The window produced two Chinese Flash models and no closed launch, so the week's story is price, access and trust
Between Oct 17 and Oct 23, 2026 the window produced two model releases, DeepSeek V4.1-Flash-1017 and MiniMax M3.1 Flash. The newest closed releases, Meta's Muse Wonder (Oct 12) and Gemini 4 Argon's general availability (Oct 13), are context. Every status is judged against Oct 24.
Five conclusions
No closed model launched; the closed news is Google's price schedule
The newest closed releases are Meta's Muse Wonder (Oct 12) and the general availability of Gemini 4 Argon through the paid API on Oct 13. Argon costs $2 / $10 per 1M tokens until Nov 30 and $4 / $20 from Dec 1, so an illustrative 60M-input, 15M-output month rises from $270 to $540. Until Nov 30 Argon is 50% of Opus 5.5's $4 / $20 list rate, and from Dec 1 the two are level.
Only DeepSeek shipped usable open weights; MiniMax's are a promise
DeepSeek V4.1-Flash-1017 arrived on Oct 17 with MIT weights, and its Search and Fetch APIs cost $0.75 per 1,000 calls. MiniMax M3.1 Flash went live on Oct 19 at $0.15 / $0.60, but its weights, announced for Oct 23, are unconfirmed and the licence is unpublished. Open-weight no longer means permissive: Kimi K3 and Muse Spark Open carry $20M revenue triggers.
Anthropic's September models still top the independent boards
On the Artificial Analysis (AA) Intelligence Index v4.3.2, Opus 5.5 scores 57.6, Sonnet 5.5 56.0 and Muse Wonder Full 55.8 (board of Oct 22, independent). Argon leads LMArena text by 1525.4 − 1507.1 = 18.3 Elo, which is within noise. Between the two window models, DeepSeek leads on AA, 45 against 43, but MiniMax costs less per task, $0.22 against $0.26.
The rate card is a poor guide; cost per task and peak hours decide the bill
Kingy's check found no Western list-price change among 17 tracked models since Sept 20 (reported, Oct 19). DeepSeek charges $0.30 / $1.20 at peak and $0.15 / $0.60 off-peak. Gemini 3.8 Flash costs about $1.24 per index task, 5.6 times MiniMax's $0.22 for one point less.
Developer feedback is about trust
AA found MiniMax M3.1 Flash correct on 56.2% of questions, but 92.8% of its misses were confident errors, or 40.6% of all questions. AI/ML API's own tests passed 49 of 51 runs at 109 tokens/s, against MiniMax's "above 200" (vendor claim). Evidence on DeepSeek is a three-task test; all of this is anecdotal except the AA numbers.
Models at a glance
| Model | Vendor | Type | Release date | Status (Oct 24) | Source |
|---|---|---|---|---|---|
| MiniMax M3.1 Flash Window | MiniMax | Open-weight (announced); licence unpublished | Oct 19, 2026 | Shipping; weights promised Oct 23, unconfirmed | MiniMax |
| DeepSeek V4.1-Flash-1017 Window | DeepSeek | Open-weight, MIT | Oct 17, 2026 | Shipping; weights public | The Decoder |
| Gemini 4 Argon Context | Closed | Announced Sep 30; API Oct 13 | GA, paid tier only; price doubles Dec 1 | Towards Data Science; AionX | |
| Meta Muse Wonder Context | Meta | Closed, no public weights | Oct 12, 2026 | Shipping; Wonder Full in Muse Code plans only | DevX |
| Claude Opus 5.5 Context | Anthropic | Closed | Sep 22, 2026 | GA | Anthropic |
| GPT-6 Astra Context | OpenAI | Closed | Sep 3, 2026 | GA | BenchLM |
| Meta Muse Spark Open Context | Meta | Open-weight, custom licence | Oct 15, 2026 | Shipping | CometAPI |
| Reflection Beam 1.0 Context | Reflection | Open-weight, Apache 2.0 | Weights Oct 16, 2026 | Shipping | Data Studios |
| Kimi K3 Context | Moonshot | Open-weight, custom licence | Jul 16, 2026 | Shipping | Fello AI |
How to read the labels
No closed model launched in the window; Google's Argon price schedule is the closed-model news
All closed models below are context (released before Oct 17), judged against Oct 24, 2026. No vendor discloses a parameter count for any of them, so size is not disclosed. Price and context cells for Google, OpenAI, xAI and Meta come from aggregators and are labelled that way.
