October 2026 AI model releases: GPT-6.1 Terra, Gemini 4 Argon, Claude 5.5 and Kimi K3.1 Pro

If you blinked in September and October, you missed about a dozen model launches. Every lab shipped something, every launch post claimed a benchmark win, and every pricing page changed.

This post cuts through it. For each release you get what is new, which problem it goes after, whether that problem is actually solved yet, and what you can build with it this week.

What actually shipped

OpenAI GPT-6.1 Terra: computer use gets cheap

What is new. Terra is OpenAI's middle-tier model, tuned to read what is on a screen and take actions across apps. Developers get it as gpt-6.1-terra in the API. Plus, Pro, Enterprise and Edu users get it in ChatGPT Work and Codex.

The pain point. Agents that click through software have been too slow and too expensive to run at volume. A workflow that touches twenty screens burns a lot of tokens.

Fixed or not? The cost part, mostly yes. Standard pricing is $1 per million input tokens and $5 per million output, which OpenAI says is 50% lower than GPT-5.6 Terra and GPT-6.1 Sol. The reliability part is still being proven. OpenAI's own error benchmark dropped from 53.6% to 16.7%, but that is a vendor number, and one in six is still a lot of errors for unattended work.

What you can do with it. Back-office automation that runs through web apps without an API: data entry, report downloads, QA passes over your own product. Keep a human checkpoint before anything irreversible. Note that tool calling needs the Responses API, and prompts above 272,000 input tokens are billed at a higher rate.

Google Gemini 4 Argon, Carbon and EmbeddingGemma 2

What is new. Gemini 4 Argon is Google's new frontier reasoning model, aimed at complex coding, financial analysis and legal work. It is available in AI Studio, the Gemini API, Google Cloud, Search and the Gemini app. Gemini 4 Carbon, which Google calls its strongest model for long-horizon engineering and cyber defence, is only in testing with the US government and vetted defenders for now.

The pain point. Long, high-stakes documents. Contracts, filings and large codebases are where cheaper models quietly miss a clause or a dependency.

Fixed or not? Not something we can confirm from launch posts alone. Treat Argon as a strong candidate for document-heavy work and test it on your own files before trusting it.

The release we would actually watch is smaller. EmbeddingGemma 2 is an open, 740 million parameter model that maps text, code, images, video and audio into one search space on everyday devices, and Google says its design cuts local vector storage by up to six times. That makes private, offline search inside a mobile app realistic, with no server and no data leaving the phone.

Google also put AI Studio on Google Play, so Android users can build personal AI apps and install them straight onto their phones. For app studios like ours, that is a signal of where small, personal software is heading.

Anthropic's Claude 5.5 family and Fable 5.1

What is new. Anthropic replaced almost its whole line in five weeks:

• Claude Fable 5.1 (1 September) is Anthropic's flagship model. It shipped alongside Mythos 5.1, which Anthropic describes as the same model with different safeguards, available only through trusted access programmes. Coverage of the launch reported about 25% lower cost on typical workloads.
• Claude Opus 5.5 (22 September) targets agentic coding and knowledge work. Anthropic says it costs 40% less to run than Opus 5 on typical workloads.
• Claude Sonnet 5.5 (28 September) is described as 30% or more faster than Sonnet 5 and up to 30% cheaper for most work.
• Claude Haiku 5.5 (7 October) is the small, fast model for summarisation, subagents and browser use.

The pain point. Coding agents that run for an hour are expensive, and teams end up rationing them.

Fixed or not? Partly. A 30 to 40% cut on every tier compounds when an agent makes hundreds of calls. It does not change the need to review what the agent wrote.

What you can do with it. Use the tiers as a team: a big model plans and reviews, small models do the repetitive steps. That split is where most of the savings come from.

Moonshot Kimi K3.1 Pro: the open-weight challenger

What is new. Kimi K3.1 Pro launched on 23 October at $2 per million input tokens and $10 per million output on Moonshot's API. Moonshot reports 65.7% on Terminal-Bench 4.0, against 66.4% for Claude Opus 5.5.

Fixed or not? Too early to say. Those scores come from Moonshot's own launch, and early independent results disagree with each other.

Update, 5 November: the open weights are now on Hugging Face, published on 2 November. They are a 1.56 TB download that needs a full 8-GPU B300-class server, so we wrote up what self-hosting Kimi K3.1 Pro really takes.

The result: what this means in practice

Strip away the benchmark charts and one trend is left: the same quality of work keeps getting cheaper. Every lab that shipped this autumn lowered prices or launched a cheaper tier. That changes product decisions:

1. Features you shelved for cost may now pay for themselves. Re-run the numbers on anything you dropped in spring.
2. Small models are now good enough for most steps. Route simple work to Haiku 5.5, Terra or a Flash-class model and save the expensive model for planning and review.
3. On-device AI is getting real. Open models like EmbeddingGemma 2 mean some features need no server bill at all.

How we would choose today

Whatever you pick, test it on twenty real examples from your own product before you commit. Launch benchmarks tell you what a model can do. Your data tells you what it will do.

Sources

• Introducing GPT-6.1 Terra, OpenAI
• OpenAI upgrades ChatGPT and Codex with GPT-6.1 Terra, 9to5Mac
• The latest AI news we announced in October 2026, Google
• Anthropic newsroom
• Introducing Claude Fable 5.1 and Claude Mythos 5.1, Anthropic
• Kimi K3.1 Pro tech blog, Moonshot AI
• moonshotai/Kimi-K3.1-Pro model card, Hugging Face