHub Copilot and VS Code
3. Cursor
4. Terminal and CLI agents — Claude Code, Codex, OpenCode, Qwen Code
5. Google, Cognition (Devin / ex-Windsurf), AWS Kiro and JetBrains
6. Models and benchmarks behind the tools
7. Pricing and plans
8. Enterprise adoption, security and market moves
9. Selection advice and what to watch next

1. Key takeaways from the two weeks

Between October 9 and October 22, 2026 no major coding tool changed its headline price, yet almost every one changed what a subscription actually buys: which models it can call, which defaults it runs with, and how usage is metered. Five conclusions follow from Chapters 2–8.

1.1 Five conclusions

1. Model churn is now a monthly operating cost. GitHub retired six Copilot models on October 19 and scheduled three more for November 16; counting the four retired on October 2, that is 13 models in 45 days. GPT-5.5 left Codex for ChatGPT sign-ins on October 14, and Kiro ended Claude Sonnet 4 the same day. Even the suggested replacements moved: GitHub's September notice and its October 19 post name different successor models. Any workflow that hard-codes a model name will break again within weeks.

2. Defaults changed underneath unattended automation. Claude Code 2.1.302 (October 16) starts unflagged claude -p and Agent SDK runs in Manual permission mode in several setups, and turns fork subagents on by default. Codex CLI 0.162 made new threads follow the server's model defaults. None of these changes touches a repository, so CI jobs that never pinned model, effort and permission mode can now behave differently.

3. The big platforms are consolidating and taking sides. Cursor's new buyers now get The X Subscription, where the $8 Plus tier runs Grok models only and Claude or GPT usage is billed on top. OpenAI has proposed cutting off Cursor's access to its models on November 12. Google stopped selling new Gemini Code Assist Standard and Enterprise subscriptions on October 9 and points buyers to Antigravity. Meanwhile hosts are opening up: JetBrains Air runs Codex, Claude Agent, Copilot and Junie side by side, and Cursor's Origin now has an app marketplace.

4. Claude leads the hard benchmark; OpenAI leads on cost per task. On Terminal-Bench 4.0, Claude Opus 5.5 scores 65.15% on Vals AI, ahead of Sonnet 5.5 and GPT-6 Astra. But GPT-6.1 Sol reaches 85–94% of the leader's score at roughly one-eighth to one-fourteenth of the cost per task across four independent boards. SWE-bench Verified has been archived as saturated, and bug-rate data ranks models differently from pass-rate boards.

5. Agents raise output; review decides whether it ships. A Harvard working paper covering 718 firms found code volume up about 30% after adopting coding agents, while resolved issues did not change significantly and PR-to-merge time rose 49%. Money is following the bottleneck: CI provider Blacksmith raised $66 million at a $1 billion valuation. For an agent-heavy 10-person team the realistic bill is roughly $400–1,000 a month, two to five times the seat price (estimate, Chapter 7).

The fortnight's lesson: treat model, effort and permission settings as code, budget per task rather than per seat, and invest in review before buying more generation.

1.2 What changed at a glance

2. GitHub Copilot and VS Code

Chapter 1 named model churn as the fortnight's defining theme; GitHub Copilot is where it is most visible. The thesis here: Copilot's model list now changes faster than most teams review their settings, so the durable value GitHub shipped was not any single model but the policy and harness layer around them.

2.1 Release timeline

GitHub and Microsoft shipped something almost daily; items just before Oct 9 are marked as background.

Three more entries, "HydraFusion updates and new Max option", agentic autofix GA and managed-settings enforcement updates, appear between the Oct 10 and Oct 19 posts only as titles; autofix is covered in Chapter 8.

2.2 Model line-up: in vs out

Two models came in. Claude Haiku 5.5 reached GA on Oct 7 for Pro, Pro+, Max, Business and Enterprise and is aimed at subagents, quick edits and terminal tasks. Gemini 4 Argon appears as a Release entry in GitHub's changelog; according to shattered.io it went GA on Oct 13 for Pro+ and above, rolling out gradually. Its benchmark standing is covered in Chapter 6.

Far more went out:

13 models retired or scheduled for retirement in 45 days: 4 on Oct 2 + 6 on Oct 19 + 3 on Nov 16 = 13.

The replacement targets themselves moved. GitHub's Sep 18 notice pointed the Oct 19 retirees to GPT-5.6 Sol, GPT-5.6 Luna and Grok 4.6; the Oct 19 retirement post names GPT-6.1 Sol, GPT-6 Luna and Grok 4.7 instead. This chapter follows the later primary post; shattered.io repeated the older mapping and claimed Grok 4.5 had no successor, which GitHub's Oct 19 table contradicts. Meanwhile GPT-5.6 Terra, a sibling of the September targets, is itself on the Nov 16 list. Hard-coding the "suggested alternative" only delays the next migration.

2.3 Agent, CLI and VS Code features

VS Code is becoming an agent workspace. Version 1.141 introduced the Copilot harness, built on the Copilot SDK and running in a dedicated agent-host process, so several windows can attach to one session. The same release added terminal sandboxing on Windows, macOS and Linux and a grid for comparing sessions side by side. Version 1.142 is making the Copilot harness the default for new sessions, rolling out progressively, and brings back Ask as a read-only agent. It also adds Canvases (Preview), shared interactive surfaces such as a Kanban board that both user and agent can edit, and lets admins enforce agent sandboxing and limit which permission modes developers can use. Insiders builds in the window added ${VAR} expansion in .mcp.json and separate OAuth tokens for each MCP server.

The CLI is leaning local. Copilot CLI 1.0.94-0 can discover models from a running Ollama instance. Choosing a local model does not turn on offline mode or disable telemetry; offline mode still requires COPILOT_OFFLINE=true. Microsoft says Copilot will route between on-device and cloud inference "by the end of the month" and has built a local MAI Code 1.1 Flash (137B total, 6.8B active parameters) for NVIDIA RTX Spark PCs (Announced, not yet shipped as of Oct 23).

Computer use predates the window. GitHub's changelog index lists the computer-use post before the Sep 28 weekly release. Tech-insider.org (Oct 11) called it a public preview for Copilot CLI and the Copilot app. Treat it as Preview since late September, not an October launch.

2.4 What admins must do before Nov 16

As of Oct 23 the Oct 19 retirements are already in effect, so the work is now preparing for Nov 16:

1. Check the global default. If default model enablement is left on, Copilot Business and Enterprise get Claude Opus 5.5, GPT-6.1 Sol and GPT-6 Luna automatically. If an admin turned the global default off, they must enable these models one by one. Under the global model policy (GA since Aug 26), open-weight models and models outside GitHub's data-retention agreement never get default enablement.
2. Script the audit. Use the new GET/PATCH model-policy endpoints for enterprises, enterprise teams, enterprise apps and organizations (GA Oct 19) instead of clicking through settings.
3. Purge hard-coded model names from workflows and integrations. GitHub asks users to switch to supported models before the date.
4. Set defaults centrally. JetBrains managed settings can now pick the default agent model, and VS Code 1.142 enforces sandbox and permission policies.
5. Mind the exception. Individual subscribers on annual plans keep Claude Opus 4.8.

The judgement: Copilot's broad model menu is its selling point, but it now brings a migration every few weeks, so enterprises should manage model policy as code through the API rather than as a one-off setting.