What the closed vendors did
Announced Sep 30, first only to vetted cyber defenders, and opened on Oct 13 through the paid Gemini API, AI Studio and Antigravity, plus the Gemini app for paid subscribers. One source still described it as gated in a page updated Oct 19; we follow the five pages that date general availability to Oct 13. There is no free tier, and core Workspace plans do not include it by default.
Unified window 1,048,576 tokens; default output 65,536, up to 1,048,576 with long decode; text, image, audio, video and PDF in, text out.
Released Oct 12 with Wonder Full. Only Wonder is on the API at $4 / $20 per 1M tokens (cached $0.20), the same as Opus 5.5; about 1M context. Wonder Full sits behind the $50 and $200 Muse Code plans with no per-token price.
The schedule was already public on Oct 13, so the Oct 19 revision is not a new change. Cache $0.10 becomes $0.20.
Opus 5.5 (Sep 22), Sonnet 5.5 (Sep 28) and Haiku 5.5 (Oct 7), all with a 1M-token window. Opus 5.5 at $4 / $20 is (5 − 4) ÷ 5 = 20% below Opus 5 at $5 / $25, and Haiku 5.5's $0.10 input is 90% below Haiku 4.5 (our arithmetic). Anthropic states different figures, "40% less to run" for Opus 5.5 and about 75% cheaper for Haiku 5.5 Vendor claim; we cannot reproduce either.
Luna $0.10 / $0.50, Sol $2 / $10, Astra $10 / $50 per 1M tokens; Astra is 100× Luna on both rates. Sol and Luna reached the API on Sep 22; the ChatGPT-wide rollout was dated Oct 7. Context shows as 1.05M in two sources and 922K in one, probably the input ceiling (our inference).
Reached the API on Sep 21 at an unchanged $2 / $6; in the Grok app's Expert mode since Oct 3. Context is not stated in the material.
Anthropic's release notes say Sep 1 and its news listing says Sep 15. We follow Sep 1, the dated changelog entry.
For buyers the closed tier is stable. The decision this week is Argon's Dec 1 doubling, not a new model.
Covered in the Google and Meta tab: every Argon rate doubles on Dec 1, and the schedule was already public on Oct 13.
QCode's calendar, last checked Oct 10, lists GitHub Copilot dropping six models on Oct 19, Google's earliest shutdown date of Oct 22 for several Veo 3.1 previews and gemini-omni-flash-preview, and OpenAI shutting down 17 older entries on Oct 23, none of them a GPT-6 model.
No closed model in the material was first released between Oct 17 and Oct 23. The newest are Muse Wonder (Oct 12) and Gemini 4 Argon's general availability (Oct 13).
Anthropic's announcements listing runs to Oct 21 and shows no model after Haiku 5.5 on Oct 7. We found no newsroom page from OpenAI, Google, xAI or Meta in the material.
Closed models at a glance
| Model | Key figures (USD per 1M tokens, in / out) | Date | Status (Oct 24) | Source |
|---|---|---|---|---|
| Google Gemini 4 Argon | Unified window 1,048,576 tokens; output 65,536 default; $2 / $10 until Nov 30, $4 / $20 from Dec 1 | Announced Sep 30; API Oct 13 | GA (Stable); paid tier only | LLM Data Hub quoting Google; Towards Data Science |
| Meta Muse Wonder | About 1M context; API $4 / $20, cached $0.20; no public weights | Oct 12 | Shipping on Meta Model API; Wonder Full only in Muse Code plans ($50, $200), no API price | DevX |
| Claude Haiku 5.5 | 1M context, 128K output; $0.10 / $0.50 (prompts up to 100K), $0.50 / $2.50 above | Oct 7 | GA | Anthropic; BenchLM; Fello AI |
| Claude Sonnet 5.5 | 1M context; $2 / $10 | Sep 28 | GA | Anthropic; BenchLM |
| Claude Opus 5.5 | 1M context, 128K output; $4 / $20; cache read $0.20 | Sep 22 | GA | Anthropic; BenchLM; vakati-tech |
| Claude Fable 5.1 | 1M context; $10 / $50; Mythos 5.1 is the same model with fewer safeguards, vetted users only | Sep 1 or Sep 15 (sources conflict) | GA; Mythos limited access | Anthropic; BenchLM |
| OpenAI GPT-6.1 Sol | 1.05M context; $2 / $10; cached $0.10 | Sep 29 | GA | BenchLM; Puter |
| OpenAI GPT-6 Sol / Luna | 1.05M context; Sol $2 / $10; Luna $0.10 / $0.50 up to 272K, $0.20 / $0.75 above | API Sep 22; ChatGPT rollout Oct 7 | GA | LLM Gateway; Puter; Shattered |
| OpenAI GPT-6 Astra | 1.05M context; $10 / $50; cached $1 | Sep 3 | GA | BenchLM; CloudZero |
| xAI Grok 4.7 | $2 / $6; context not stated in the material | Sep 21 | GA on xAI API | Fello AI; Activepieces |
All rows are context (released before the window). Parameter counts are not disclosed by any of these vendors.
