Monthly roundup
September 2026 in AI models
A release-wave summary for readers who want the important changes without pretending every launch is a reason to switch.
The signal
21 releases plus 4 material status updates
Google, OpenAI, Anthropic, LongCat, Xiaomi, SpaceXAI, Qwen, StepFun, Moonshot AI, DeepSeek, Inception, InclusionAI, Institute of Foundation Models (MBZUAI), Meta account for 21 tracked release events. 23 catalogued events link to a detailed comparison profile. 11 launch guides are available for deeper trade-off coverage. There is no single winner because the constraints are different.
At a glance
- Release events
- 21
- Profile links
- 23
What shipped and changed
Limited previewGemini 4 Argon
Google
Google announces Gemini 4 Argon, initially restricted to trusted cyber defenders in Fairwind. Broader access is planned; no general API endpoint is confirmed.
Date note: Announced introductory API rates are $2/$10 per million tokens, then $4/$20 after the introductory period. These are future launch rates, not a currently generally purchasable API price. Google reports 77.9% on DeepSWE 1.1 using its own evaluation setup.
Google Argon announcement
ReleaseGPT-6.1 Sol
OpenAI
OpenAI releases GPT-6.1 Sol at the same $2/$10 Standard token rates as GPT-6 Sol, with 1.05M context, 128K output and lower cached-input pricing ($0.10).
Date note: Reasoning effort starts at low; none and minimal are unsupported. Above 272K input, the whole request uses the higher $4/$15 rate. Tool calling requires Responses rather than Chat Completions.
OpenAI API changelog
ReleaseClaude Sonnet 5.5
Anthropic
Anthropic releases Sonnet 5.5 at $2/$10 per million tokens, with 1M context and 128K standard output. Adaptive thinking is enabled by default.
Date note: Haiku 5.5 remains announced for the coming weeks. Vendor Terminal-Bench 4.0 (70.6%) and Artificial Analysis's independent mini-swe-agent run (63.6% at max effort) use different evaluation setups.
Anthropic Sonnet 5.5 announcement- Limited preview
LongCat-2.5-Preview
LongCat
Official API changelog announces coding and image-understanding preview availability.
Date note: Current $0.30/$1.20 USD per million and $0.006 cached input are limited-time direct rates without a confirmed end date.
LongCat official API changelog - Release
Claude Opus 5.5
Anthropic
Anthropic ships Claude Opus 5.5 at $4/$20 per MTok (cheaper than Opus 5's $5/$25) with a 1M-token context, and its docs now say to start new work with it. It leads the Artificial Analysis Intelligence Index v4.3.2 (58, publisher-rounded) and CursorBench 4.0 (57.8% at Max effort).
Date note: At this launch, Sonnet 5.5 and Haiku 5.5 were announced for the coming weeks. Sonnet 5.5 subsequently launched September 28; Haiku 5.5 is still announced as of October 1. Opus 5 moves to Active (legacy).
Anthropic — Claude Opus 5.5 - Release
GPT-6 Luna
OpenAI
OpenAI's new GPT-6 budget tier: GPT-6 Luna at $0.10/$0.50 per MTok (cached $0.01), with a 1.05M-token context. Its predecessor GPT-5.6 Luna lists $0.20/$1.20: input is halved and output falls by about 58%.
OpenAI — Introducing GPT-6 Sol and Luna - Release
GPT-6 Sol
OpenAI
OpenAI launches GPT-6 Sol at $2/$10 per MTok — half GPT-5.6 Sol's price — with a 1.05M-token context, in the API, ChatGPT Work and Codex the same day. OpenAI reports 68.8% on DeepSWE 1.1 at max effort.
OpenAI — Introducing GPT-6 Sol and Luna 
ReleaseXiaomi MiMo-V2.6-Pro + V2.6-Flash
Xiaomi
Xiaomi releases MiMo-V2.6-Pro (1.02T/42B MoE, 1M context, MIT weights) and V2.6-Flash with omni-modal input at unchanged API prices ($0.435/$0.87 and $0.14/$0.28 per MTok).
Xiaomi dated model release log
ReleaseGrok 4.7
SpaceXAI
xAI releases standard Grok 4.7 at $2/$6 per MTok and 500K context, with API, Cursor and Grok Build access. GitHub announces a gradual Copilot rollout on eligible tiers. The separate Fast option is restricted to Cursor and Grok Build.
xAI — Grok 4.7- Status update
Gemini 2.5 access restriction
Google
Google limits Gemini 2.5-series API access to existing users; new projects are pointed to Gemini 3.8 Flash or 3.5 Flash-Lite. No Gemini API shutdown date is published.
