AI model release roundup: September and October 2026

admin
AI model release roundup: September and October 2026 cover graphic

Late September and October 2026 brought a dense run of AI model releases: new Claude 5.5 models, a long-awaited Gemini 4 flagship, an open-weight model from Meta, a cheaper OpenAI model for computer use and dozens of smaller launches. Keeping track is hard, and release trackers do not always agree on dates.

This roundup lists what we could find in dated sources, flags where the sources conflict and says what is a vendor claim and what is independently reported. It is a news summary, not a ranking. Benchmarks quoted by vendors are not verified by us, and we have not used these models for this article.

Key takeaways

  • Anthropic released Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 within roughly three weeks, ending with Haiku 5.5 on 7 October according to press reports.
  • Google announced Gemini 4 Argon on 30 September to a limited group, then widened access on 13 October through its API and subscriber products.
  • Meta released the weights for Muse Spark Open on 15 October under a custom licence with conditions, including attribution and a revenue-based agreement for some model-hosting companies.
  • OpenAI released GPT-6.1 Terra at the end of October, priced from $1 per million input tokens according to a news report quoting OpenAI.
  • Trackers disagree on some dates and details. Use official vendor pages to confirm anything you plan to depend on.

Summary

The releases fall into three groups. Frontier assistants from the three largest Western labs arrived in quick succession and were framed around coding, agents and computer use. Open-weight models kept appearing, with Meta’s Muse Spark Open the most prominent in our sources. And a long tail of specialist models covered voice, image, video and translation.

The common thread in the vendor messaging we read is cost and speed: several launches were described as faster or cheaper than predecessors. That matters for buyers because a cheaper model is only a better value if it does your task well enough. The table below gives the dated record, and the sections after it add detail.

Release timeline

Dates are as stated in the cited reports. Where a report gives only a day of the week, we say so.

Date (2026) Vendor and model What was reported
Week before 28 September Anthropic, Claude Opus 5.5 Launched “last week” per Reuters coverage; priced at $4 per million input and $20 per million output tokens
28 September Anthropic, Claude Sonnet 5.5 Described as a faster, lower-cost complement to Opus 5.5; $2 input and $10 output per million tokens, unchanged from Sonnet 5
30 September Google, Gemini 4 Argon Announced; provided first to select cybersecurity partners with no public timeline given at that point
7 October Anthropic, Claude Haiku 5.5 Priced at $0.10 input and $0.50 output per million tokens for prompts under 100,000 tokens
13 October Google, Gemini 4 Argon Access widened to the Gemini API, AI Studio and subscriber products
15 October Meta, Muse Spark Open Open weights released under the Muse Model License
26 to 27 October OpenAI, GPT-6.1 Terra Announced 26 October per news coverage, available from 27 October in ChatGPT Work and Codex and through the API

Anthropic: the Claude 5.5 family

Opus 5.5 and Sonnet 5.5

Reuters coverage republished on 29 September described Sonnet 5.5 as the second model in the Claude 5.5 family, released on a Monday that appears to be 28 September, and as a faster, lower-cost complement to Opus 5.5. Opus 5.5 had launched the previous week; the article gives no exact date. Pricing for Sonnet 5.5 was $2 per million input tokens and $10 per million output tokens, the same as Sonnet 5, and $4 and $20 for Opus 5.5.

Paul Thurrott’s report of 28 September adds that Anthropic says Sonnet 5.5 runs more than 30 percent faster than its predecessor and costs up to 30 percent less for most work because it typically needs fewer tokens for the same result. Those are Anthropic’s claims as relayed by the press. The model was reported as available on Amazon Web Services, Google Cloud and Microsoft Azure.

Haiku 5.5

Reuters coverage dated 8 October said Claude Haiku 5.5 launched on a Wednesday, which matches 7 October, as the third model in the family within about a month. It is priced at $0.10 per million input tokens and $0.50 per million output tokens for prompts under 100,000 tokens, and $0.50 and $2.50 for longer prompts. Anthropic is quoted as saying it costs 75 percent less than Haiku 4.5. The intended uses named are classification, summarisation and extraction, including live support and voice agents, and the report calls it the first Haiku with built-in safeguards for a narrow set of high-risk cybersecurity requests.

