Between 1 September and 13 October 2026, Anthropic, OpenAI, Google and Mistral released new models or opened them to a limited group of users. For companies, three things change: how much text a model can take in at once, how far it works on its own with tools, and who is allowed to use it at all. This overview sticks to what the vendors document themselves.

As of 28 October 2026

Which models were released?

Anthropic: Claude Fable 5.1 and the 5.5 family

Anthropic released four models in just over five weeks. According to the vendor's model overview, they are positioned as follows:

  • Claude Fable 5.1 (1 September) is aimed at demanding reasoning and long agentic tasks, with strengths in coding agents, multi-step research and work on documents, spreadsheets and slides.
  • Claude Opus 5.5 (22 September) targets long agentic coding and knowledge work; Anthropic recommends it as the starting point for most workloads.
  • Claude Sonnet 5.5 (28 September) offers, in Anthropic's words, the best combination of speed and intelligence.
  • Claude Haiku 5.5 (7 October) is built for classification, extraction, routing and sub-agents.

All four handle a context of up to one million tokens and produce up to 128,000 tokens of output. They accept text and images and respond in text. Their knowledge ends in June 2026.

OpenAI: GPT-6

Since 7 October 2026, GPT-6 Sol and GPT-6 Luna have been available to all ChatGPT users, replacing GPT-5.6 Sol and Luna. OpenAI documents this in its Deployment Safety Hub. That page does not state a context window.

Google: Gemini 4 Argon

Google released Gemini 4 Argon in September, initially only for selected cyber-defence organisations in its Fairwind programme. According to Google's September recap, developers, businesses and consumers will follow later. Since 13 October 2026, Gemini 4 Argon has been generally available in the Gemini API according to Google's changelog (gemini-4-argon). On the same day, according to the Gemini app release notes, a staged rollout to Google AI Pro and Ultra subscribers began. Whether and in which regions the model can also be booked through Vertex AI is not clear from the sources reviewed.

Mistral: Large 4 in preview

Mistral Large 4 has been available as a public preview in Mistral Studio and via the API since 6 October 2026. The documentation lists a one-million-token context, function calling, agents and structured outputs, and describes the model as "open-weight". The weights have not been published yet, though, and no licence is stated. Reports say the weights are due at the end of October; there is no official date.

What will a company notice?

Longer context

All four Claude models and Mistral Large 4 now offer a one-million-token context. For GPT-6 and Gemini 4, the sources reviewed give no figure. In practice, a long context means that extensive contracts, manuals or project files fit into a single request. That does not replace a well-maintained document base. The more text a model receives, the longer the answer takes and the more it costs. It therefore makes sense to give the model only what matters for the task.

A second point is easy to miss: the Claude models' knowledge ends in June 2026. They only know about later events if you supply them through web search or your own documents.

Agents and tools

Anthropic and Mistral put agentic work front and centre: the model calls tools, searches, reads files and works on a task over many steps. Fable 5.1 shows readable progress updates between tool calls, as a beta feature. Haiku 5.5 is explicitly designed for sub-agents, meaning parts of a task that a larger model hands out. Mistral Large 4 supports function calling and agents.

For companies, this shifts the question. It is less "Which model is the smartest?" and more "Which systems may the agent touch, and who signs off its results?". Permissions, logs and approval steps decide the value more than the choice of model.

Availability

Not every announced model can be booked by a company. Gemini 4 Argon has been generally available in the Gemini API since 13 October 2026; companies should check pricing and regions before deciding. Anthropic offers Mythos 5.1 only to verified organisations. Mistral Large 4 is a preview, and previews can change before the final release. The Claude models are available through Anthropic's API and, according to the vendor, through Amazon Bedrock, Google Cloud, Microsoft Foundry and the Claude Platform on AWS. For many companies that is the simplest route: through a cloud they already have contracts with.

What role do region and open weights play in the EU?

For companies in the EU, two questions matter: where is the data processed, and who controls the model?

With cloud models, the region decides. Anthropic's overview lists regional endpoints with guaranteed data routing on Amazon Bedrock. Whether a specific model is available in an EU region, and what commitments apply there, is set out in each platform's documentation. The model pages reviewed for this article say nothing about EU data residency. Check this per model and per platform before you plan, and get it confirmed in the contract.

Open-weight models are the second option. Their weights can be downloaded and run on your own hardware or with a European hosting provider. The company then decides on location, access and updates itself. The trade-off is operational effort: someone has to host, monitor and update the model. Mistral Large 4 may become such an option once weights and licence are published. Read the licence carefully before you build on it.

What should companies pay attention to now?

  • Do not switch models with every release. Build your application so that the model can be swapped.
  • Put together a handful of typical tasks from your own work as a test set, and test each new model against it rather than against leaderboards.
  • Clarify region, contract and data processing before you start a pilot.
  • Do not plan preview models into production.

The AI consulting page describes how a structured model comparison for a specific project works.