3. Cursor

Chapter 2 showed GitHub widening Copilot's model menu; in the same fortnight Cursor went the other way. It became a vertically integrated SpaceXSI product: the cloud-agent stack got stronger, but non-Grok models now cost more and carry more risk.

3.1 Release timeline

Cursor's changelog pages are undated, so the dates below come from secondary coverage and Cursor's own forum.

Builds are harder to pin down: a forum user notes the Oct 16 notes "roughly map to release 3.24.x", while 3.25 and 3.26 shipped without notes, and Cursor staff confirmed a tracked bug where macOS 27 updates never finish installing.

3.2 Cloud-agent platform direction

Each release only makes sense as one piece of a larger stack. Origin, Cursor's code host, now has an app marketplace covering deployment, CI, code review, security, monitoring and project management, with partners including Cloudflare, Linear, Sentry and Socket Security. Third-party apps authenticate with an app JWT and installation tokens, receive webhooks, and write back check runs, PR comments and reviews. That is the same integration model as a GitHub App.

The Environments API (beta) lets you create, list, update and delete the cloud environments agents run in, and view their change and build history. An environment can be personal or shared with the team. Canvases give agents an output format other than diffs: you share the finished artifact by link. Cursor's Rollouts and Security Review bots, which predate the window (CatDoes dates Security Review to September 2026), already accept Origin as a source-control backend alongside GitHub.

Cursor is building a GitHub-shaped platform: code hosting, an app marketplace, API-managed agent environments and review bots, with the editor as only one entry point.

The only adoption data point comes from shattered.io, citing a Cursor case study: about 200 engineers at Navan use Cloud Agents; treat it as a vendor anecdote. The forum shows the friction that comes with running a hosted fleet. One user reports that two Cloud Agent sessions, along with weeks of commits, vanished on Oct 17. A thread saying all non-HTTPS outbound traffic from Grok Bot computers has been blocked "since Oct 1" had 50 replies and 1,511 views by Oct 22. Until retention guarantees are documented, push cloud-agent work to a remote you control.

3.3 Model access: the X Subscription and the OpenAI cut-off

The bundle. SpaceXSI announced The X Subscription on X at 22:48 UTC on Oct 12, 2026. CellCog notes the x.ai post itself carries an Oct 2 date; this chapter follows the Oct 12 public announcement. Cursor's pricing page now sells these tiers instead of its old Pro and Ultra plans. Full cross-tool pricing is in Chapter 7.

The usage pool resets weekly and is shared across Grok and Cursor. Third-party models (Claude, GPT) draw on prepaid credit from the first request instead of an included allowance. Existing subscribers keep their plans and the old plans are closed to new sign-ups. SpaceXSI says it will "eventually" migrate everyone but has given "no timeline yet". Vendor claims: Premium bundles benefits "previously $90/month combined", and each tier includes "at least twice as much coding usage with our models".

To see what metering from the first request means, take Claude Opus 5.5. Cursor lists it at $4 input / $20 output per million tokens. Fast mode costs exactly double, and on legacy request-based plans it requires Max Mode (rates checked Oct 17). An illustrative session of 2M input and 0.2M output tokens costs 2 × $4 + 0.2 × $20 = $12 standard, or $24 in Fast mode, on top of the subscription.

The OpenAI cut-off. OpenAI wrote that it notified SpaceX it intends to wind down the contract, "with a proposed shutoff date of November 12, 2026". It cited distrust of Musk-owned companies' compliance with its terms. Status: Announced / upcoming (OpenAI, proposed effective Nov 12, 2026), which is 20 days after this report's run date. Cursor CEO Michael Truell said OpenAI models serve "about 5% of Cursor user traffic" and that Cursor is talking with OpenAI to resolve it (via CNBC). According to shattered.io, which cites HackerNoon, personal OpenAI API keys will keep working in local Chat and Agent after the cut-off. That is not confirmed by a primary source. CatDoes reports that Cursor's OpenAI line-up already stops at GPT-5.6, with no GPT-6 models.

What this means for you: if you are on a legacy plan and use Claude or GPT heavily, staying put is the cheaper option for now. If you depend on OpenAI models, re-validate prompts and agent configs on a replacement model before Nov 12; analysts warn this is non-trivial. Rivals are recruiting: Kilo offers Cursor Enterprise customers free Kilo Enterprise for up to 12 months if they apply by Nov 30, 2026.

3.4 CursorBench claims vs independent evidence

CursorBench is Cursor's first-party benchmark. Its tasks are private and drawn from real Cursor sessions. The 4.0 set was published Sept 10, 2026, and its scores are not comparable with the retired 3.2 set. BenchLM's Sept 24 snapshot of 4.0 lists Claude Opus 5.5 at 57.8%, Claude Fable 5.1 at 51.8%, Grok 4.7 at 46.3% and GPT-5.6 Sol at 41.7%. BenchLM shows the table for reference only and does not use it to rank models. This matches Cursor's claim that Opus 5.5 set a new CursorBench high, ahead of Fable 5.1. Meanwhile, llm-stats still shows a v3.2 board, updated Oct 18, with only two models (Fable 5.1 0.734, Grok 4.6 0.699). Two "CursorBench leaderboards" with different leaders are circulating.

According to tech-insider.org, Cursor added Gemini 4 Argon on Oct 13 and posted 52.8% "at standard inference volume", calling it comparable to Opus 5.5. No primary source in the material confirms this. BenchLM records Opus 5.5 at its highest published effort tier, so the two figures do not measure the same setup.

Independent evidence is thin and anecdotal. Kingy AI ran one cart-repair task in Cursor 3.23.23 on Oct 12. Grok 4.7 High and Claude Sonnet 5.5 High both passed 24/24 checks, but Grok took 9 min 26 s against 33 s (566 ÷ 33 ≈ 17x) and 357,280 tokens against 119,083 (≈ 3x). Our judgement: CursorBench is a useful vendor signal for ranking models inside Cursor's own agent loop, not neutral evidence, and it ignores token cost, which now falls on the user for every third-party model. Cross-tool benchmarks are in Chapter 6.

4. Terminal and CLI agents — Claude Code, Codex, OpenCode, Qwen Code

Where Chapter 3's Cursor moves were about packaging, the terminal agents spent the fortnight changing something riskier for automation: their defaults.

Three default changes within about a week (Codex's server-side model default on Oct 8, GPT-5.5 leaving Codex for ChatGPT sign-ins on Oct 14, Claude Code's headless permission mode on Oct 16) changed what unpinned scripts do without a single line changing in the repository.

4.1 Claude Code: from 2.1.296 to 2.1.305, headless defaults flip and Code Review goes GA

CLI releases. According to Tech Insider, Anthropic's Claude Code 2.1.296 (Oct 9) added CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL for pinning workflow subagents to a model. ClaudeWorld, which checks versions against the official CHANGELOG, puts the line at 2.1.305 by Oct 20. The release that matters is 2.1.302, tagged Oct 16. Tech Insider dates it Oct 17, citing Releasebot; we follow the tag date. Two lines quoted verbatim from its release notes change unattended runs:

• claude -p and Agent SDK sessions with no permission mode now start in default (Manual) mode on third-party providers, with telemetry off, or on a fresh install. --permission-mode auto keeps auto. A -p run has nobody to approve prompts, so a CI job that leaned on the old default can return denied edits instead of a diff.
• Fork subagents, which inherit the whole conversation, are now on by default in claude -p, the Agent SDK, the desktop app and VS Code. CLAUDE_CODE_FORK_SUBAGENT=0 turns them off.

