Anthropic Ships Claude Haiku 5.5 at One-Tenth the Price, OSWorld Jumps to 72.4%
新智元 · wechat · 2026-10-08
Anthropic released Claude Haiku 5.5, priced at one-tenth of Haiku 4.5 and one-twentieth of Sonnet 5.5, matching OpenAI's GPT-6 Luna. On official benchmarks it beat Luna on all six tasks where Luna reported scores, and closed in on Sonnet on several.
- Scores: OSWorld computer use jumped from 15.7% to 72.4%, just 11.5 points behind Sonnet 5.5; Terminal-Bench 4.0 went from zero to 39.2%; GDPval-AA for knowledge work hit 1620, more than double the prior generation; HLE without tools rose from 10.2% to 45.9%. Coding remains the weak spot, with Sonnet 5.5 at 70.6% on Terminal-Bench.
- Pricing: $0.10 per million input tokens and $0.50 output (prompts under 100K tokens); above 100K it's $0.50 input / $2.50 output; cache reads $0.01, batch halves it again. Anthropic estimates 75% savings versus Haiku 4.5 and roughly four times the calls for the same budget. Luna is cheaper on long context (price rises only around 270K tokens). Sonnet 5.5 cache reads drop from $0.20 to $0.10. Max and Team subscriptions get monthly API credits starting this week: $100 for Max 5x, $200 for Max 20x, up to $500 shared for Team.
- Positioning: built for high-volume, fast, cheap work such as summarization, context compression, database queries and classification, and it's the fastest Claude model; it can serve as a sub-agent under Opus/Sonnet. Devin Fusion pairs Opus 5.5 with Haiku 5.5 sub-agents to hold FrontierCode at 66.2.
- Availability and migration: live in the Claude apps, Claude Code and on AWS, Google Cloud, Azure, Cursor and OpenRouter, model ID claude-haiku-5-5, 1M-token context, up to 128K output tokens, and the first Haiku with five thinking-effort levels. Migration notes: 30% more tokens for the same text, budgettokens replaced by adaptive thinking plus Effort levels, temperature/topp/topk removed.
Related event: Anthropic Launches Claude Haiku 5.5 with ~75% Lower Running Costs(59 posts)→
More from Models
- 'Compression of the RH proof has begun': AI circles buzz over unverified Riemann claim — teortaxesTex · 2026-11-22
- GLM 5.3 and DeepSeek V4.1 Flash resold at 60% off: the economics of open-source giveaways — lxfater · 2026-11-22
- GPT-6.2 Astra and Sol answer more directly but score lower on transparency than Claude — The Decoder · 2026-11-22
- Baseline model test: no persona, no forced formatting, no safety bypasses observed — Sauers_ · 2026-11-22
- 2-day Gemini stress test: $4.99 Google AI Pro nearly maxed, orchestrator UI revealed — lipeng0820 · 2026-11-22
- Timed eval task asks users to deliberately make Claude fail at real work in 12 minutes — Icy-Measurement10 · 2026-11-22