Mistral Large 4 brings a trillion parameters to Europe as Google and OpenAI make AI cheaper to run

Mistral Large 4 "Le Chonk" puts a one trillion parameter model into public preview

If you have wanted a serious European alternative to the big US and Chinese models, this is the most important release Mistral has made in a long time. Mistral Large 4 is a mixture-of-experts model with one trillion parameters in total and 49 billion active at a time. It reads images as well as text, supports more than 160 languages and offers a context window of up to one million tokens on OpenRouter. It is live now through Mistral's API and OpenRouter at $1.36 per million input tokens and $4.18 per million output tokens, which undercuts most frontier models by a wide margin.

Mistral is pitching it hard at cybersecurity, finance and coding, and says it was trained on around 4,000 Nvidia Grace Blackwell GPUs in its own European data centres, the kind of build-out its 3 billion euro raise in September was meant to pay for. The weights are promised later this month (VentureBeat reports 27 October) under a custom Mistral licence rather than a fully permissive one, and the strong benchmark figures are Mistral's own, so wait for independent testing before you switch. Coming a day after Reflection's Beam and two days after Aleph Alpha's Kolibri, it caps a remarkable run for Western open-weight models.

Nano Banana 2.1 upgrades Google's image model for editing and consistent characters

If you use Gemini for product shots, thumbnails or illustrations, Nano Banana 2.1 fixes the things that usually force a dozen re-rolls. Google says it beats its previous image models across the board, with clear gains in visual design, mask-based editing and keeping a subject looking the same from one image to the next. Developers get up to 14 reference images per request, grounding with Google Search, adjustable thinking levels and cleaner output at 1K, 2K and 4K, including very wide panoramas that used to show tiling artefacts.

It is listed as a stable model in the Gemini API and went live on OpenRouter the same day at roughly 3.4 cents per 1K image and 5 cents per 2K image, cheap enough to build into everyday workflows. It is the first meaningful refresh since Google brought Nano Banana 2 into its products in the spring.

EmbeddingGemma 2 brings multimodal search to your phone as an open model

If you are building search, recommendations or retrieval into an app, you can now do it on the device across text, code, images, audio and video without sending anything to a server. EmbeddingGemma 2 is a 740 million parameter open model that maps all of those formats into one shared vector space, so a spoken question can find a matching photo or code snippet. It is modular, so a text-only build uses about 270 million parameters, and Google says the quantised model needs roughly 191MB of memory for text and 567MB with every encoder loaded. Weights are out under Apache 2.0 on Hugging Face and Kaggle, with Ollama, llama.cpp and LiteRT builds ready on day one.

OpenAI's Decisions API opens to every developer as a public beta

If your app spends a lot of tokens just deciding things (which model to call, whether a message is spam, which tool to use next), the Decisions API is built for exactly that job. Rather than writing free text, it answers questions you define with a typed result and a confidence score: a yes or no probability, a choice from a list, or a score on a scale. It runs on GPT-6 Luna, and OpenAI says it decides up to ten times faster than calling GPT-6 Luna through the Responses API, which matters when a decision sits in front of every user request.

Unite.AI reports pricing of 10 cents per million input tokens with output free, and it accepts images as well as text, with Zero Data Retention and European data processing available. It was first shown in limited preview at DevDay last week, and the public beta puts OpenAI squarely into the small, fast decision-model market that Cloudflare and AWS entered at the start of October.

OpenAI API rate limits get simpler, with the top tier now unlocked at $500

If you are a small team or solo builder who has hit OpenAI's rate limits during a launch, this helps you directly. OpenAI has collapsed its five paid usage tiers into three, called Build, Launch and Grow, and you can now reach the highest tier, Grow, after $500 in total API payments rather than the previous $1,000. Existing paid organisations have been moved across automatically, so there is nothing to do except check your limits page. A small policy change on paper, but it removes one of the most common blockers to scaling a new product on the API.

