Ali Asif

The AI log, frontier · open weights · image, video, world · agents and coding

Covers 5 Oct 2026 to 11 Oct 2026

Mistral Large 4, Claude Haiku 5.5 and a wave of decision models

Week of 5 October 2026: Mistral launches Large 4, Anthropic ships Haiku 5.5, FLUX 3 arrives, and OpenAI, Microsoft and Liquid release decision models.

By , published

What shipped

  1. Mistral AI

    Mistral Large 4

    A 1.05-trillion-parameter mixture-of-experts model with 49B active parameters, a vision encoder and a 1M-token window, in public preview.

  2. Anthropic

    Claude Haiku 5.5

    The small Claude is cheaper at $0.10 in and $0.50 out per million tokens under 100K, and gets effort controls and adaptive thinking for the first time.

  3. Black Forest Labs

    FLUX 3 Image

    One endpoint for text-to-image and editing, with up to ten reference images, layout boxes in the prompt and output up to 4K.

  4. Google DeepMind

    Nano Banana 2.1

    Google's image generation and conversational editing model is generally available, with up to 14 reference images and search grounding.

  5. OpenAI

    GPT-6 Luna Decisions

    A Decisions API in public beta where you send input and typed questions and get calibrated probabilities back, billed on input tokens only at $0.10 per million.

  6. Microsoft

    Microsoft-Decision-1

    A Qwen3.5-9B model post-trained for decision scoring that writes no text and returns a probability for yes/no, choice and score questions.

  7. Liquid AI

    d1-3B

    An open 3.12B decision model that Liquid reports at 8 milliseconds per decision on an RTX 4090.

  8. Mistral AI

    Voxtral Mini 4B Realtime Arabic

    An open streaming speech-to-text model for Modern Standard Arabic and Gulf, Levantine, Nile Valley and Maghrebi dialects, built to handle code-switching.

  9. Google DeepMind

    EmbeddingGemma 2

    A 740M open embedding model that maps text, code, images, video and audio into one vector space, across more than 100 languages.

  10. JetBrains

    Mellum2.1 Thinking

    A 12B open coding model with 2.5B active parameters whose repository-level reinforcement learning lifts SWE-bench Verified from 2.0 to 47.0.

Two flagships and a new category in the same week. The category surprised me most.

Mistral Large 4

Mistral launched Large 4 in public preview on 6 October. It is a mixture-of-experts model with 1.05 trillion total parameters, 49B active, a 1.6B vision encoder and a 1M-token context window, taking text and images. The docs list structured outputs, function calling, document question answering, batching and built-in agent tools. For a European lab to land in the trillion-parameter range is the headline, and I'd check the license terms before building on it.

Claude Haiku 5.5

Anthropic released Haiku 5.5 on 7 October, calling it its cheapest, fastest and most capable small model. It is built for high-volume work: summaries, classification, routing, subagents and live support. It is the first Haiku with effort controls and adaptive thinking. Pricing is $0.10 per million input tokens and $0.50 output for prompts up to 100K tokens, and $0.50 and $2.50 above that.

Decision models

Over six days a new kind of model showed up in several places. Instead of generating text, a decision model takes input plus typed questions (yes/no, a choice, a score) and returns calibrated probabilities in one pass. OpenAI shipped a Decisions API in public beta on 6 October, powered by GPT-6 Luna, billed on input tokens only. Microsoft announced Microsoft-Decision-1 on 9 October, generally available in Foundry. Liquid AI opened d1-3B on 7 October, reporting 8 ms per decision on a consumer GPU. Perplexity and Meta released theirs the same week, and Cloudflare had opened Clef a few days earlier. I haven't read those closely.

Images and embeddings

Black Forest Labs released FLUX 3 Image on 5 October, and Google made Nano Banana 2.1 generally available the next day. Google also opened EmbeddingGemma 2, a 740M model that puts text, code, images, video and audio in one 768-dimension space.

Arabic and code

Mistral released Voxtral Mini 4B Realtime Arabic on 9 October, an open streaming speech model for Modern Standard Arabic and several dialects, tuned for switching languages inside one conversation. JetBrains released Mellum2.1 Thinking on 8 October, a 12B open coding model that reports a jump on SWE-bench Verified from 2.0 to 47.0 under the same setup.

What I'd watch next

Whether decision models become a standard API shape across vendors, and whether anyone publishes an Arabic benchmark for them.

Sources

Topics: Anthropic, OpenAI, Google DeepMind, Claude, Mistral AI, Mistral, Liquid AI, Nano Banana