Gemini API changelog 
ReleaseQwen3.8-Omni-Flash
Qwen
Alibaba adds Qwen3.8-Omni-Flash to Model Studio: text, image, audio and video input with a 1M-token context at $0.15/$0.47 per MTok. A -realtime variant followed on 2026-09-21.
QwenCloud model changelog
Limited previewStep 5 Preview
StepFun
StepFun opens API access to Step 5 Preview ($1.00/$2.70 per MTok, 1M context) and promises open weights on 2026-10-15.
Date note: StepFun's launch page is undated; the date is Artificial Analysis's listing date.
StepFun platform pricing (step-5-preview)
Limited previewKimi K2.8 Preview
Moonshot AI
Moonshot rolls out Kimi K2.8 Preview in Kimi Code under the existing `kimi-for-coding` alias, with up to 1,048,576 context tokens; no pay-as-you-go token tariff is verified.
Date note: Preview product; Kimi Code membership/quota access is distinct from the K3 platform token tariff.
Kimi Code — what's new- Status update
ChatGPT Pro 20x paused for new sign-ups
OpenAI
OpenAI stopped new sign-ups to the $200 Pro tier on this date; the $100 tier stayed open. Pro 200 subsequently reopened with a lower included allowance for new subscribers (confirmed October 1).
OpenAI Help — About ChatGPT Pro tiers 
Status updateDeepSeek V4-Flash retired
DeepSeek
DeepSeek retires V4-Flash and V4-Flash-Vision-Exp. The legacy ids temporarily route to V4.1-Flash and bill at V4.1-Flash rates.
DeepSeek — V4.1-Flash release notes- Status update
DeepSeek V4-Pro phase-out announced
DeepSeek
DeepSeek says it is phasing out V4-Pro in favour of V4.1-Flash.
Date note: DeepSeek later reversed the planned 2026-09-14 re-route and still serves V4-Pro.
DeepSeek API changelog - Release
DeepSeek V4.1-Flash
DeepSeek
DeepSeek releases V4.1-Flash (MIT weights, 1M context, image input) as its new default at $0.30/$1.20 per MTok peak ($0.15/$0.60 off-peak) under the API id `deepseek-flash`, and says it beats V4-Pro on performance, cost and speed.
DeepSeek — V4.1-Flash release notes 
ReleaseInception Mercury 2.5
Inception
Inception releases the Mercury 2.5 diffusion LLM at $0.20/$0.75 per MTok list (80%-off launch promo) with a 260K context and 65,536-token output.
Date note: September 8 is the publisher’s exact Mercury 2.5 related-article date; the page extraction’s September 29 date belongs to Mercury Voice. Current discounted prices remain $0.04/$0.15, with no stated end date.
Inception Mercury 2.5 announcement and dated publisher card
Availability updateLing 3.0 Flash VL native-platform launch
InclusionAI
Date note: Native chat/API availability. This date does not establish initial downloadable-weight publication or a later third-party hosting date.
Ant Ling official dated platform changelog- Release
GPT-6 Astra
OpenAI
OpenAI launches GPT-6 Astra, its most capable model and the first gated at the Preparedness Framework's 'Critical' cybersecurity threshold — $10/$50 per MTok (≤272K input) with a 1.05M-token context. It debuted at 61.2 on the Artificial Analysis index (v4.1.1; 53, publisher-rounded, on the rebased v4.3.2) and reported 74.1% on DeepSWE 1.1, which Datacurve's board confirms at rank 1.
OpenAI — Introducing GPT-6 Astra - Release
K2-Horizon Fleet (0.9B to 375B)
Institute of Foundation Models (MBZUAI)
The Institute of Foundation Models releases six fully open (Apache 2.0) K2-Horizon models from 0.9B to 375B-A23B, with weights, code, training data and methodology. The 375B-A23B flagship has a 524,288-token context; MoVA attention applies to the 36B-A4B model only.
Date note: Artificial Analysis lists K2 Horizon 375B A23B at 30.5 on Intelligence Index v4.3.2; the earlier '47' tracker figure was not verified.
Institute of Foundation Models — K2-Horizon press release - Release
Gemini 3.8 Flash + Flash Cyber
Google
Google unveils Gemini 3.8 Flash workhorse with 89.4% on Terminal-Bench 2.1 (Google's table) and companion defensive cyber edition via the Fairwind Program.