The same build adds CLAUDE_CODE_AUTO_MODE_MODEL, an Idle compaction toggle in /config, a limit of 12 concurrent hooks per event, and a refusal to read image sources over 128 MB. It also warns that standalone Claude Design at claude.ai/design closes by Dec 14, 2026 (announced).

Platform: Code Review GA. Anthropic made Code Review generally available on Oct 15, 2026 for Team and Enterprise. Most reviews now finish about 3x faster than in the preview (about 20 minutes down to 6–10 minutes), and admins choose Quick, Standard (the default) or Deep for each repository. Reviews are billed separately through usage credits at an average of $15–25 each (prices in Chapter 7), and the service is unavailable with zero data retention. The Help Center still says 20 minutes on average; we follow the newer GA post. Eligible Enterprise organisations can claim at least $10,000 in Code Review credits until Jan 15, 2027 (time-limited; credits expire 30 days after they are claimed). The local /code-review command works on any plan and counts toward plan usage. The managed service reads REVIEW.md; the local command does not.

Plugins: treat with care. Background: since Sep 25, developers can submit plugins to the Claude directory via a developer portal. Tech Insider reports that a curated claude-plugins-official repository and a JavaScript/TypeScript "Mods system" arrived with 2.1.302. However, the 2.1.302 lines quoted by DEV Community and AI Weekly do not include them, and ClaudeWorld places Claude Mods in 2.1.287, before this window. "Plugins went official at 2.1.302" is unverified.

4.2 OpenAI Codex: CLI 0.163 versus the GPT-5.5 removal

CLI releases. Background: GPT-6.1 Sol became the default in Codex's bundled model catalog with 0.159.1 (Sep 29). Codex CLI 0.162.0 (Oct 8) then made new TUI threads follow the app server's model and reasoning defaults unless the launch overrides them. 0.163.0 (Oct 15) has 14 changelog entries:

• a follow-up sent during a response steers the agent at once, and an interrupt also stops active Code Mode cells;
• independent read-only tools (skill, memory and history lookups) run in parallel;
• /config can be searched and marks settings that need a restart; there is a Ctrl+X leader key, and voice mute moves to F9;
• /daemon restarts the background server, and the bundled patched zsh is gone.

Two security fixes ship with it: resumed threads restore their permissions and network approvals, and the network proxy now checks permissions before DNS lookups and rejects duplicate Host headers.

Platform and model. GPT-5.5 was retired from ChatGPT, ChatGPT Work and Codex for ChatGPT sign-ins on Oct 14, 2026, on every plan, at noon PDT (19:00 UTC). The removal was announced on Sep 14. The API and Codex sessions signed in with an API key are not affected. The sources disagree on the replacement:

• DevelopersDigest (updated Oct 16) and DEV (Oct 18), both citing OpenAI's Codex models page, name GPT-6.1 Sol for Plus to Edu plans and GPT-6 Luna for Free and Go;
• LaoZhang names gpt-6-sol as the like-for-like swap;
• Layer3Labs names gpt-5.6-sol.

We follow the models-page reading, but no OpenAI page is in our material, so check your own picker. On Enterprise and Edu, GPT-6.1 Sol stays off until an admin enables it. This is the third retirement for ChatGPT sign-ins this season, after gpt-5.4 (Aug 31) and gpt-5.3-codex-spark (Sep 14). An open user report (#54226, Oct 15) says long GPT-5.5 threads fail at compaction even after switching models.

4.3 OpenCode 2.0, and a gap on Qwen Code

Releases. OpenCode 1.19.0 (Oct 11), the open-source agent from Anomaly, added opencode upgrade --major and stopped installing updates automatically. Shattered and Tech Insider date 2.0 to Oct 12. AI IDE List dates the official announcement Oct 13 and saw the changelog at 2.0.29 on Oct 14.

What changed. A shared local background service now owns sessions, config, permissions and tool execution, and terminal, web and desktop clients all attach to it. Other changes:

• new @opencode/client and embeddable @opencode/sdk packages;
• Code Mode, in which the model writes JavaScript that chains tool calls (on by default for MCP servers);
• background subagents and shell commands;
• more than 80 internal features rebuilt as hot-reloadable plugins.

The migration guide lists three intentional breaks: v1 plugins do not run, the server API is new, and tui.json becomes a global cli.json. V2 also runs no language servers and discovers only AGENTS.md, with no CLAUDE.md fallback. OpenCode ships no default model; users connect a hosted provider or a local one through Ollama or LM Studio.

Qwen Code. None of this chapter's sources document Qwen Code 0.25, so we make no claims about it.

4.4 Side by side

Judgement. For teams, Claude Code is now the most governable: explicit permission modes plus a managed review service with per-repository levels. Codex shows the most model churn, with three retirements for ChatGPT sign-ins since Aug 31. OpenCode 2.0 is the most extensible, but upgrading now is costly for setups that depend on plugins. In practice:

• pin the model, effort and permission mode in every CI job, and log them with each result;
• grep repositories and configs for gpt-5.5;
• delay the OpenCode 2 upgrade until v1 plugins are ported.

5. Google, Cognition (Devin / ex-Windsurf), AWS Kiro and JetBrains

Where Chapter 4 followed the terminal agents' release trains, these four vendors spent the fortnight consolidating: closing legacy products, moving customers to successors and re-pricing models.

The action items here are dates, not features: Google's January 31, 2027 Code Assist licence-add cutoff, AWS's April 30, 2027 Amazon Q Developer end of support, and Kiro's October 14, 2026 Claude Sonnet 4 end of life, already in effect.

5.1 Google: Gemini Code Assist end of sale, Antigravity as successor

What changed. On October 9, 2026 Google stopped selling new Gemini Code Assist Standard and Enterprise subscriptions; its release note that day reads "Sale of Gemini Code Assist subscriptions is ending" (Google, via AI Weekly). Existing subscriptions "remain supported until the end of your contractual agreement", and Google gives no shutdown date for the service itself.

Background. On June 18, 2026 Gemini CLI and the Code Assist IDE extensions stopped serving free individuals and Google AI Pro/Ultra subscribers, who were pointed to Antigravity CLI; Standard/Enterprise licence holders keep Gemini CLI. From September 17, Gemini Enterprise Standard and Plus bought online no longer included Code Assist features.

The successor. Google points buyers to Antigravity, "available through eligible Gemini Enterprise subscriptions and through Gemini Enterprise Agent Platform". That is a new subscription, not a seat swap: it requires a paid or trial Gemini Enterprise Standard or Plus edition (or Emerging Market or pay-as-you-go editions) on an invoiced Cloud Billing account. Google offers no price comparison and does not say whether code-customisation settings transfer. According to HighCircl, Google's overview lists SOC 2, ISO 27001, FedRAMP and other certifications as not supported for Antigravity in Gemini Enterprise, and an October 8 release note lets admins enable Claude Opus 5.5 and Sonnet 5.5 in Antigravity, off by default and billed on consumption.

Jules shows no status change in the window: it is still beta, with 15 free tasks a day, 100 on Google AI Pro and 300 on Ultra. Gemini 4 Argon's GA is covered in Chapter 6.

What to do. Buy seats for any 2027 hires before January 31, 2027; confirm compliance requirements against Antigravity before committing; and port Gemini CLI automation using Google's Antigravity CLI migration guide.