Only DeepSeek shipped usable open weights in the window; MiniMax's are a promise dated Oct 23
Two models count as window releases on the open-weight side. Muse Spark Open, Reflection Beam 1.0 and Kimi K3 shipped earlier and are context baselines.
The two window models
1M context; takes text, image and video, returns text; allows 524,288 output tokens and always reasons, at five effort levels. Access is the MiniMax API, M Plan, MiniMax Code and OpenRouter. Size is not disclosed: some pages say 428B, but that is M3's size. MiniMax says the anonymous "Space Bunny" model on OpenRouter, which processed over 200 trillion tokens, was an early version.
Vals AI lists $0.60 / $2.40 and Oct 20; we follow MiniMax's post and OpenRouter, Oct 19 and $0.15 / $0.60. MiniMax's post says weights come "later" while its X post, relayed by others, says Oct 23. The material shows no weights or licence, so the status is announced.
OpenRouter's free window ended Oct 22 at 12:00 UTC, so list price is in effect. MiniMax's "70% lower agent cost" is a vendor claim about cost per task, not per token. Artificial Analysis found a 93% hallucination rate (module 04).
A "minor update" to the Sept 10 V4.1-Flash, with MIT weights on Hugging Face and ModelScope. DeepSeek's pages, read by CodePick on Oct 18, give 1M context and 384K max output. They disagree on size: a 552B backbone with 8B active on input in the Sept 10 post, versus 763B on the Hugging Face cards, unexplained.
DeepSWE v1.1 74.2 → 77.0 and Terminal-Bench 4.0 31.2 → 42.3 against the Sept build, with no independent confirmation yet.
The headline $0.15 / $0.60 is the off-peak rate. MiniMax's flat price equals DeepSeek's off-peak and is half its peak. Token prices were unchanged by the Oct 17 update.
$0.75 per 1,000 calls each, capped at 5 QPS per account. Server-side calls through DeepSeek Harness, Claude Code and Codex are free until Oct 31 at 16:00 UTC. AI Weekly says "November 1", the same moment, since Nov 1 00:00 Beijing time is Oct 31 16:00 UTC.
An agent making 1,000 searches and 1,000 fetches adds 2 × $0.75 = $1.50 before token costs. 20 searches per task cost 20 × $0.00075 = $0.015, about 8% of DeepSeek's chart figure of under $0.20 per DeepSWE task.
Cap search spend before Oct 31, when free server-side calls end. DeepSeek is the only window model you can download today, but price a daytime Beijing workload at the peak rate.
Open-weight and Chinese models at a glance
| Item | Key figures | Date | Status (Oct 24) | Source |
|---|---|---|---|---|
| MiniMax M3.1 Flash Window | 1M context; size not disclosed; $0.15 / $0.60 per 1M in / out | Oct 19 | Shipping; weights announced for Oct 23, unconfirmed | MiniMax |
| DeepSeek V4.1-Flash-1017 Window | 1M context; 552B or 763B (sources differ); $0.15 / $0.60 off-peak, $0.30 / $1.20 peak; MIT | Oct 17 | Shipping; weights public | The Decoder; CodePick |
| DeepSeek Search and Fetch Window | $0.75 per 1,000 calls | Oct 17 | Live | The Decoder |
| Meta Muse Spark Open Context | 816B / 42B active; 1M context; hosted $1.25 / $4.25; 256 experts, 8 routed per token | Oct 15 | Shipping; Together's free promotion ended Oct 18 | CometAPI; TechPilot |
| Reflection Beam 1.0 Context | 501.3B / 23.4B active (4.67% active); 1M; Apache 2.0 | Oct 16 (weights) | Shipping; own API is a free beta with no official price | Data Studios |
| Kimi K3 Context | 2.8T / 104B active; 1M; $3 / $15; download 1.56 TB | Jul 16 (weights Jul 27) | Shipping; no new release in the material | Fello AI |
Reflection's "3 to 4 times less compute than GLM-5.2" is a vendor claim that counts only generation compute. Self-hosting Muse Spark Open needs about 530 GB even in NVFP4.
