Anthropic ships Claude Haiku 5.5: +26 points in a year, 10x cheaper than predecessor
ArtificialAnlys · x · 2026-10-08
Anthropic has released Claude Haiku 5.5, scoring 43 on the Artificial Analysis Intelligence Index — up 26 points in one year — and it's the first Haiku with effort settings, adaptive thinking, and tiered pricing.
Key takeaways
- Leading small-model performance: at max effort it edges out GLM-5.3 Flash (42), Gemini 3.8 Flash (41) and GPT-6 Luna (38), is comparable to open-weights Kimi K3 (44), and trails Claude Sonnet 5.5 (56) by 13 points.
- Heavy token usage: 162k output tokens per Intelligence Index task at max, roughly 3x GPT-6 Luna; moving from xhigh to max buys 2 points for 1.8x tokens.
- Strong agentic knowledge work: 1578 Elo on AA-Briefcase, ahead of Kimi K3 and GLM-5.3.
- Terminal use jumps: 33% on Terminal-Bench 4.0, up from 0% for Haiku 4.5.
- Narrower knowledge, fewer hallucinations: 36% AA-Omniscience accuracy vs 55% for Gemini 3.8 Flash, but hallucination rate is lower (40% vs 55%).
- AutomationBench likely understated: 35% due to a pre-release over-refusal safety bug that Anthropic is fixing.
Pricing & specs: $0.10/$0.50 per 1M input/output tokens under 100k context (10% of the previous Haiku's price), stepping up 5x above 100k; context window expanded from 200k to 1M tokens; text and image input.
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