AI model releases September 2026: everything that shipped, including Claude Opus 5.5 and GPT-6 Sol

The short answer: in the first three weeks of September 2026, Anthropic, Google, Meta, OpenAI, DeepSeek, Moonshot, TypeSafe AI, xAI and Xiaomi all shipped something. The month ended with a double release on 22 September: Claude Opus 5.5 at $4 / $20 per 1M tokens, which an independent index now ranks first overall, and GPT-6 Sol ($2 / $10) and GPT-6 Luna ($0.10 / $0.50), which replace GPT-5.6 Sol and Luna at half the price. Earlier in the month the two $10 / $50 flagships arrived: Claude Fable 5.1 (1 September) and GPT-6 Astra (3 September). DeepSeek V4.1 Flash remains the cheapest open-weight option. In the last week of the month Claude Sonnet 5.5 (28 September) and GPT-6.1 Sol (29 September, at DevDay) shipped, and Google announced Gemini 4 Argon (30 September), which you cannot buy yet.

Every row below links to one official source: the lab's own announcement, docs or pricing page. We re-checked each source on 23 September 2026, and the late-September additions on 1 October 2026. If we could not find an official source for a release, we left it out. Benchmark scores are labelled with who measured them.

Correction, 23 September 2026: the first version of this page, published on 22 September, called GPT-6 Sol a rumour and pointed to DevDay on 29 September. OpenAI released GPT-6 Sol and GPT-6 Luna the same day. The sections below now treat both as shipped, with sources.

What shipped, 1 to 22 September 2026

Prices are per 1 million tokens, input / output, on the lab's own API, unless the row says otherwise.

What shipped, 23 to 30 September 2026 (added 1 October)

OpenAI also announced dots at DevDay on 29 September: always-on agents powered by GPT-6 Astra, rolling out to ChatGPT Pro users outside the EEA, Switzerland and the UK and to Business Premium users. They are a ChatGPT feature, not a model, so they are not in the table.

What each release means for a small business

Claude Fable 5.1 (1 September). Anthropic kept the rate card at $10 / $50, the same as Fable 5, and cut cached input by 75% to $0.25 per 1M. That only helps if you re-send the same long context repeatedly. It shipped the same day on the Claude API, AWS, Google Cloud and Azure. A sibling model, Claude Mythos 5.1, is gated behind verification programmes, so most buyers cannot get it.

Gemini 3.8 Flash (2 September). The launch price is introductory. Google has already published the date it doubles: 1 January 2027. If you are choosing a model to build on, budget for $1.50 / $7.50, not the launch rate. It is available in Google AI Studio, the Gemini API, Gemini Enterprise and the Gemini app on Google AI Pro and Ultra.

Muse Spark 1.3 (2 September). Meta's flagship is available in Muse Code and the Meta Model API. Meta says open weights are on its roadmap but gives no date.

GPT-6 Astra (3 September). OpenAI's new flagship has a 922,000-token input window and a 128,000-token output limit. It shipped to enterprises first, with the API and paid ChatGPT plans following over the next few days. Artificial Analysis (independent) says Astra ties Claude Fable 5.1 on its Intelligence Index at about 40% of the cost to run the index. Astra is worth its price when your bottleneck is a long chain of clicks and forms. It is not worth it for plain drafting.

GPT Image 2.5 (8 September). Two image models: Flare, which OpenAI's docs describe as "fast, high-quality everyday image generation," and Sunburst, the slower, more detailed option. For product shots and social posts, start with Flare.

DeepSeek V4.1 Flash (10 September). This became the new volume tier. It costs $0.15 / $0.60 per 1M off-peak, double at peak hours (01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays). Weights are on Hugging Face. The old deepseek-v4-flash id now routes to it, so switch your config to deepseek-flash rather than relying on the redirect. We covered it in detail in what changed with DeepSeek V4.1 Flash.

Kimi K2.8 Preview (11 September). Moonshot shipped this inside Kimi Code on the same kimi-for-coding model id, with a 1M-token context for every membership tier and adjustable thinking effort (low, high, max). It is labelled a preview, and Moonshot has not said when it becomes stable or whether pricing changes. If you already pay for Kimi Code, you have it. For Moonshot's cheaper flagship, see our Kimi K3 guide.

Jev (15 September). Jev is not a chatbot. It answers a typed question with a choice, a score or a yes/no, together with a probability. The price is very low, and output tokens are free. But it is early access, pricing may change at general availability, and it only matters if someone already runs automations for you. See what Jev is.

Grok 4.7 (21 September). It has the same price and the same 500K context as Grok 4.6, and scores better on coding. Independent tests also found it writes more than twice as many output tokens per task, so the bill per job rises even though the rate card does not. Full detail in Grok 4.7: what changed.