Google DeepMind — Gemini 3.8 Flash announcement 
ReleaseMuse Spark 1.3
Meta
Meta Superintelligence Labs ships Muse Spark 1.3 with a self-reported 75.4% on DeepSWE 1.1 (max thinking mode) and a 1M-token context for autonomous agent workflows.
Meta Superintelligence Labs — Muse Spark 1.3 announcement- Availability update
Qwen3.8-Max-0902 snapshot
Qwen
Alibaba lists a dated qwen3.8-max-0902 snapshot on Model Studio; Artificial Analysis scores it 45.4 on Intelligence Index v4.3.2.
QwenCloud model changelog - Release
Claude Fable 5.1 + Mythos 5.1
Anthropic
Anthropic refreshes its flagship: at launch Fable 5.1 topped the Artificial Analysis Intelligence Index (66 on v4.1.1) at unchanged $10/$50 pricing, with cache reads cut to $0.25; restricted-access Mythos 5.1 arrives alongside.
Date note: Claude Opus 5.5 (2026-09-22) now leads the rebased AA index v4.3.2 (58 vs Fable 5.1's 53, publisher-rounded).
Anthropic — Claude Fable and Mythos 5.1
Read the important launches
OpenAI
GPT-6.1 Sol: the newer Sol release at the same base price
GPT-6.1 Sol is the newer Sol release at the same $2/$10 Standard short-context rates, with cached input at $0.10 per million tokens.
Anthropic
Claude Sonnet 5.5: everyday coding with breaking API changes
Sonnet 5.5 is released at Sonnet 5's $2/$10 token price, with 1M context and 128K standard output.
Gemini 4 Argon: restricted preview, broader API access pending
Google initially restricts Argon to trusted cyber defenders in Fairwind while preparing broader safeguards and access.
Anthropic
Claude Opus 5.5: Anthropic's recommended starting point for many workloads at a lower price than Opus 5
Claude Opus 5.5 replaces Opus 5 as Anthropic's recommended starting point and costs less per token ($4/$20 vs $5/$25). Published benchmark rows are dated, effort- and harness-specific snapshots, not a universal model ranking.
OpenAI
GPT-6 Sol and Luna: lower API rates than the GPT-5.6 tiers
OpenAI's September 22, 2026 release added Sol and Luna below Astra. Sol's $2/$10 rate is half GPT-5.6 Sol's $4/$20 promotional rate; Luna's $0.10/$0.50 is half its predecessor's input rate and about 58% lower on output ($0.50 vs $1.20).
xAI
Grok 4.7: a larger base model at Grok 4.6's price
Grok 4.7 is xAI's flagship since 2026-09-21 — a larger new base model with longer reinforcement learning on multi-hour tasks, sold at Grok 4.6's $2/$6 rate card.
DeepSeek
DeepSeek V4.1-Flash: the hosted successor to retired V4-Flash
DeepSeek's September release is an MIT-licensed multimodal model. DeepSeek reports comparisons against V4-Pro; those vendor results are not an independently matched ranking. The same announcement retired V4-Flash API aliases.
OpenAI
GPT-6 Astra: OpenAI's computer-use flagship arrives gated
OpenAI's first GPT-6-family model is a computer-use flagship — and the first model it gates at the Preparedness Framework's 'Critical' cybersecurity threshold.
Anthropic
Claude Fable 5.1 and Mythos 5.1: the September refresh
Fable 5.1's 2026-09-01 launch announcement reported a leading position on the Artificial Analysis Intelligence Index v4.1.1 at $10/$50 pricing, with a restricted-access Mythos 5.1 sibling for cyber and bio research. Later index snapshots use different versions and model rows; this guide does not claim a current top-ranked model.
Gemini 3.8 Flash: September 2026 launch and dated coding results
Google's September 2, 2026 launch positioned Gemini 3.8 Flash for coding and agentic work. In Datacurve's independent DeepSWE 1.1 snapshot dated 2026-09-22, its score was within 0.3 points of GPT-6 Astra at about half the reported task cost; this is a dated benchmark comparison, not a current cross-model rank.
Meta
Muse Spark 1.3: Meta's self-reported DeepSWE result
In its 2026-09-02 release, Meta reported 75.4% on DeepSWE 1.1 at $1.25/$4.25 per MTok; this is a vendor-reported result, not a universal leaderboard ranking. The same announcement says max reasoning is available on Muse Code and Meta Model API.