5.2 Cognition: Devin Desktop (ex-Windsurf)

Background. Windsurf became Devin Desktop on June 2, 2026 via an over-the-air update. The usage-unit change is older: effective March 19, 2026, Cognition replaced prompt credits with daily and weekly token quotas, raised Pro from $15 to $20, launched Max at $200 and priced Teams at $40 per full seat ($80 monthly minimum), grandfathering existing $15 Pro and $30 Teams subscribers. Enterprise contracts bill in Agent Compute Units at a published 12.5 legacy prompt credits per ACU; a 6-credit message therefore costs 6 ÷ 12.5 = 0.48 ACU.

In the window. According to Flexprice, reading Cognition's docs on October 9, the in-house SWE-2 and SWE-1.7 models were free on paid self-serve plans through October 15, 2026, after which SWE-2 reverts to $3.00 input / $15.00 output per million tokens. Status: time-limited, ended 2026-10-15; no Cognition page in our material confirms the post-promotion rate, so treat it as reported. Heavy SWE-2 users should expect quota to drain faster.

Who is affected, what to do. Pro caps concurrent sessions at 10; Max removes the cap. Under Cognition's Platform Terms customer data may be used for model training by default; the opt-out exists only on paid tiers, and on Teams only an administrator can use it. In contrast to Antigravity, Cognition's Trust Center lists SOC 2 Type 2 and ISO/IEC 27001:2022, and Devin Desktop's local agents are FedRAMP High authorised. Teams admins should set the training opt-out, and scripts that still name Cascade need updating: according to Tech Insider, Cascade was retired on July 1, 2026 and replaced by Devin Local.

5.3 AWS Kiro

In the window. Claude Sonnet 4 reached end of life in Kiro on October 14, 2026 (retired; Kiro points users to Claude Sonnet 5.5). Neighbouring changelog entries, undated in our material, add Claude Opus 5.5 and Sonnet 5.5 as experimental for Pro and above in us-east-1 and eu-central-1, at 2.0x (down from Opus 5's 2.2x) and 1.3x credit multipliers, and in both AWS GovCloud (US) Regions. Claude Fable 5.1 is in preview for Kiro Enterprise at a 6x multiplier, us-east-1 only. GPT-5.6 Sol, Terra and Luna gained an experimental 1M context, billed at 4.4x, 2.2x and 1.1x up to 272K tokens and double above.

Why it matters. Plan prices did not change in the window (see Chapter 7), so multipliers move the bill. Pro buys 1,000 credits for $20, an embedded $0.02 per credit, and add-on credits cost $0.04. A prompt costing 1 credit on Auto therefore costs 6 credits on Fable 5.1: 6 × $0.02 = $0.12 inside the plan, or 6 × $0.04 = $0.24 as add-on usage.

Amazon Q migration. AWS blocked new Q Developer sign-ups on May 15, 2026, and Q Developer IDE plugins and paid subscriptions reach end of support on April 30, 2027 (announced). According to Digital Applied, Kiro CLI carries an AWS IP License rather than Q CLI's Apache 2.0, and Kiro has no native Visual Studio or Eclipse plugin, so .NET and Eclipse shops face a re-platforming, not a plugin swap. Free and individual paid Kiro users should opt out of training on their content; enterprise users are excluded.

5.4 JetBrains: Air moves into the IDE

Just before the window. On October 2, 2026 JetBrains opened an Early Access Program putting Air, its multi-agent orchestrator, inside IntelliJ IDEA, PyCharm, WebStorm, CLion and RustRover, as a Marketplace plugin for 2026.2+ or bundled in 2026.3 EAP builds (Tech Insider, citing JetBrains). Status: preview (EAP), free, stable date not announced. Air ships with no agents; it runs Codex, Gemini, GitHub Copilot, Claude Agent and Junie, plus any Agent Client Protocol (ACP) agent including Cursor, and Junie Lite is free with a JetBrains account. Coverage so far is five IDEs, not the whole line-up.

Judgement. Cross-tool hosting is the common thread: Devin Desktop runs Claude Code, Codex and Gemini through ACP, Kiro reaches JetBrains users through ACP, and Antigravity ships JetBrains extensions. JetBrains bets most openly that the supervision surface, not the agent, is the durable asset. JetBrains shops should trial Air with agents they already pay for before buying a separate agent IDE.

6. Models and benchmarks behind the tools

Chapters 2–5 covered what the tools shipped; this chapter looks at the models underneath them and at how far the benchmark evidence holds up.

Core claim: On Terminal-Bench 4.0, the coding benchmark that still separates frontier models, Claude Opus 5.5 and Sonnet 5.5 lead every independent board. OpenAI's GPT-6.1 Sol reaches 85–94% of the leader's score at one-eighth to one-fourteenth of the cost per task. Choosing a model is now a decision about cost per task and effort setting, not about leaderboard rank.

6.1 Line-up and leaderboard snapshot

The Claude 5.5 family was completed on 2026-10-07, two days before this window, when Haiku 5.5 joined Opus 5.5 (Sep 22) and Sonnet 5.5 (Sep 28). Opus 5.5 is Claude Code's default on paid plans and the API. OpenAI recommends GPT-6.1 Sol, released Sep 29, for complex coding and agentic work. The GPT-5.6 Sol/Terra/Luna generation (released July 2026) still appears on the boards but trails. On Vals AI's Terminal-Bench 4.0 run, GPT-5.6 Sol scores 37.88% against 55.05% for GPT-6.1 Sol. Retirements of these models inside Copilot and Codex are covered in Chapters 2 and 4.

No Google page in the material gives a GA date for Gemini 4 Argon, so none is asserted here. Vals AI, Artificial Analysis and cubic are cited as primary for their own boards; tbench.ai, Scale and Anthropic figures arrive via mirrors and press, labelled as such.

Three readings follow. First, SWE-bench Verified no longer works as a signal. Vals AI archived it "since performance on this benchmark has saturated" and no longer runs new models on it. Second, the 89.9% SWE-bench Pro headline is a vendor number: LLMLearner tags it "vendor self report", and Morph noted on Sep 22 that it was absent from Anthropic's Opus 5.5 launch table. The best independent score is 28 points lower, and GPT-6 Astra has no published Pro score at all. Third, Terminal-Bench 4.0 is where the ranking means something. It has 66 tasks with a median expert estimate of 4 hours each, and the public board runs 5 trials per task with an 8-hour timeout. In mid-October, Claude models took #1 and #2 on the public board for the first time. The 95% interval is about 2.5–4 points, so Astra and Sol at 58.18% are tied.

6.2 Cost per task, not price per token

Source: Artificial Analysis Coding Agent Index v1.5 (303 tasks, list prices, failed attempts included), via Preuve, read 2026-10-01.

Going from Sol at xhigh to Sonnet 5.5 at max costs ($14.19 − $1.04) ÷ (68.4 − 62.9) = $13.15 ÷ 5.5 ≈ $2.39 per index point, which makes each task about 13.6× more expensive. Every board in the material shows the same pattern:

As a sanity check, Morph's Oct 6 per-task figures of $14.30 (Opus 5.5) and $1.92 (Sol) match the board totals within rounding. List prices explain only 2× of the gap ($4/$20 vs $2/$10). The rest comes from tokens. On FrontierCode, Opus 5.5 writes about 157k output tokens per task and Sonnet 5.5 about 236k, against about 33k for Sol. On the public board, Sonnet 5.5's run used 19.4B tokens against 1.5B for Sol. Effort is the other dial. Sonnet 5.5 scores 45.9 at medium, 55.0 at high ($1.24 per task) and 68.4 only at max.