Atlassian brings OpenAI models deeper into Jira, Confluence and Rovo

If your team lives in Jira and Confluence, expect OpenAI's models to show up more often in your day. The expanded partnership has GPT-6 Astra and the GPT-5.6 series powering agents across Atlassian's Rovo platform, drawing on its Teamwork Graph of projects, documents and decisions, while ChatGPT and Codex plugins connect straight to Jira work items and Confluence pages. Atlassian says teams will soon be able to assign Jira work directly to AI agents and track their progress like any other ticket. VentureBeat notes Atlassian is committing spend to OpenAI but keeping its platform multi-model, so this adds choice rather than lock-in.

An unreleased OpenAI model publishes a batch of new maths results

This is not a product you can use yet, but it is a strong hint at what OpenAI's next model can do. OpenAI has published a broad set of new mathematical results produced by an internal frontier model, in a public GitHub repository with many proofs formalised in Lean so a computer can check them. It reports an average of around three hours of ChatGPT Pro-level thinking per result and says it took advice from the Institute for Advanced Study's advisory group on how to release AI-produced maths responsibly. It follows the Navier-Stokes proposal from an internal OpenAI system last month; the model itself is not available, and OpenAI only says it is working to release it responsibly.

Anthropic's Cyber Verification Program expands to three tiers with access to Mythos 5.1

If you work in security and have found Claude refusing legitimate defensive tasks, this is the route to fewer blocks. Anthropic has rebuilt its Cyber Verification Program into three tiers: Defense Access for incident response and malware analysis, Red Team Access for authorised penetration testing, and Specialized Access for vetted critical infrastructure organisations. Every tier covers Claude Opus 5.5, Sonnet 5.5 and Claude Mythos 5.1, the model usually kept behind tighter safeguards, and SiliconANGLE reports that the earlier Project Glasswing scheme has been folded in.

The trade-off is that enrolled organisations must allow data retention so Anthropic can watch for misuse, although a self-hosted retention option is promised later this year. It mirrors the approach Google took with Gemini 4 Argon last week: the most capable cyber models go to verified defenders first.

Claude for Startups opens to more founders with a bundled "Startup Stack"

If you are starting a company on Claude, the startups programme is now open to more founders and is worth a look before you pay full price for anything. Members can get a year of Claude Team, API credits, office hours with Anthropic's Applied AI team and a new Startup Stack of offers from companies building on Claude, including Linear, Lovable, ElevenLabs, Granola and Hex. The post was one of the most-liked AI announcements of the day on X, with around 11,000 likes, and most replies were founders asking how to apply.

Grok 4.7 is now live on Microsoft Foundry

If your organisation buys AI through Azure, you can now use xAI's latest model without a separate contract. SpaceXAI announced that Grok 4.7, launched in September as its strongest model for coding and knowledge work, is available on Microsoft Foundry, putting Grok on all three of the big clouds. On the same day, a post from Elon Musk claiming Grok 4.7 is excellent with large code repositories drew around 15,000 likes; treat that as a claim until independent tests catch up. At the time of writing the Foundry news had only been announced on X.

OpenAI's GPT-6.1 Sol arrives on Snowflake Cortex AI in public preview

If your company's data already lives in Snowflake, you can now point OpenAI's newest flagship at it without moving anything out. Snowflake says GPT-6.1 Sol is in public preview across Cortex Inference, Snowflake CoCo, Snowflake CoWork and Cortex Agents. For data teams that means the latest OpenAI reasoning model inside existing governance and access controls, which is usually the hard part of getting AI past a security review.

Scale AI open-sources AgentEnv, the framework behind its agent training environments

If you train or evaluate agents, Scale has just handed over the tooling it uses internally. AgentEnv is a framework for building reinforcement learning environments, the simulated worlds where agents practise tasks, with 48 built-in step types and support for AWS and Google Cloud plus Docker, Modal and E2B sandboxes. Scale says every RL environment it builds runs on it, so this is production-tested code rather than a research demo.