Anthropic launches Claude Haiku 5.5: 90% cheaper input, OSWorld jumps to 72.4%
赛博禅心 · wechat · 2026-10-08
A detailed Chinese write-up of Anthropic's newly released Claude Haiku 5.5 and companion updates.
- Pricing: input drops from $1 to $0.10/M tokens (≤100K context tier), output $5→$0.50; the >100K tier is $0.50/$2.50. Roughly 75% cheaper than Haiku 4.5 overall, half again with Batch API. Caveat: new Opus 5.5-shared tokenizer means 30% more tokens for the same text
- Benchmarks: OSWorld 15.7%→72.4%, Terminal-Bench 0%→39.2%, GDPval-AA 735→1620, HLE (no tools) 10.2%→45.9%, Chartography 6.4%→46.4%
- First Haiku with thinking levels: Low/Med/High/Xhigh/Max, default Med via API; 1M-token context, 128K max output
- Positioning: Anthropic says complex agentic coding still belongs to Sonnet 5.5/Opus 5.5; Haiku targets summarization, context compression, classification, real-time support, browser use, or subagent duty
- Demo: on Claude ManagedAgents, an egg-drop engineering task — Opus 5.5 alone took 3m37s/$0.47; Opus leading 10 Haiku subagents took 58s/$0.14, both passing the 32m test
- Also: Sonnet 5.5 cache reads halved to $0.10 (≈20% cheaper agent workloads); Max/Team subscriptions get monthly API credits ($100/$200/$500); migration notes cover adaptive thinking, removal of prefill, and a new computer toolset
- Available on Claude Platform, Claude Code, AWS, Google Cloud, and Azure
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