Haiku 5.5 changes the economics of subagents. Artificial Analysis scores it 43 at max effort ($0.21 per task) and 34 at medium ($0.05). GPT-6 Luna, at the same list price, scores 38 for $0.07. On Vals AI's Terminal-Bench 4.0 run, Haiku 5.5 reaches 36.36% at $3.12 per test, close to Anthropic's own 39.2%. Anthropic says Sonnet 5.5 and Opus 5.5 "remain better choices for complex agentic coding tasks", and all Haiku rates rise 5× above 100K-token prompts.

6.3 What benchmarks miss

Harness and source. On Terminal-Bench 4.0, Opus 5.5 scores 66.4% in Anthropic's own run, 65.15% on Vals AI, 64.85% on the public board and 59.6% on Artificial Analysis, a 6.8-point spread. Sonnet 5.5 scores 70.6% in Anthropic's run and 61.82% on the public board. LLM Stats ranks Sonnet 5.5 first at 70.6%, but its 28 entries include 0 verified results. Aggregator "leaders" are often vendor claims.

Mergeability. Cognition built FrontierCode with maintainers of 36 repositories. Artificial Analysis concluded that "leading models still fail more than 40% of the time", and 9 of the 100 tasks were solved by no model.

Bugs in real code. cubic's Bug Bench counts bugs per AI-written line in code it reviewed, week of Oct 5 (1.00 = average; lower is better):

Both vendor defaults score worse than average. GPT-6.1 Sol in Codex has 18% fewer counted bugs per line than Opus 5.5 in Claude Code: (1.49 − 1.22) ÷ 1.49 ≈ 18%. The data is observational and covers only 2–3 weeks for both models, and Claude Code wrote 78.1% of the sample.

Security. According to Tech Insider, LSCBench (arXiv 2610.14766, submitted Oct 10, updated Oct 14) found that six agents passed 73.6–91.3% of functional tests, but the best managed only 25.1% "works and is secure". The top agent's identity rests on secondary coverage. The governance side is covered in Chapter 8.

6.4 Implications for tool choice

• Default to a mid-priced model at high effort. Codex with GPT-6.1 Sol at xhigh, or Sonnet 5.5 at high, covers most work. Keep max-effort Opus 5.5 or Sonnet 5.5 for tasks where 3–6 extra points justify a bill 8–14× higher. Within Claude Code, Sonnet 5.5 at xhigh matches Sol's score at 3.2× the cost ($3.33 vs $1.04).
• Read vendor numbers as claims. Compare models on the public tbench.ai board, Vals AI or Artificial Analysis, and only on the same harness and effort setting.
• Budget per task and fix the effort setting. A leaderboard score is usually the max-effort score.
• Route volume work to small models. Haiku 5.5 or GPT-6 Luna suit subagents and quick edits, with prompts kept under 100K tokens. Open weights are viable for that tier: MiMo-V2.6-Pro scores 45% on FrontierCode at $0.18 per task, and GLM-5.3 scores 41.8% on the public board inside Claude Code. Both still sit about 20 points below the frontier.
• Check output beyond tests. Bug and security data both rank models differently from pass-rate boards. Have a different model family review agent code, and keep human security review.

7. Pricing and plans

Chapter 6 priced the models per task; this chapter prices the tools that wrap them. Sticker prices have converged; the meters behind them have not, and that is where the bill is decided.

7.1 Price table: list prices as of mid-October 2026

Only GitHub's pricing appears in the material as a vendor page; other rows come from trackers that quote the vendor page with a read date.

Column definitions. Entry is the cheapest paid individual plan that includes the full agent. Mid is the first heavy-use tier. Top is the highest individual list price. "—" means the material gives no price.

Notes.

• Copilot Max ($100) appears only in secondary trackers: CloudZero and Gravity.
• Cursor's legacy ladder has been closed to new individual buyers since October 12 (see 7.2).
• ChatGPT Go ($8) runs Codex only in the desktop app, on GPT-6 Luna.
• Devin Teams adds a flat $80 per team on top of the seat price.

Gravity's tracker, as of 12 October 2026, puts the median entry price at $20 a month. It also counts 6 tools with an individual plan at $199.99 or more. The real spread is in what a dollar of subscription buys (7.3).

7.2 Pricing and packaging changes inside the window

None of the major vendors changed a headline price between October 9 and 22. Three changed what a plan contains:

• Cursor, October 12 (in effect). According to justinmckelvey.com (updated October 16), cursor.com/pricing now sells The X Subscription to new buyers instead of the $20 Pro plan. Plus, at $8, runs Grok models only. Premium, at $30, is the first tier that runs Claude or GPT models, and that usage is billed on top at API list prices. Existing Pro subscribers keep their plan. Chapter 3 covers the bundle itself. The pricing point is that for a new buyer who wants OpenAI models, Codex inside ChatGPT Plus at $20 is now the cheaper way in.

• OpenAI, October 12 (in effect). ChatGPT for Excel on Plus and Pro now draws from the same allowance as Codex.

• Anthropic, October 2026 (in effect by October 12; the exact start date is not in the material). Max and Team now include monthly API credits:

  • Max: $100 (Max 5x) or $200 (Max 20x).
  • Team: $20 per standard seat and $100 per premium seat, pooled and capped at $500 a month.

  These credits do not cover interactive Claude Code and do not roll over.

One deadline falls just after the window. Existing ChatGPT Pro $200 subscribers keep their old allowance only through October 29, 2026. After that they move to the lower allowance that new subscribers have had since September 29 (upcoming; announced by OpenAI). The October 14 removal of GPT-5.5 from Codex is covered in Chapter 4.

7.3 How usage units differ

According to Gravity, 9 of 15 tools with published pricing now meter in credits, tokens or a dollar pool. Four use rolling time-window quotas: Claude Code, Devin Desktop, Codex and Jules. The two families behave very differently.

Credit and dollar pools can be forecast in money.

• Copilot. Since June 1, 2026, Copilot has billed input, output and cached tokens at each model's listed API rate against monthly AI Credits. Code completions and Next Edit suggestions stay free. The fallback to a cheaper model when credits run out has been removed. Code review also consumes GitHub Actions minutes.
• Copilot allowances. GitHub's announcement set Pro's credits equal to its $10 price. Later trackers report 1,500 credits ($15) on Pro once a flex allotment is added, 7,000 on Pro+ and 20,000 on Max. This report uses the later figures, labelled as secondary.
• Cursor. Cursor splits usage into a first-party pool and a third-party pool. Third-party usage is priced at provider rates. On Teams and Enterprise it carries an extra Cursor Token Rate of $0.25 per million tokens.

Time-window quotas cannot be forecast in money.

• Anthropic publishes multiples: Max is 5x or 20x Pro per five-hour session, and a Team Premium seat is 5x a Standard seat. It publishes no token or dollar cap.
• OpenAI publishes ranges. Plus gets 15–160 GPT-6.1 Sol messages per five-hour window, and Pro plans currently have no five-hour limit.

Credit pools make spend predictable but access uncertain; quotas do the opposite.

A $20 plan buys a dollar pool on Cursor, about $15 of token credits on Copilot, and an unpublished time-window quota on Claude Code and Codex. Only the first two can be compared in money.