Model Input price per million tokens Output price per million tokens Positioning in the reports
Opus 5.5 $4 $20 Top model of the 5.5 family
Sonnet 5.5 $2 $10 Faster, lower-cost complement for everyday work
Haiku 5.5 (prompts under 100,000 tokens) $0.10 $0.50 High-volume tasks such as classification and extraction

Prices are as reported in September and October and may have changed.

Google: Gemini 4 Argon

Khaleej Times reported on 1 October that Google announced a new top-tier model called Argon, anchoring the Gemini 4 generation, on the previous day, 30 September. A spokesperson said it is larger than Google’s previous Pro models. At that stage it was provided to select cybersecurity partners, Google gave no public release timeline, and the report says Google no longer plans to release Gemini 3.5 Pro. The report also says Google’s performance claims are self-reported and that Argon trailed on two of four coding benchmarks Google included.

A report published on 13 October says Google then expanded access. It names the Gemini API, Google AI Studio and the Gemini platform for developers and enterprises, with the stable model ID gemini-4-argon, and says Google AI Ultra subscribers get access in Antigravity, while Ultra and Pro subscribers get access in the Gemini app and AI Mode in Search. The documented combined input and output context limit is 1,048,576 tokens. Introductory API pricing is $2 per million input tokens and $10 per million output tokens through 30 November, and the report says standard pricing will be higher afterwards without giving figures.

OpenAI: GPT-6.1 Terra

Interesting Engineering reported on 26 October that OpenAI released GPT-6.1 Terra, available from 27 October to Plus, Pro, Enterprise and Edu users through ChatGPT Work and Codex, and to developers through the API as gpt-6.1-terra. API prices are reported as $1 per million input tokens, $0.05 for cached input and $5 per million output tokens, which OpenAI is quoted as describing as 50 percent lower than for GPT-5.6 Terra and GPT-6.1 Sol.

The same report lists OpenAI’s own claims, including faster computer-use tasks and lower factual error rates at higher reasoning effort. We repeat none of the benchmark numbers here because they come from the vendor and we have not verified them. The report also states that OpenAI’s preparedness evaluations place the model at the Critical threshold for cybersecurity, which is a point to read in the vendor’s documentation if security matters to your use.

Open-weight and other releases

Meta Muse Spark Open

Unite.AI reported that Meta released Muse Spark Open weights on 15 October. The model is a mixture-of-experts design with 816 billion total and 42 billion active parameters and a context window of 1 million tokens, and it accepts text, image, video and audio input and produces text. The Muse Model License, dated 15 October, allows use, modification including fine-tuning, distillation and distribution subject to conditions. Licensees must display “Built with Muse Spark Open” in their product, and companies offering AI models as a service with more than $20 million in aggregate gross revenue over a consecutive 12-month period must notify Meta and sign a separate written agreement. Meta claims no ownership of outputs.

Weights were reported as available through several cloud and model hosting platforms, and running the model locally was reported to need at least 700GB of VRAM for the NVFP4 version. For readers new to the idea, our guide on open-source AI models and who they are for explains what open weights do and do not give you. Note that a custom licence with conditions is not the same as a standard open-source licence.

A long tail of other launches

A release tracker from Digital Applied, covering 1 to 27 October and updated 28 October, lists many smaller launches. We have not checked each against vendor pages, so we list them as reported by that tracker only.