Licence and access risk
Open-weight no longer means permissive. DeepSeek is MIT and Beam is Apache 2.0, but several models carry custom licences with revenue triggers. Internal use is exempt, so the triggers bite hosting businesses.
| Model | Licence | Trigger |
|---|---|---|
| DeepSeek V4.1-Flash-1017 | MIT | None |
| Reflection Beam 1.0 | Apache 2.0 | None |
| Kimi K3 | Custom | $20M model-as-a-service threshold |
| Meta Muse Spark Open | Muse Model License | Separate agreement above $20M revenue for model-as-a-service; "Built with Muse Spark Open" credit |
| Qwen3.8 | Custom | $50M revenue |
| MiniMax M3.1 Flash | Unpublished | Not known |
Bars federal civilian agencies from buying DeepSeek products; notice published Oct 20 per PYMNTS. It names no other lab.
A draft list dated Oct 16 covers large models; weights already released are reportedly exempt.
The Decoder separately reports Beijing may tighten rules on publishing new open-weight models.
No window model tops a headline board, and vendor tables are claims to reproduce
This module separates independent results from vendor claims. Snapshots are dated, and scaffolds, effort settings and runs differ between boards, so compare only like with like.
Artificial Analysis Intelligence Index
All six are context models. Muse Wonder Full is 57.6 − 55.8 = 1.8 points behind Opus 5.5 but limited to Muse Code subscribers on $50 and $200 plans. Argon is 5.0 points back, though AA ran it only at "high" and Opus 5.5 at "max", so the gap is not a controlled comparison.
Opus 5.5 57.6 vs Opus 5 at 50.8 and Opus 4.8 at 41.8; Sonnet 5.5 56.0 vs Sonnet 5 at 38.2 (+17.8); GPT-6 Astra 52.7 vs GPT-5.6 Sol at 47.0.
Source: Artificial Analysis via AICoder (board updated 2026-10-22); Gemini 4 Argon 52.6 from Fello AI, page of Oct 13. Independent measurement.
Window models against peers
| AA measure (as of Oct 21) | Intelligence Index | Cost per index task | Humanity's Last Exam | Hallucination rate (lower is better) |
|---|---|---|---|---|
| MiniMax M3.1 Flash Window | 43 | $0.22 | 48.0% | 92.8% |
| DeepSeek V4.1 Flash 1017 Window | 45 | $0.26 | 43.3% | 66.8% |
| GLM-5.3 Flash Context | 42 | $0.28 | 39.9% | 27.6% |
| Gemini 3.8 Flash Context | 41 | $1.24 | 48% | n/a |
| MiniMax M3 Context | 29 | $0.51 | 39.0% | 18.4% |
Source: Artificial Analysis via aimlapi, as of Oct 21, 2026 · independent measurement. AA's headline score for MiniMax is 42, up 13 from M3; OpenRouter's listing shows 42.7, which this table rounds to 43. Cost per task fell (0.51 − 0.22) ÷ 0.51 = 57% from M3.
Vendor claims against independent results
| Measure | Vendor figure | Independent figure | Gap |
|---|---|---|---|
| MiniMax output speed | Above 200 tokens/s Vendor claim | 93 tokens/s (AA) IndependentAI/ML API's own test: 109 | About half the claim |
| MiniMax, Terminal-Bench 4.0 | 38.1 Vendor claim | 25.8 (AA) Independent | 12.3 points |
| DeepSeek, Terminal-Bench 4.0 | 42.3 (own card) Vendor claim | 33.8 (AA) Independent | 8.5 points |
| DeepSeek, DeepSWE v1.1 | 77.0, up from 74.2 Vendor claim | No independent confirmation yet | Not available |
Harness differences explain part of the gaps, so vendor tables are claims, not rankings. MiniMax's self-run DeepSWE of 73.7 trails DeepSeek's 77.0. On Vals Index v2.2, MiniMax scores 53.78%, 15th of 22.