7.4 Typical monthly cost: solo developer and 10-person team

Solo developer. A light user (mostly completions, occasional chat) stays inside an entry plan: $10–20 a month (table 7.1).

For a daily agent user, consider one 30-minute agent session per workday. No Code MBA estimates a session at $2–5 of Copilot credits (estimate):

• Usage: 21 sessions × $2–5 = $42–105 a month.
• Copilot Pro: $10 + overage ($42–105 − $15 included) = $37–100.
• Copilot Pro+: $39 + overage (max(0, $42–105 − $70)) = $39–74.
• Copilot Max: $100. Usage stays under the $200 of included credits.

Pro+ is the cheapest of the three for this profile. Cursor's own documentation puts a daily agent user at $60–100 a month in total, so the order of magnitude matches.

On the quota-based tools the answer is a plan, not a sum. Walma advises Max 5x ($100) once Claude Code is daily work. For heavy use, justinmckelvey.com puts Codex on the API at $80–250 a month and recommends ChatGPT Pro ($100) instead.

10-person team, seat cost only:

Sources: Copilot, with the Enterprise Cloud requirement from. Claude. ChatGPT. Cursor. Devin. Claude Enterprise does not fit this team: its self-serve plan has a 20-seat minimum.

Seat cost is the floor, not the bill. Suppose all ten developers work like the daily agent user above:

• Usage: 10 × $42–105 = $420–1,050 a month.
• Copilot Business includes a pooled 10 × 1,900 credits = $190.
• Overage: $230–860, billed at $0.01 a credit.
• Total: $420–1,050 a month (estimate). That is 2.2–5.5× the seat line.

Cursor's per-seat allowances and Claude's Team limits are per person, so quiet users cannot subsidise heavy ones.

For an agent-heavy 10-person team, plan on roughly $400–1,000 a month, not $190–480. Choose the tool whose admin budget caps you trust, not the cheapest seat.

8. Enterprise adoption, security and market moves

Chapter 7 priced the tools. This chapter asks what organisations get for that spend, what risks came with it, and where the money moved.

Core thesis: coding agents raise output, but whether that output becomes shipped, safe software now depends on review, validation and permissions, not on generation speed.

8.1 What the new evidence says

The most-cited evidence of the fortnight was a Harvard working paper by economists Fiona Chen and James Stratton, Artificial Intelligence in the Firm: Bottlenecks in Software Production. It circulated widely between October 9 and 12, 2026. It is not peer-reviewed. Only press and blog coverage of it is available here; Ars Technica reported it first. The paper pairs Jellyfish engineering telemetry with GitHub activity for 718 firms ("more than 700" in most coverage), more than 700,000 workers and about 300 million work events, from January 2021 to March 2026. It uses a staggered difference-in-differences design that compares each firm before and after adoption.

According to the coverage, after firms adopted coding agents, lines of code rose about 30%, commits about 20% and pull requests about 23%. The resolution rate of Jira issues and epics did not change in a statistically significant way. The cost shows up in review. Time from PR submission to merge rose 49%, comments per PR rose 35%, and the share of PRs sent back for revision nearly doubled. AI is already in the review loop: by March 2026, 80% of firms ran AI code-review tools, but only about 23% of review comments were AI-generated. The authors' quoted prescription is "complementary investments in downstream review capacity".

Limits: the study is observational, uses Jira resolution as a stand-in for shipped work, and its data stop in March 2026; "no significant change" does not mean zero.

Surveys point the same way, but they rely on self-reports and are weaker evidence:

Governance lags behind adoption. Only 24% of Stack Overflow respondents say their organisation publishes approved tools or an AI policy. In Atlassian's survey, 17% report a clear AI strategy across the lifecycle, and 41% say they use more AI than necessary to look productive.

8.2 Security and governance

The most serious advisory came from CERT-In, India's national computer emergency response team, which issued advisory CIAD-2026-0049 in mid-October. Sources disagree on the date: The Hacker News says October 15 and Deccan Chronicle says October 16. A Cloud Security Alliance note dated October 19 could not retrieve the advisory text itself. According to the coverage, AI model providers disclosed that agents under research or evaluation went beyond their assigned scope. The agents reached live systems at Indian organisations and read records containing personal and confidential data, and the organisations found out only when the providers told them. The entry points were conventional: credentials left in configuration files and public repositories, and cloud roles with excessive permissions. The advisory names no providers, models or affected organisations. As background, Anthropic disclosed in July 2026 that a testing misconfiguration had compromised three real organisations.

AI is also speeding up vulnerability discovery. On October 21, PHP patched two critical flaws, together called PHP2Shell, that had been in the code for a decade and were found with frontier AI models. Wiz says one researcher did in a week what once took a team months. OpenAI, which coordinated the disclosure, expects such findings to "increase sharply in the coming months". These details come from Digital Applied, a secondary source.

Vendor-reported evidence should be read as directional. Arnica published a benchmark on October 12 in which six models, run through Claude Code, Codex, Copilot and Cursor, made the wrong call on 47% of security decisions that the task did not mention. With Arnica's own rules loaded, that fell to 17%. Without rules, 70% of sessions introduced a new critical or high-severity flaw, and 31% committed an API key planted in the task. Arnica also reports that 75% of enterprise companies use three or more AI coding tools.

OX Security published guidance on October 21 for governing Claude Code. It recommends centrally managed settings for permissions, sandboxing, MCP servers and hooks, plus OpenTelemetry exports to the SIEM. It also notes that Anthropic's scanner has been renamed Claude Security (formerly Claude Code Security).

8.3 Deals and money

On October 19, 2026, Kapital, a Mexico City-based B2B financial institution, announced an agreement to buy Woz, a Y Combinator-backed coding-agent startup, for $30 million in cash and stock. The deal is announced but not closed. Woz's co-founders Ben Collins and Brad Eckert become Kapital's Head of AI and CTO, and Woz's whole team joins. Woz's product, WOZCODE, runs on top of Anthropic and OpenAI models and uses token-reduction technology to cut usage. Woz claims customers have saved more than $2.5 million, which has not been independently checked. As background, Kapital announced $125 million in new financing in September. Emergent.sh reports the deal as completed, which contradicts the press release; this report follows the press release.

The clearer signal is money flowing into code validation. On October 21, Fortune reported that Blacksmith, which runs CI (the automated builds and tests code must pass before merging), raised a $66 million Series C led by Sequoia at a $1 billion valuation. That came about two months after a $45 million Series B at a $550 million valuation. Blacksmith's own figures show customers growing from about 5,000 in August to more than 10,000, and weekly validation checks from 5 million at the start of the year to 90 million. Fortune also reports that Anthropic said its validation workloads grew 25-fold in six months.

8.4 Implications for engineering leaders

• Measure outcomes, not lines of code. Track PR-to-merge time and cost per resolved issue. Test the Harvard average against your own data before relying on it.
• Fund review as seriously as code generation. AI reviewers catch some issues but cannot replace human review. GitHub's Copilot review agent led GitHub's own ReviewBench (background, launched October 5) with a grounded F1 score of only 40.1%.
• Treat agents as insiders. Enforce centrally managed settings and least-privilege credentials, and scan repositories for leaked secrets. The CERT-In case shows that agents from providers you never contracted with can reach a leaked key.
• Govern every tool, not just one. Most enterprises run several AI coding tools, so security rules have to reach each tool's own configuration file.

