Claude Haiku 5.5 Launches at $0.10 Per Million Tokens as Mistral Previews Large 4

This Week in LLMs, Monday, October 12, 2026. Covers October 5–11.

TL;DR

• Anthropic released Claude Haiku 5.5 on October 7 at $0.10 input and $0.50 output per million tokens for prompts up to 100,000 tokens (Anthropic).
• Mistral opened a public preview of Mistral Large 4, an open-weight model whose weights are due by the end of October (Mistral).
• Anthropic published a new Usage Policy effective November 12, and OpenAI began offering text watermarking to meet EU rules (Anthropic, OpenAI).

Top Story: Haiku 5.5 Makes Small Models Serious

Anthropic's Claude Haiku 5.5 is its fastest and cheapest model, released October 7 on Amazon Web Services, Google Cloud, Microsoft Azure and the Claude Platform. Prompts over 100,000 tokens cost $0.50 input and $2.50 output per million. Cache reads are $0.01 per million.

Artificial Analysis scored it 43 on its Intelligence Index at max effort, up 26 points from the previous Haiku, and reports a 1 million token context window. That puts it near Kimi K3 (44) and ahead of GPT-6 Luna (38). It is also the first Haiku with effort settings and adaptive thinking.

There is a trade-off. At max effort it used about 162,000 output tokens per test task, roughly three times Luna's 50,000. At high effort it scores 38 and uses about 55,000. Artificial Analysis says its cost figures are provisional because its site does not yet reflect the tiered pricing.

Why it matters: Business users should price a finished task, not a token; a cheap model that thinks longer can cost more than the sticker suggests. Builders should test high effort before max, since the score drops from 43 to 38 while token use falls by about two-thirds.

Leaderboard Watch

Sources: Arena text leaderboard (page dated October 8), Artificial Analysis. Arena does not publish prior ranks on the page we could read, so we do not state them. Claude Opus 5.5 (max) still leads Artificial Analysis at 58. We did not verify new releases from Meta, xAI, DeepSeek or Qwen this week.

Model & Pricing Changes

• Mistral Large 4 preview: $1.36 input and $4.18 output per million tokens; a 1-trillion-parameter mixture-of-experts model with 52 billion active parameters. Mistral names no license yet (Mistral).
• Independent check: Artificial Analysis gives the preview a 38, below Haiku 5.5's 43 (Artificial Analysis).
• GPT-6 in ChatGPT: Plus, Pro, Business and Enterprise users get GPT-6 Sol with Intelligent UI from October 7. Free and Go users get Luna starting the next day. No pricing was given (OpenAI).
• Gemini 4 Argon pricing: Introductory $2 input and $10 output per million tokens, rising to $4 and $20 later. Output limit rises to 1 million tokens from 64,000. Announced September 30; Google gives no firm wider-release date (Google).
• Haiku 5.5 over-refusals: Artificial Analysis says an automation test likely understates the score because of a safety-refusal issue, and will re-run it after Anthropic's fix (Artificial Analysis).

Policy Watch

• Anthropic Usage Policy: Takes effect November 12. It adds rules on fake accounts and influence operations, tightens limits on surveillance and weapons, and sets conditions when Claude controls hardware (Anthropic).
• EU text watermarking: OpenAI will add an invisible watermark to eligible ChatGPT and Codex text for EU users in coming weeks. API customers worldwide can opt in from October 5 (OpenAI).
• Cyber access: Anthropic expanded its Cyber Verification Program on October 6, giving qualifying security professionals advanced cyber capabilities with fewer blocking classifiers (Anthropic).

Tools Worth a Look

• Copilot local sandboxing: Now generally available. It limits what Copilot-launched commands can reach on your machine, including files, network and credentials, on Windows, macOS and Linux (GitHub).
• Gemini agent: Google Cloud introduced a work agent that plans tasks and returns output in documents, inboxes and developer tools. The post does not state availability (Google).
• Deno team joins Cloudflare: The team is merging its self-hosting work into Cloudflare's open-source workerd runtime, announced October 9 (Cloudflare).

Events (Next 30 Days)

• Oct 15: Conversational AI Innovation Summit, New York (calendar)
• Oct 20–21: AI & Big Data Expo Europe, Amsterdam (calendar)
• Oct 27–29: ODSC AI West, Burlingame (calendar)
• By Oct 31: Mistral says Large 4 weights arrive (Mistral)
• Nov 4–5: Future of AI Summit, London (calendar)
• Nov 10–12: Agentic AI Week USA, Atlanta (calendar)

Keep Up With the Rankings

See every model on the LLM1 leaderboard, and subscribe to the LLM1 newsletter for this roundup every Monday.