MiMo-V2.6 (22 September, China time). Xiaomi released two open-weight models, Pro and Flash, with weights on Hugging Face and API prices unchanged from V2.5. Xiaomi reports that MiMo-V2.6-Pro scores 46 on the Artificial Analysis Intelligence Index, the highest of any open-source model, while admitting a gap to Claude Fable 5.1 and GPT-6 Astra. That score is Xiaomi's claim about a third-party index. Check it on Artificial Analysis before relying on it.

Claude Opus 5.5 (22 September). Anthropic's successor to Opus 5 costs less per token ($4 / $20 against $5 / $25) and Anthropic reports it well ahead of Fable 5.1 on coding: 66.4% on Terminal-Bench 4.0 against 55.8%. Artificial Analysis (independent) scores it 58 on its Intelligence Index, first of 212 models, ahead of Fable 5.1 and GPT-6 Astra at 53 each. Anthropic now lists Opus 5, Fable 5 and Opus 4.8 as legacy models: still available, not retired. Sonnet 5.5 followed on 28 September; Haiku 5.5 followed on 7 October (see the update below). If you pay for Opus 5 or Fable 5.1 today, test Opus 5.5 first.

GPT-6 Sol and GPT-6 Luna (22 September). OpenAI released both on its API as gpt-6-sol and gpt-6-luna, and in ChatGPT Work and Codex for paid plans. They cost half of what GPT-5.6 Sol and Luna did. Both have a 1,050,000-token window, but a prompt over 272K tokens is billed at double the input rate and 1.5 times the output rate for the whole request. OpenAI reports Sol at 33.2% on AutomationBench, above Claude Opus 5, and Luna at 66.6% on DeepSWE v1.1. GPT-6 Astra stays OpenAI's flagship. Full comparison in Claude Opus 5.5 vs GPT-6 Sol vs GPT-6 Luna.

Claude Sonnet 5.5 (28 September). Same $2 / $10 as Sonnet 5, with a 1M-token window billed at the standard rate. Anthropic reports 70.6% on Terminal-Bench 4.0, ahead of Opus 5.5's 66.4% on that one test (Anthropic's own figures). Artificial Analysis (independent) scores it 56 but labels it very verbose, so its cost per task at max effort came out higher than Opus 5.5's. See Sonnet 5.5 vs Opus 5.5 vs GPT-6 Sol.

GPT-6.1 Sol (29 September). OpenAI's DevDay upgrade to GPT-6 Sol at the same $2 / $10, with cached input halved to $0.10 and the same 272K long-prompt rule. OpenAI reports it matching GPT-6 Astra on DeepSWE v1.1 at roughly one-fifth of the cost (OpenAI's own figures). Artificial Analysis (independent) scores it 52, up from GPT-6 Sol's 48, at about $0.72 per index task, which makes it the best frontier model per dollar on our model tracker. It is in the API, ChatGPT Work and Codex, not yet in regular ChatGPT chat.

Gemini 4 Argon (30 September, announced only). Google's new frontier model has a 1M-token output limit and Google reports 77.9% on DeepSWE v1.1 (Google-reported). Artificial Analysis (independent) scores it 53, behind Claude Opus 5.5's 58. But it is only rolling out to vetted cyber defenders; Google says paid API customers and Google AI Ultra subscribers come first when it opens, with no date. Do not plan around it yet. Which of these to use for coding is in our best AI model for coding guide.

Still unconfirmed

This section is for names people are searching that no lab has released. Nothing here goes on our model tracker until the lab publishes it.

When was GPT-6 Sol released? On 22 September 2026, together with GPT-6 Luna. It did not wait for OpenAI DevDay on 29 September 2026. Its upgrade, GPT-6.1 Sol, is the model that came at DevDay.

Do not confuse GPT-6 Sol with GPT-5.6 Sol. GPT-5.6 Sol became generally available on 9 July 2026 and is still sold, at a promotional $4 / $20 that OpenAI says runs at least until 21 November 2026. GPT-6 Sol is the newer model, at $2 / $10.

Retirements coming

GPT-5.5 leaves ChatGPT and Codex on 14 October 2026. OpenAI's pricing documentation says "GPT-5.5 retires from ChatGPT, ChatGPT Work, and Codex on all plans on October 14, 2026." It also says "The OpenAI API isn't affected." So if your team picks GPT-5.5 in the ChatGPT model menu or in Codex, move to a newer model before then. If you call GPT-5.5 through the API, nothing changes on that date.

A retirement that did not happen: DeepSeek V4 Pro. On 10 September, DeepSeek said all deepseek-v4-pro traffic would route to V4.1 Flash from 14 September. It then reversed that decision. V4 Pro is still listed on DeepSeek's pricing page with its own prices. If you switched away in a hurry, nothing broke.

After DevDay (29 September)

Added 1 October 2026, from OpenAI's DevDay 2026 recap.