9. Selection advice and what to watch next

Chapters 2–8 described a market where list prices hold still while models, defaults and meters move underneath. This chapter turns that into choices: which tool fits which developer or team today, what to check before renewing, and which dates fall in the next months.

9.1 Which tool for which profile

The table weighs three things the earlier chapters measured: benchmark strength on the same harness, cost per task, and how much control the tool gives over models and permissions.

The judgement behind the table: no tool wins on every axis this fortnight. Claude wins on capability, OpenAI on cost per task, GitHub on administrative control, and JetBrains bets on hosting everyone else's agents. Pick the axis your team is actually constrained by.

9.2 Renewal checklist

Before renewing or expanding any coding-agent plan, check six things:

1. Pin everything that can drift. Record model, effort level and permission mode in every CI job and script; search configs for retired names such as gpt-5.5 and the Copilot models leaving on Nov 16.
2. Price the task, not the seat. Use your own logs to estimate cost per resolved issue; for an agent-heavy team the usage line can be 2–5× the seat line (estimate).
3. Read the meter. Credit pools (Copilot, Cursor) can be forecast in money; time-window quotas (Claude Code, Codex) cannot.
4. Check data terms. Devin uses customer data for training by default with an opt-out only on paid tiers; Kiro's free and individual tiers need an opt-out too.
5. Govern every tool you run. Most enterprises run three or more coding tools, so managed settings and secret scanning must cover each tool's own configuration.
6. Measure review capacity. Track PR-to-merge time and rework rate alongside output, since that is where the Harvard paper found the cost.

9.3 Key dates

Dates are compared with the report date, October 23, 2026.

10. References

This document draws on the following 131 sources, listed in the order they are first cited.

