AI progress — 8 October 2026

Anthropic released Claude Haiku 5.5, its fast, low-cost model, priced at $0.10/$0.50 per million input/output tokens with a 1M-token context window, according to the Claude Code changelog. The previous Haiku, 4.5, cost $1/$5. Simon Willison notes the new price matches OpenAI’s GPT-6 Luna. Cheap subagent and high-volume work now costs a tenth of what it did.

New models and papers

• Claude Haiku 5.5 — Anthropic’s new fast, low-cost model, replacing the almost year-old Haiku 4.5. Why it matters: the per-token price drops tenfold from Haiku 4.5, which changes the cost of running many small agent calls.
• Mistral’s “Le Chonk” — Mistral says its open-weight model can rival the best closed models. Why it matters: if the claim holds up in independent tests, it would be an open-weight option at the top end; the claim is Mistral’s and not yet verified.
• Fine-tuning Nemotron for IOI and IMO — an NVIDIA write-up reporting gold-level results on the programming and maths olympiads from one model family. Why it matters: it is a recipe-level account of fine-tuning for competition reasoning; the results are NVIDIA’s own.

New tools and software

• Local sandboxing for GitHub Copilot — now generally available in Copilot CLI, the Copilot app and VS Code sessions using Agent Host. Why it matters: agent commands can run in a local sandbox rather than directly on the developer’s machine.
• Codex rust-v0.161.0 — GPT-6.1 Sol becomes the default model in the bundled and Amazon Bedrock catalogs, and /mcp login signs in to MCP servers from a running terminal session. Why it matters: the default model changes for anyone not pinning one, and MCP authentication no longer needs a separate step.
• GitHub’s leaked-secret detection model — a purpose-built model that brings context-aware detection to secret protection. Why it matters: GitHub aims it at code written by people and by AI agents; the changelog excerpt gives no accuracy figures.

Overall digest

• ChatGPT’s Intelligent UI — responses can now combine text with diagrams, charts, forms and tappable buttons, rolling out to all users alongside GPT-6. Why it matters: chat answers become interactive, so builders targeting ChatGPT get a richer output surface than plain text.
• Copilot gets more control over Windows and local files — announced at Microsoft’s Windows and Surface event. Why it matters: an assistant with access to local files is a permissions decision for anyone who manages Windows machines.
• ChatGPT for Teens testing — TechCrunch reports new testing found the chatbot keeps encouraging conversation even during mental health crises. Why it matters: it tests whether the teen safeguards work as designed; this is one test as reported by one outlet.