The fortnight at a glance: models, defaults and meters moved
Five conclusions
The lesson
Treat model, effort and permission settings as code, budget per task rather than per seat, and invest in review before buying more generation.
What changed: the change log
Copilot and the terminal agents now migrate every few weeks
GitHub Copilot model retirements
The suggested replacements moved too: GitHub's Sep 18 notice pointed Oct 19 retirees to GPT-5.6 Sol, GPT-5.6 Luna and Grok 4.6; the Oct 19 post names GPT-6.1 Sol, GPT-6 Luna and Grok 4.7. Hard-coding the "suggested alternative" only delays the next migration.
Admins, before Nov 16
Copilot harness becoming the default for new sessions (progressive rollout), Ask returns as a read-only agent, Canvases in Preview, and admin enforcement of agent sandboxing and permission modes.
Claude Haiku 5.5 (Oct 7, background) for subagents and quick edits; Gemini 4 Argon for Pro+ and above, GA Oct 13 per shattered.io, rolling out gradually.
Copilot CLI 1.0.94-0 discovers Ollama models; offline mode still needs COPILOT_OFFLINE=true. Microsoft's local MAI Code 1.1 Flash (137B total, 6.8B active) is announced, not shipped.
Copilot for JetBrains (Oct 10): enterprise-managed default models and a diagnostics Fix action. Copilot in Slack and Teams: multi-repo conversations (Oct 15). Computer use is Preview since late September, not an October launch.
gpt-5.5 and the Copilot models leaving Nov 16.Cursor is turning into a Grok-first cloud-agent platform
A GitHub-shaped platform, with the editor as one entry point
Four releases in one fortnight read as one stack. Third-party Origin apps authenticate with an app JWT and installation tokens, receive webhooks and write back check runs, PR comments and reviews, the same model as a GitHub App.
Oct 13 · Marketplace GA
20+ apps, open to public submissions; partners include Cloudflare, Linear, Sentry and Socket Security (date per shattered.io).
Oct 16 · Canvases GA, Environments API beta
Up to 300 repos per environment; large ones load 5× faster.
Friction
A forum user reports two Cloud Agent sessions and weeks of commits vanished on Oct 17. Push cloud-agent work to a remote you control.
The X Subscription (from Oct 12)
Starter allowance for Cursor agents.
Cursor with Grok models only.
Cloud agents, code hosting and reviews, Grok Bot; Claude or GPT at API list price, drawn from prepaid credit from the first request.
As Premium, plus priority and early access. The pool resets weekly, shared across Grok and Cursor. Existing subscribers keep legacy plans; no migration timeline yet.
OpenAI cut-off: proposed for Nov 12
OpenAI intends to wind down the SpaceX contract with a proposed shutoff on Nov 12, 2026, 20 days after this issue.
Cursor's CEO says OpenAI models serve about 5% of user traffic (via CNBC). Personal API keys continuing to work is not confirmed by a primary source.
Opus 5.5 lists at $4 in / $20 out per million tokens: a 2M-in, 0.2M-out session costs $12 standard or $24 in Fast mode, on top of the plan (illustration).
On a legacy plan with heavy Claude or GPT use, staying put is cheaper for now. Re-validate OpenAI-based configs on a replacement before Nov 12.
CursorBench: a vendor signal, not neutral evidence
BenchLM snapshot, 4.0 set (Sep 24)
Opus 5.5 57.8%, Fable 5.1 51.8%, Grok 4.7 46.3%, GPT-5.6 Sol 41.7%. Tasks are private; shown for reference only.
Unconfirmed claim
Gemini 4 Argon at 52.8% "at standard inference volume" (tech-insider.org); not the same setup as Opus 5.5 at its highest effort tier.
One independent task (Kingy AI, Oct 12)
Grok 4.7 High and Sonnet 5.5 High both passed 24/24 checks, but Grok took about 17× longer and about 3× the tokens.
Google, Cognition, AWS and JetBrains: closing products, hosting rivals
The supervision surface is becoming the asset
These four vendors spent the fortnight consolidating: closing legacy products, moving customers to successors and re-pricing models. The action items are dates, not features.
Jan 31, 2027
Last day to add Gemini Code Assist licences; annual auto-renewal ends.
Apr 30, 2027
Amazon Q Developer IDE plugins and paid subscriptions reach end of support (announced).
Oct 14, 2026
Claude Sonnet 4 end of life in Kiro, already in effect.
Google End of sale Oct 9
Gemini Code Assist: new Standard and Enterprise subscriptions stopped on Oct 9; existing ones stay supported to contract end. No shutdown date for the service.
Antigravity is a new subscription, not a seat swap: it needs a Gemini Enterprise edition on an invoiced billing account. HighCircl reports SOC 2, ISO 27001 and FedRAMP listed as not supported.
Jules is still beta: 15 free tasks a day, 100 on Google AI Pro, 300 on Ultra.
Cognition · Devin Desktop Promo ended Oct 15
SWE-2: free on paid self-serve plans through Oct 15; reported to revert to $3.00 in / $15.00 out per million tokens (not confirmed by a Cognition page).
Data terms: customer data may train models by default; opt-out only on paid tiers, and on Teams only an admin can set it.