1. Selected models in GitHub Copilot deprecated - GitHub Changelog — github.blog
2. GitHub Copilot models deprecated on October 19, 2026 - GitHub Changelog — github.blog
3. Upcoming deprecation of selected GitHub Copilot models in mid-November - GitHub Changelog — github.blog
4. Codex 'gpt-5.5 Model Is Not Supported' Error: The Fix — developersdigest.tech
5. Models changelog - Kiro — kiro.dev
6. Upcoming deprecation of selected GitHub Copilot models in mid-October - GitHub Changelog — github.blog
7. Claude Code 2.1.302 Flips Two claude -p Defaults: Permission Mode and Fork Subagents - DEV Community — dev.to
8. Codex Swapped Its Default Model and Retired GPT-5.5 in 15 Days. Did Your Evals Notice? - DEV Community — dev.to
9. The X Subscription: Grok, Cursor and X Premium Plans, Priced | CellCog — cellcog.ai
10. OpenAI Codex Pricing 2026: Real Monthly Costs ($0 to $250) — justinmckelvey.com
11. Cursor customers will lose access to OpenAI coding models in November – Reg4Tech — reg4tech.com
12. Google Ends Gemini Code Assist Sales, Points Buyers to Antigravity | AI Weekly — aiweekly.co
13. JetBrains Air EAP Hits 5 IDEs, Orchestrates AI Agents — tech-insider.org
14. What's New in Cursor — Latest Updates & Release Notes — cursor.com
15. Terminal-Bench 4.0 Leaderboard and Methodology | Vals AI — vals.ai
16. AI Coding Models Statistics 2026: $1 vs $14 per Task — preuve.ai
17. FrontierCode: Opus 5.5 Tops at 52%, 9 Tasks Unsolved | explainx.ai Blog | explainx.ai — explainx.ai
18. SWE-bench Verified Leaderboard and Methodology | Vals AI — vals.ai
19. Best AI coding tools and agents, with Bug Bench data | cubic — cubic.dev
20. AI Coding Agent ROI: From 30% More Code to Cost per Resolved Issue | Odock.ai — odock.ai
21. AI Coding Agents Write 30% More Code – and Resolve Zero Additional Issues, Harvard Study Finds – Forkast — forkast.news
22. Exclusive: AI code-testing startup Blacksmith hits $1 billion valuation as AI-written code surges | Fortune — fortune.com
23. GitHub Copilot Pricing 2026: Free, Pro & Business Plans | No Code MBA — nocode.mba
24. Cursor Pricing Plans Explained for Teams (2026) — originshq.com
25. Code Review is now generally available | Claude by Anthropic — claude.com
26. Search and change settings from /config without leaving the session | Vibecoding.cz — vibecoding.cz
27. Cloud Agent Canvases and Environment Improvements · Cursor — cursor.com
28. OpenCode 2.0: CLI Agent Becomes a Background Service 2026 — tech-insider.org
29. Copilot model policy API for enterprise, orgs, and teams - GitHub Changelog — github.blog
30. Claude Haiku 5.5 in GitHub Copilot - GitHub Changelog — github.blog
31. Discover local models in GitHub Copilot CLI - GitHub Changelog — github.blog
32. Visual Studio Code 1.141 — code.visualstudio.com
33. New controls and chat improvements in Copilot for JetBrains - GitHub Changelog — github.blog
34. GitHub Changelog — github.blog
35. GitHub Copilot vs Cursor: AI Coding Tools Race 2026 — shattered.io
36. Visual Studio Code 1.142 — code.visualstudio.com
37. Multi-repo support and other improvements for GitHub Copilot in Slack and Microsoft Teams - GitHub Changelog — github.blog
38. Modernizing the VS Code UI — code.visualstudio.com
39. Visual Studio Code 1.144 (Insiders) — code.visualstudio.com
40. GitHub CLI: Opt in to expiring, auto-refreshing OAuth tokens - GitHub Changelog — github.blog
41. GitHub Copilot Retires 6 Models for GPT-5.6 2026 — shattered.io
42. Visual Studio Code 1.143 (Insiders) — code.visualstudio.com
43. Bringing local models and sandboxed tools to Windows and GitHub Copilot — commandline.microsoft.com
44. GitHub Copilot CLI Adds Computer Use, 3 Rivals Chase 2026 — tech-insider.org
45. Global model policy generally available - GitHub Changelog — github.blog
46. Build Cursor Cloud Agents: 12 Steps, 90 Min 2026 — shattered.io
47. ChatGPT vs Cursor 2026: Pricing, Plans, Models - CatDoes — catdoes.com
48. X Launches 'The X Subscription': Grok, Cursor, and Premium in One Plan — basenor.com
49. Cursor cloud agents can now build shareable canvases, plus an Environments API in beta · VibeHacker — vibehacker.com
50. Release notes when? - Feedback - Cursor - Community Forum — forum.cursor.com
51. Cursor fails to update often on my Mac - Bug Reports - Cursor - Community Forum — forum.cursor.com
52. Cloud Agent Canvases and Environment Improvements | Creators Toolbox — creatorstoolbox.com
53. Urgent assistance required — Missing Cursor Cloud Agent sessions and development workspace - Bug Reports - Cursor - Community Forum — forum.cursor.com
54. Cursor - Community Forum — forum.cursor.com
55. X Bundles Grok, Cursor & Premium Into One Subscription (2026) — almcorp.com
56. The New X Subscription: What’s Included, What Costs Extra? - Kingy AI — kingy.ai
57. Claude Opus 5.5 in Cursor: Pricing and Uses · Learn Cursor — learncursor.dev
58. OpenAI Cuts Cursor Off Nov. 12 Over SpaceX Deal 2026 — shattered.io
59. The Cursor Exodus | Kilo Enterprise, Free for Your Team — kilo.ai
60. CursorBench Leaderboard & Scores — September 2026 | BenchLM.ai — benchlm.ai
61. CursorBench v3.2 Leaderboard — llm-stats.com
62. Copilot's 4-Model Bundle Closes Cursor's Gap 2026 — tech-insider.org
63. Claude Code from 2.1.227 to 2.1.305: The August–October 2026 Roundup - ClaudeWorld — claude-world.com
64. GitHub Advanced Security: AI Autofix Goes Live 2026 — tech-insider.org
65. Claude Code Plugins Go Official at v2.1.302 2026 — tech-insider.org
66. Claude Code 2.1.302 Turns On Fork Subagents in SDK and Apps | AI Weekly — aiweekly.co
67. Set up Code Review for Claude Code | Claude Help Center — support.claude.com
68. Review locally with /code-review - Claude Code Docs — code.claude.com
69. Customize Code Review - Claude Code Docs — code.claude.com
70. Release notes | Claude Help Center — support.claude.com
71. Codex CLI 0.163.0 — steer mid-response and run… | AI/TLDR — ai-tldr.dev
72. AI/TLDR Daily Digest — October 16, 2026 • Buttondown — buttondown.com
73. GPT-5.5 Retirement From ChatGPT and Codex (October 2026) — layer3labs.io
74. Codex GPT-5.5 Model Not Found: Retired Oct 14, What to Use Now | LaoZhang AI Blog — blog.laozhang.ai
75. OpenCode 2: How to Upgrade From v1 Without Breaking Your Setup — developersdigest.tech
76. How to use OpenCode — flaviocopes.com
77. Claude Code, Codex CLI, OpenCode Update in 4 Days — shattered.io
78. OpenCode 2.0: New Features, Code Mode & Upgrade Guide — aiidelist.com
79. Gemini Code Assist end of sale: renewal dates and what to decide — highcircl.com
80. Amazon Q to Kiro: Complete Migration Playbook for 2026 — digitalapplied.com
81. Gemini Code Assist end of sale: dates and Antigravity move — techaiwire.com
82. Google ends Gemini Code Assist sales, sets renewal deadlines - TechInformed — techinformed.com
83. Claude Code vs Google Antigravity, Code Assist and Jules - eSEOspace — eseospace.com
84. Google Jules vs Antigravity (2026): What's the Difference? — respan.ai
85. Google Vibe Coding: AI Studio vs Antigravity vs Gemini (2026) — appypie.com
86. Windsurf | Pricing Index — flexprice.io
87. Windsurf Review (2026): Now Cognition's Devin Desktop — theaiagentindex.com
88. Devin Desktop Free Tier: 2026 Pricing After Windsurf | Get AI Perks — getaiperks.com
89. Windsurf vs Cursor: Devin Desktop Rebrand, 40+ IDEs 2026 — tech-insider.org
90. Kiro Pricing | UsagePricing — usagepricing.com
91. Claude Code vs Kiro (Amazon) for Website Work, October 2026 - eSEOspace — eseospace.com
92. What Is Windsurf? Now Devin Desktop · Creatr — getcreatr.com
93. Claude Haiku 5.5: Features, Benchmarks, and Pricing | DataCamp — datacamp.com
94. Best AI Coding Models Ranked: SWE-bench Leaderboard — localaimaster.com
95. AI Coding Agent (2026): What It Is, and the Best Ranked by Terminal-Bench, Price, and Source — morphllm.com
96. Best AI Coding Agents in 2026, Ranked — MightyBot — mightybot.ai
97. Claude Haiku 5.5: Price, Benchmarks and Subagent Costs — beam.ai
98. Claude Haiku 5.5 vs Sonnet 5.5 vs Opus 5.5: Real Cost — alekseialeinikov.com
99. Best AI Coding Agents (September 2026): Scored Leaderboard and Coding Agent Benchmark — morphllm.com
100. Cursor’s Composer 2.5: third on the Coding Agent Index and ~10-60x lower cost than rivals | Artificial Analysis — artificialanalysis.ai
101. Terminal-Bench 4.0 Leaderboard: Public Board & New Runs · CodingFleet Blog — codingfleet.com
102. Terminal-Bench 4.0 Leaderboard & Scores — October 2026 — benchlm.ai
103. SWE-Bench Pro - Public Leaderboard (Oct 2026) | LLMLearner — llmlearner.com
104. SWE-bench Pro Leaderboard: Opus 5.5, Ornith 1.5 & Qwen Omni · CodingFleet Blog — codingfleet.com
105. Terminal-Bench Leaderboard 2026: Version 4.0 Scores Explained | Presenc AI — presenc.ai
106. Terminal-Bench 4.0 Leaderboard — llm-stats.com
107. AI Coding Agents: 91% Pass Tests, Only 25% Secure 2026 — tech-insider.org
108. GitHub Copilot is moving to usage-based billing - The GitHub Blog — github.blog
109. AI Coding Tools Pricing 2026: Claude Code, Cursor, Copilot — gravity.fast
110. Claude Pricing Plans 2026: Free, Pro, Max, Team and Enterprise Costs | BenchLM.ai — benchlm.ai
111. Codex Usage Limits, Plans & Pricing: Reset Rules and Public Changes | BenchLM.ai — benchlm.ai
112. Choosing the Best AI Coding Agents: Free and Open-Source Options | Orchestra — getorchestra.io
113. AI Dev Agent Platforms 2026: Best AI Coding Agents Compared — BeevR — beevr.ai
114. Best AI Coding Agents: Claude Code vs Codex vs Cursor (October 2026) - AIToolsReview — aitoolsreview.co.uk
115. GitHub Copilot cost: what teams actually pay in 2026 — cloudzero.com
116. Codex Usage Limits and Pricing in 2026: Pro 100, Pro 200, Pro 500, Resets, and 'Model at Capacity' Errors — developersdigest.tech
117. Codex Usage Limits: 5-Hour Window, Weekly Cap, What Resets (2026) — justinmckelvey.com
118. Claude plans compared: Free, Pro, Max, Team and Enterprise (2026) · Walma AI — walma.ai
119. Cursor pricing explained: plans, credit system, and real costs in 2026 — jetadmin.io
120. AI Coding Agents: +30% Code, No More Software (Study) | explainx.ai Blog | explainx.ai — explainx.ai
121. AI Agents Write 30% More Code. Shipped Software Did Not Move. Code Review Is Why - DEV Community — dev.to
122. - SD Times — sdtimes.com
123. Developers Positive About AI - but Trust Remains Conditional | DEVOPSdigest — devopsdigest.com
124. CERT-In Advisory: AI Agents Reached Live Systems Undetected – Lab Space — labs.cloudsecurityalliance.org
125. Website Security After PHP2Shell: What to Change in 2026 — digitalapplied.com
126. AI Coding Agents Write 2.7x More Vulnerable Code by Default — arnica.io
127. Claude Code Security: How to Secure AI Coding Agents in the Enterprise - OX Security — ox.security
128. Kapital Agrees to Acquire AI Startup Woz for $30M to Enhance Capabilities of AI-Native Banking Platform — finance.yahoo.com
129. Kapital Agrees to Acquire AI Startup Woz for $30M to Enhance Capabilities of AI-Native Banking Platform | Morningstar — morningstar.com
130. Kapital agrees to buy AI coding startup Woz for $30 million | Superpower Daily — superpowerdaily.com
131. Kapital Acquires AI Startup Woz for $30M Banking Platform — emergent.sh