Other boards and coding evidence
| Board | Leaders | Basis and date |
|---|---|---|
| LMArena text | #1 Gemini 4 Argon 1525.4; #2 Opus 5.5 1507.1 | Human-vote Elo · mirror updated 2026-10-22. The 18.3 Elo gap is within noise (a guide treats shifts under 25 as noise). |
| LMArena Code (WebDev) | #1 Opus 5.5 1813; #2 GPT-6 Astra 1786.1 | Human-vote Elo · fetched 2026-10-12 |
| Vals Index | Argon 68.90%; Sonnet 5.5 67.04%; Opus 5.5 66.97% | Date not stated. Another Vals board has Sonnet 5.5 first at 70.8%; we do not mix the two. |
| Source conflicts | Mythos 5 at 1531; Opus 4.6 (Fast) first at 1500 | Two other tables disagree with the LMArena text order; we follow the two sources that agree Argon leads text. |
| Model | Mercor extended pass@1 | Original-set score | Change (points) |
|---|---|---|---|
| DeepSeek V4.1-Flash | 82.7% | 82.9% | −0.2 |
| Claude Opus 5 | 82.0% | 93.5% | −11.5 |
| Kimi K3 | 78.7% | 83.3% | −4.6 |
| Claude Opus 5.5 | 66.0% | 98.2% | −32.2 |
| Claude Sonnet 5.5 | 57.3% | 96.7% | −39.4 |
| GPT-6.1 Sol | 45.3% | 87.3% | −42.0 |
Mercor via BenchLM, Oct 19, 2026 capture, 150 samples per row, display-only. Extended-set margins are ±9 to 13 points per row, so the top three (83.3%, 82.7%, 82.0%) are effectively tied. Whether the 32 to 42 point falls on the newest rows reflect saturation of the original set is not established.
| Run | Score | Cost per attempt | Cost per solved task |
|---|---|---|---|
| GPT-6 Astra, extra-high | 74.1% (±2.9 pp) | $4.43 | $4.43 ÷ 0.741 = $5.98 |
| Gemini 3.8 Flash, high | 73.8% (±1.4 pp) | $2.36 | $2.36 ÷ 0.738 = $3.20 |
| GPT-6 Astra, low | 67.0% | $1.60 | Top setting buys 7.1 points for 2.8× the cost |
DeepSWE v1.1 (Datacurve, mini-swe-agent, 113 tasks, results of 2026-09-22) is the cleanest like-for-like view; the top two scores are tied, so cost decides. Scale Labs' SWE-Bench Pro V2 HARD board (Oct 19 snapshot) mixes model, scaffold and effort, so no generation claim is safe.
| Model | AA Intelligence Index | Note |
|---|---|---|
| GLM-5.3 | 45 | Context |
| DeepSeek V4.1 Flash 1017 Window | 45 | Oct 17 release |
| Kimi K3 | 44 | Context |
| Meta Muse Spark Open | 39 | Reproduced by Meta's card; VentureBeat found no independent listing at launch |
| Reflection Beam 1.0 | 30 | Vendor-reported 81.6% on SWE-bench Verified contrasts with a DeepSWE v1.1 of 44.4 against 74.2 for DeepSeek's September Flash |
One AA table of Oct 19, 2026 · independent measurement. Rank on AA index plus cost per task, check coding on boards with matched scaffolds, and treat every vendor table as a claim to reproduce.
The rate card is a poor guide: cost per task, peak hours and a dated Dec 1 doubling decide the bill
Between Oct 17 and Oct 23 prices moved in three places: DeepSeek's Search and Fetch fees, MiniMax's cheap M3.1 Flash, and a Google price page, revised Oct 19, that lists Gemini 4 Argon doubling on Dec 1 (the schedule was already public on Oct 13). OpenAI and Anthropic list prices show no dated change in the window.