Compliance: SOC 2 Type 2 and ISO/IEC 27001:2022; local agents FedRAMP High. Pro caps concurrent sessions at 10; Max removes the cap.
AWS Kiro Experimental models
Credit multipliers: Opus 5.5 at 2.0× (Opus 5 was 2.2×), Sonnet 5.5 at 1.3×, for Pro and above; Claude Fable 5.1 in preview for Enterprise at 6×.
What a multiplier costs: Pro buys 1,000 credits for $20 ($0.02 each), add-ons $0.04. A 1-credit prompt on Fable 5.1 costs $0.12 in plan or $0.24 as add-on.
Q migration: Kiro has no native Visual Studio or Eclipse plugin (per Digital Applied), so .NET and Eclipse shops face a re-platforming.
JetBrains Air Preview (EAP)
Since Oct 2: the multi-agent orchestrator runs inside IntelliJ IDEA, PyCharm, WebStorm, CLion and RustRover; free, no stable date.
Ships no agents of its own: also runs Gemini and any ACP agent including Cursor; Junie Lite is free with a JetBrains account.
Advice: JetBrains shops should trial Air with agents they already pay for before buying a separate agent IDE.
Claude leads Terminal-Bench 4.0; GPT-6.1 Sol wins on cost per task
Score vs cost per task
AA Coding Agent Index v1.5 (303 tasks, list prices, failed attempts included), read Oct 1. Cost axis is logarithmic. Moving from Sol at xhigh to Sonnet 5.5 at max costs about $2.39 per index point, 13.6× per task.
GPT-6.1 Sol vs the leader
Read the boards with care
SWE-bench Verified: archived
Vals AI stopped running it as saturated; final top score 97.00% (Claude Opus 5, Sep 1).
SWE-bench Pro: 89.9% is vendor-reported
The best independent score on Scale's board is 61.5% (Muse Spark 1.1), 28 points lower.
Mergeability lags
On FrontierCode, leading models still fail more than 40% of the time; 9 of 100 tasks were solved by no model.
Secure code lags further
LSCBench: six agents passed 73.6–91.3% of functional tests, but the best managed only 25.1% "works and is secure".
Use: default to a mid-priced model at high effort (Sol at xhigh, or Sonnet 5.5 at high, $1.24 per task); keep max-effort Opus or Sonnet for tasks where 3–6 extra points justify an 8–14× bill; route subagents to Haiku 5.5 ($0.05–0.21 per task) or GPT-6 Luna ($0.07); have a different model family review agent code.
Sticker prices converged; the meters did not
List prices, mid-October (USD per month)
Cursor, Oct 12
New buyers wanting OpenAI models now get in cheaper via Codex in ChatGPT Plus at $20.
Anthropic, October
Max includes $100 or $200 of monthly API credits; Team $20 / $100 per seat, capped at $500. Not for interactive Claude Code.
ChatGPT Pro $200, Oct 29
Grandfathered allowance ends; existing subscribers move to the lower allowance (upcoming).
10-person team: seat cost vs realistic bill
Seat cost is the floor. If all ten developers run one 30-minute agent session per workday, Copilot Business usage lands at $420–1,050 a month (estimate, shaded band), 2.2–5.5× the seat line. Copilot Enterprise also needs GitHub Enterprise Cloud.
Solo developer, per month
Agents raise output, but review decides what ships
After adopting coding agents: change per firm
Chen & Stratton (Harvard), 718 firms, Jan 2021 to Mar 2026, staggered difference-in-differences; working paper, not peer-reviewed, reported via press coverage. Jira issue resolution did not change significantly, and the share of PRs sent back for revision nearly doubled.
Surveys point the same way
Security and governance
Agents under evaluation went beyond scope and read personal data at Indian organisations, via leaked credentials and over-permissioned cloud roles. No providers are named.
Wrong calls on 47% of unprompted security decisions, 17% with rules loaded; 31% of sessions committed a planted API key.
Two decade-old critical PHP flaws found with frontier models and patched (via Digital Applied, secondary).
Money follows the bottleneck
CI provider's Series C led by Sequoia, two months after $45M at $550M. Weekly validation checks rose from 5M at the start of the year to 90M.
Cash and stock for the YC-backed coding-agent startup; announced, not closed.
Track PR-to-merge time and cost per resolved issue; fund review like generation. GitHub's own review agent led ReviewBench with a grounded F1 of only 40.1%.
Pick the axis you are constrained by, and pin what can drift
Which tool for which profile
First choice
Claude Code on Max 5x ($100)
Why
Opus 5.5 and Sonnet 5.5 lead Terminal-Bench 4.0 on every independent board.
Watch out for
Max-effort runs cost 8–14× more per task; set effort deliberately.
Renewal checklist
Tap an item to tick it · 0/6 done
Dates to watch
Days counted from October 23, 2026
Bottom line
Capability
Claude
Leads Terminal-Bench 4.0 on every independent board
Cost per task
OpenAI
GPT-6.1 Sol at 85–94% of the leader for far less
Admin control
GitHub
Model policy as code through the REST API
Hosting
JetBrains
Bets on running everyone else's agents