Date (2026) Vendor and model What the tracker says
1 October Microsoft AI, MAI-Voice-2.1 and MAI-Transcribe-2-Streaming Text-to-speech in 23 languages; streaming speech-to-text across 60 languages
14 October xAI, Grok Voice 3.0 Speech-to-speech model for the Voice Agent API
17 October DeepSeek, V4.1-Flash-1017 Update that shortens outputs; MIT weights and 1M context
20 October Mistral, Voxtral 3 family Four public preview models for transcription, plus text-to-speech
22 October Ideogram, Vector 1.0 Hosted text-to-SVG and image-to-SVG generation
24 October Moonshot, Kimi K3.1 Pro API available; full weights promised by 4 November
27 October Google DeepMind, Genie 4 Interactive world simulation; trusted-tester preview only

Where sources disagree

Treating a tracker as the final word is risky. Two examples from our own reading:

  • Gemini 4 Argon dates. One tracker treats general availability as 13 October, another lists it on 15 October. A news report of 13 October describes the expanded access.
  • Sonnet 5.5 and GPT-6.1 Terra dates. One tracker lists Sonnet 5.5 on 1 October, while Reuters and Thurrott place it on 28 September. A tracker lists GPT-6.1 Terra on 29 October, while news coverage dates it 26 to 27 October.
  • Single-source items. One tracker also lists a Claude Duet research preview, Gemini 4 Carbon and a change making GPT-6 the default in ChatGPT for all users. We found no second source, so we do not describe them as confirmed.

The practical rule: take the date from the vendor’s announcement or a news report that names it, and treat tracker rows as pointers to check.

What it means for tool buyers

  • A new model is not automatically your best option. Cheaper models can be right for high-volume tasks, and larger ones for hard reasoning. Match the model to the job.
  • Prices are part of the story. Several launches were framed around lower cost per task. Price a real job, not just a rate card.
  • Access differs by plan and region. Argon reached some products before others, and Terra was tied to specific ChatGPT plans.
  • Open weights come with terms. Read the licence before building a product on one.
  • Security notes are worth reading. Both Haiku 5.5 and Terra reports mention cybersecurity safeguards or thresholds.

If you are new to what sits behind these names, what is a large language model is a short primer, and how to choose an AI tool covers selection.

How to evaluate a new release

Question Where to look
What exactly was released, and when? The vendor’s announcement and model documentation
Who can use it today? Plan, region and API availability notes
What does it cost for your task? The pricing page, plus a small real-world test
Which claims are the vendor’s own? Look for independent evaluations before relying on benchmark claims
What are the data and licence terms? Terms of service, data-use settings and, for open weights, the model licence

To run your own comparison, follow how to test an AI tool before you pay for it and keep notes as described in how to compare AI tools side by side fairly.

What to watch next

  • Whether Gemini 4 Argon pricing changes when the introductory rate ends on 30 November.
  • Whether Moonshot publishes the Kimi K3.1 Pro weights it promised by 4 November. We did not confirm this.
  • Whether Genie 4 and other trusted-tester previews reach wider availability.
  • Independent evaluations of the new models, which were mostly not available in the sources we read.

These are open questions, not predictions.

Frequently asked questions

Which new model is the best?

We cannot say. We did not run tests, and most performance numbers in the reports come from vendors. The right choice depends on your task, budget and data requirements.

What does open-weight mean?

The trained model weights can be downloaded and run by others, usually under a licence. That is different from the full training data or code being open, and licence terms vary.

Why do release dates differ between sources?

Trackers may record an announcement, a limited preview or general availability as the release. Check the vendor’s announcement for the date that matters to you.

Are the prices here current?

They are as reported on the dates shown. Introductory pricing, such as Gemini 4 Argon’s, can end. Confirm on the vendor’s page.

Should I switch tools because of a new release?

Only if a small trial on your own work shows a clear gain at an acceptable cost. Switching has its own costs, including retraining prompts and checking outputs.

Conclusion

The period from late September to October 2026 was crowded, but the verified core is manageable: three Claude 5.5 models, Gemini 4 Argon, Muse Spark Open and GPT-6.1 Terra, with many smaller releases around them. The sensible response is to confirm details with the vendor, price your own task and test before switching.

Sources

Keep reading