Argon month explorer
Illustrative Assumed usage: 60M input and 15M output tokens a month. Prices reported by AionX; batch and priority rows are secondary listings.
DeepSeek and MiniMax rate explorer
Reported CodePick's re-check of DeepSeek's pages, Oct 18; MiniMax release post and OpenRouter, Oct 19. Token prices did not change with the Oct 17 update.
Cost per Artificial Analysis index task
Gemini 3.8 Flash costs $1.24 ÷ $0.22 ≈ 5.6 times more per task than MiniMax M3.1 Flash, for one point less on the index (41 vs 42 in this table).
Haiku 5.5 and GPT-6 Luna both list $0.10 / $0.50, yet an index task costs $0.21 on Haiku and $0.07 on Luna, three times as much, because Haiku writes far more output. Muse Wonder costs $4.52 per task for 52.4 against $0.72 for GPT-6.1 Sol at 51.8, about 6.3× for 0.6 more points.
Source: Artificial Analysis via VentureBeat (index shown in brackets; MiniMax M3.1 Flash is 42 here and 43 in the AI/ML API table, as 42.7 rounded) · independent measurement · as of Oct 21, 2026 · approximate values.
Price and context at a glance
| Item | USD per 1M tokens (in / out) and context | Date | Status (Oct 24) | Source |
|---|---|---|---|---|
| DeepSeek V4.1-Flash-1017, 1M context | Peak $0.30 / $1.20; off-peak $0.15 / $0.60 | Oct 17 (window) | In effect; token prices unchanged | CodePick, reading DeepSeek pages |
| DeepSeek Search and Fetch APIs | $0.75 per 1,000 calls, each | Oct 17 (window) | In effect; free in Harness, Claude Code, Codex to Oct 31 16:00 UTC | The Decoder |
| MiniMax M3.1 Flash, 1M context | $0.15 / $0.60; cached $0.015 | Oct 19 (window) | Shipping | MiniMax release post; OpenRouter |
| Google Gemini 4 Argon, 1,048,576 unified | $2 / $10 intro; $4 / $20 from Dec 1 | GA Oct 13 (context); price page revised Oct 19, schedule public Oct 13 | Intro in effect; rise announced | LLM Data Hub quoting Google; AionX; Shattered |
| OpenAI GPT-6 Astra / 6.1 Sol / Luna, 1.05M each | $10 / $50; $2 / $10; $0.10 / $0.50 | Astra Sep 3; 6.1 Sol Sep 29 | Context; in effect | BenchLM |
| Anthropic Fable 5.1 / Opus 5.5 / Sonnet 5.5 / Haiku 5.5, 1M each | $10 / $50; $4 / $20; $2 / $10; $0.10 / $0.50 | Sep 1 / Sep 22 / Sep 28 / Oct 7 | Context; in effect | Fello AI |
| Meta Muse Wonder | $4 / $20; cached $0.20 | Oct 12 | Context; in effect | DevX |
| Meta Muse Spark Open, hosted, 1M | $1.25 / $4.25; cached $0.15 | Oct 15 | Context; in effect | eesel AI |
1M tokens for all four Claude models and 1.05M for the GPT-6 tiers. Argon shares one pool of 1,048,576 tokens across input, thinking and output; its long-decode ceiling is 1,048,576 ÷ 65,536 = 16× the earlier 64K output limit.
Haiku 5.5 charges $0.50 / $2.50 past 100,000 prompt tokens, five times its base rate. GPT-6 Astra reprices the whole request past 272K input tokens (CloudZero, Sept 4 update): a 300K-token prompt costs 0.3 × $20 = $6.00 for input, not $3.00. Argon and Muse Wonder list no tiers.
Kingy's Oct 19 check found no list-price change among 17 tracked models since Sept 20. The small-model median fell $0.53 → $0.26 (−50.0%) from earlier replacements, not cuts. OpenAI's older GPT-5.6 Sol sits at a promotional $4 / $20 "through at least November 21" per BenchLM; sources conflict.
The week's sharpest complaint was trust, not speed or price
All feedback here is anecdotal (small tests, blog posts, comment threads) unless marked independent measurement, and none of it replaces module 04's benchmark evidence.
Right more often, wrong with confidence
AA's AA-Omniscience "hallucination rate" is the share of non-correct responses that were confidently wrong rather than partial or refused. With 56.2% correct, 43.8% were not, and 0.928 × 43.8 = 40.6% of all questions got a confident wrong answer.