• Shipped with a model id and a price: GPT-6.1 Sol, as gpt-6.1-sol at $2 / $10 per 1M tokens. See the section above.
• Announced and rolling out, no separate price: dots, always-on ChatGPT agents powered by GPT-6 Astra, for ChatGPT Pro users outside the EEA, Switzerland and the UK and for Business Premium users. The first dot is included in those plans. A new ChatGPT Pro 500 tier at $500 a month was also added.
• Previewed, not yet out: GPT-6.1 Sol Ultrafast, which OpenAI says brings up to 8x faster token generation in Codex "in the coming days".

What to do this month

1. Test Claude Opus 5.5 before paying flagship rates. It costs less than half of Fable 5.1 or Astra and ranks higher on an independent index. Keep Astra for work that means operating software.
2. Compare GPT-6 Luna and DeepSeek V4.1 Flash for volume work. Luna has the lower list price; V4.1 Flash off-peak is cheaper for prompts over 272K tokens, and its weights can run in-house.
3. Switch any gpt-5.6-sol or gpt-5.6-luna calls to gpt-6.1-sol or gpt-6-luna. The new ones cost half as much. Move gpt-6-sol calls to gpt-6.1-sol too, at the same price.
4. Budget Gemini 3.8 Flash at its 2027 price, not the launch price.
5. Check the ChatGPT model menu before 14 October if anyone on the team still uses GPT-5.5.

Keeping that choice current every month is what a Business AI Plan is for. If you want the right mix named for your own workflow, start with the free 2-minute AI plan finder.

Frequently asked questions

What AI models were released in September 2026? Between 1 and 22 September 2026: Claude Fable 5.1 (Anthropic, 1 September), Gemini 3.8 Flash (Google, 2 September), Muse Spark 1.3 (Meta, 2 September), GPT-6 Astra (OpenAI, 3 September), GPT Image 2.5 Flare and Sunburst (OpenAI, 8 September), DeepSeek V4.1 Flash (10 September), Kimi K2.8 Preview (Moonshot AI, 11 September), Jev (TypeSafe AI, 15 September), Grok 4.7 (xAI, 21 September), and on 22 September MiMo-V2.6 (Xiaomi), Claude Opus 5.5 (Anthropic), and GPT-6 Sol and GPT-6 Luna (OpenAI). At the end of the month came Claude Sonnet 5.5 (Anthropic, 28 September) and GPT-6.1 Sol (OpenAI, 29 September). Google announced Gemini 4 Argon on 30 September but has not released it to the public.

What is the GPT-6 Sol release date? 22 September 2026. OpenAI released GPT-6 Sol and GPT-6 Luna together, on its API and in ChatGPT Work and Codex for paid plans. GPT-6 Sol costs $2 input / $10 output per 1M tokens.

Is GPT-6 Sol the same as GPT-5.6 Sol? No. GPT-5.6 Sol has been generally available since 9 July 2026 and costs $4 / $20 at its current promotional rate. GPT-6 Sol is its successor at half that price.

Which new AI model is cheapest? Among the text models above with a published API price, GPT-6 Luna has the lowest list price at $0.10 input / $0.50 output per 1M tokens. For prompts over 272K tokens Luna costs $0.20 / $0.75, and there DeepSeek V4.1 Flash off-peak, at $0.15 / $0.60, is cheaper. Jev is cheaper per token, but it is a decision model, not a general text model, and its price is early-access pricing.

Is GPT-5.5 being shut down? Only in ChatGPT, ChatGPT Work and Codex, on 14 October 2026. OpenAI says the API is not affected.

Which September model is best for a small business? None of them is a reason to change the chat app you already use. On the API side, GPT-6 Luna or DeepSeek V4.1 Flash suits cheap bulk work, Claude Opus 5.5 suits coding and multi-step agent work, GPT-6.1 Sol is the best frontier model per dollar on our tracker, and GPT-6 Astra is worth its price only for operating software.

Is Gemini 4 Argon available? Not to the public. Google announced it on 30 September 2026 and is rolling it out only to vetted cyber defenders. Google says paid API customers and Google AI Ultra subscribers come first, with no date.

Sources: Anthropic — Claude Opus 5.5, Anthropic — models overview, OpenAI — Introducing GPT-6 Sol and Luna, OpenAI API pricing, Artificial Analysis — Claude Opus 5.5, Anthropic — Claude Fable 5.1, Google — Gemini 3.8 Flash, Meta — Muse Spark 1.3, OpenAI — GPT-6 Astra, OpenAI — GPT Image 2.5 Flare, DeepSeek pricing, DeepSeek V4.1 Flash announcement, Kimi Code — what's new, TypeSafe AI — Jev, xAI — Grok 4.7, Xiaomi — MiMo-V2.6, OpenAI — DevDay 2026, OpenAI — ChatGPT and Codex pricing, Artificial Analysis — Benchmarking GPT-6 Astra, Anthropic — Claude Sonnet 5.5, OpenAI — Introducing GPT-6.1 Sol, OpenAI — Introducing dots, OpenAI — DevDay 2026 recap, Google — Gemini 4 Argon, Artificial Analysis — GPT-6.1 Sol, Artificial Analysis — Gemini 4 Argon. Checked 1 October 2026.