M3.1 Flash gets about 13 points more questions right than DeepSeek but is wrong on a similar share (40.6% vs 38.1%), and on twice GLM-5.3 Flash's share (20.0%), because it attempts nearly everything. AA's overall AA-Omniscience score is 16.
Source: Artificial Analysis via AI/ML API, as of Oct 21, 2026 · independent measurement. The wrong and abstained shares are our arithmetic: abstained or partial = 100 − correct − wrong. Values: MiniMax M3.1 Flash 56.2 / 40.6 / 3.2; DeepSeek V4.1 Flash 1017 42.9 / 38.1 / 19.0; GLM-5.3 Flash 27.5 / 20.0 / 52.5; MiniMax M3 (predecessor) 16.7 / 15.3 / 68.0.
Feedback at a glance
| Model and signal | Key figure | Date | Status | Source |
|---|---|---|---|---|
| MiniMax M3.1 Flash WindowFactual reliability | 56.2% correct, but 92.8% of misses were wrong answers, not abstentions | As of Oct 21 | Independent measurement | AI/ML API; VentureBeat |
| MiniMax M3.1 FlashHands-on run | 49 of 51 runs passed; 109 tokens/s vs MiniMax's "above 200"; the model "often invented facts" | Oct 21 | Anecdotal test | AI/ML API |
| MiniMax M3.1 FlashDeveloper mood | "Split": "benchmaxxing" and context drift vs preferred to rival Flash models; no counts | Oct 21 | Anecdotal | kie.ai |
| DeepSeek V4.1-Flash-1017 WindowThree-task test | Accuracy 3/3, level with GPT-5.4-mini; median latency 2.5 s vs 0.7 s | Not stated | Anecdotal; build not stated | Activepieces |
| Gemini 4 Argon ContextWriting test | 180/190 rule checks vs 190/190 for Gemini 3.8 Flash; 140 runs | After Oct 13 | Anecdotal | Puter |
| Gemini 4 Argon ContextComment threads | 518 comments; 29% positive, 55% mixed, 16% skeptical | c. Oct 11 | Aggregator summary | daily.dev |
| Meta Muse Spark Open ContextLaunch thread | 914 Hacker News points, 477 comments on day one | Oct 15 | Reported | eesel AI |
Model notes
AI/ML API measured 109 tokens/s against AA's 93 and MiniMax's "above 200"; another write-up puts AA's figure near 105. Either is about half the vendor claim. A developer reading the docs found that disabling thinking returns an HTTP 400 error, so every request pays for reasoning.
OpenCode says the "Space Bunny" stealth run ranked first for 12 consecutive days and served 106T tokens; a routing marketplace reported 56.7T. That totals 106 + 56.7 = 162.7T, short of MiniMax's "over 200T". Space Bunny was only an early build. Miraflow (blog opinion) advises checkable, high-volume agent steps only.
DeepSeek shipped the checkpoint on Oct 17 behind the unchanged deepseek-flash alias, so API users needed no code change. The worry is process: according to AI Weekly, DeepSeek ships updates "as a config change" without formal deprecation windows. Shattered says secondary reports described an outage of about five hours before the release; DeepSeek's changelog does not confirm it. OpenRouter shows 99.56% availability over 24 hours (date not stated).
DeepSeek's own claim is DeepSWE v1.1 of 77.0 at just under 20 cents per task, against about 50 cents before; independent validation was not available. Activepieces compared it with GPT-5.4-mini on three tasks: equal accuracy (3/3), median latency 2.5 s vs 0.7 s (≈ 3.6×) and cost per task $0.000358 vs $0.000127 (≈ 2.8×). Simple tasks cannot confirm or refute the 20-cent claim.
Puter's 140-run test found it followed style rules but missed length targets: 1 of 5 articles hit 550–650 words (Gemini 3.8 Flash: 5 of 5) and 9 of 15 replies stayed under 80 words (15 of 15), though Argon was faster on the article (median 16 s vs 85 s). A one-day reviewer rated it best for large inputs and numbers-heavy work, with Claude Opus 5.5 still the coding default. NeuralTrust: 82 short questions cost $0.220 on Argon vs $0.069 on Opus 5.5 (≈ 3.2×), yet a robot-arm task cost $0.72 vs $1.09.
GPT Proto, a reseller of both, recommends Opus for production coding agents and Kimi for cheap, testable first drafts (opinion). One Hacker News commenter read Muse Spark Open's flat hosted pricing as a side effect of separate licensing.
Read the week by reader, watch the dated items, and treat promises and drafts as risks
This module turns the earlier ones into implications by reader and a dated timetable. It uses only facts from modules 01–06 and is not investment or procurement advice.
What changed, and what it means by reader
Builders (agents)
DeepSeek V4.1-Flash-1017 shipped on Oct 17 with MIT weights; Search and Fetch cost $0.75 per 1,000 calls, free in Harness, Claude Code and Codex until Oct 31 at 16:00 UTC.
Cap search spend before Oct 31; 1,000 searches plus 1,000 fetches add 2 × $0.75 = $1.50 before token costs.
Builders (self-hosting)
MiniMax M3.1 Flash is live, but weights and licence are not confirmed; Kimi K3 and Muse Spark Open carry $20M revenue triggers.
Treat DeepSeek as the only window model you can download today; read licences before building a hosted product.
Builders (quality)
AA measured a 92.8% hallucination rate on M3.1 Flash; hands-on tests passed 49 of 51 runs but the model "often invented facts".
Use cheap Flash models for checkable first passes and pilot on your own workload.
Buyers (Google)
Argon is GA at $2 / $10 until Nov 30, then $4 / $20.
Budget $540, not $270, for the illustrative 60M / 15M month from Dec 1.
Buyers (all vendors)
Western list prices did not change in the window; cost per task varies more than rate cards.
Compare cost per task: $0.22 for M3.1 Flash against $1.24 for Gemini 3.8 Flash.
Investors
The window's launches came from Chinese labs at $0.15–$0.30 input; Anthropic still leads the independent index at 57.6.
Watch whether price pressure reaches Western list prices; none has so far. This is a description, not investment advice.
Watch-list for the coming week
The Oct 23 date has passed, and the material shows no weights or licence. Until they appear, the model is "announced" as open-weight.
Free server-side calls end Oct 31 at 16:00 UTC, the same instant as Nov 1 00:00 Beijing time.
DeepSeek's DeepSWE v1.1 score of 77.0 has none yet, and Muse Spark Open's AA score of 39 comes from Meta's card with no independent listing at launch.
The Dec 1 rise is 38 days away (Oct 24 to Dec 1), so migrations to Batch at the new rates ($2 / $10) should be planned now.
The material has no Qwen or Kimi release in the window and nothing on Fibo-2.0.
Key dates
| Date | Event | State on Oct 24, 2026 | Source |
|---|---|---|---|
| Aug 16, 2026 | DeepSeek peak and off-peak price split starts | In effect | CodePick |
| Oct 12, 2026 | Meta Muse Wonder and Wonder Full released | In effect | DevX |
| Oct 13, 2026 | Gemini 4 Argon opens through the paid Gemini API | In effect | Towards Data Science |
| Oct 17, 2026 | DeepSeek V4.1-Flash-1017 weights; Search and Fetch fees start | In effect | The Decoder |
| Oct 19, 2026 | MiniMax M3.1 Flash goes live at $0.15 / $0.60 | In effect | MiniMax |
| Oct 22, 2026 | OpenRouter free window for M3.1 Flash ends at 12:00 UTC | In effect | OpenRouter |
| Oct 23, 2026 | MiniMax weights promised; OpenAI shutdown of 17 older entries scheduled | In effect as dates; completion not confirmed | Miraflow; QCode |
| Oct 31, 2026 | DeepSeek Search and Fetch free period ends, 16:00 UTC | Upcoming | The Decoder |
| Nov 21, 2026 | GPT-5.6 Sol promotional $4 / $20 "through at least" this date (sources conflict) | Upcoming | BenchLM |
| Nov 30, 2026 | Argon introductory $2 / $10 rate ends | Upcoming | AionX |
| Dec 1, 2026 | Argon rates double to $4 / $20 | Upcoming | AionX |
| Jan 1, 2027 | Gemini 3.8 Flash and 3.6 Flash double from $0.75 / $3.75 to $1.50 / $7.50 | Upcoming | AionX |