Claude Haiku 5.5 cuts small-model prices 90% below the…

Anthropic's October 7 release drops per-request costs into the range where small teams can finally run always-on sales and support subagents.

Anthropic released Claude Haiku 5.5 on October 7, 2026, and the pricing move matters more than the model card. For prompts under 100,000 tokens, input runs $0.10 per million and output $0.50 per million, with cache reads at $0.01 and cache writes at $0.125. Above that threshold, pricing steps up to $0.50 input, $2.50 output, $0.05 cache reads, and $0.625 cache writes. Haiku 4.5, by comparison, listed at $1.00 input, $5.00 output, and $0.10 cache reads.

That’s a 90% cut on the sub-100K tier and 50% above it. Anthropic’s own footnote, flagged by Unite.AI, notes that roughly 90% of Haiku 4.5 traffic lived in the cheaper bucket. The sub-100K input price now matches OpenAI’s GPT-6 Luna, which is the real story. The small-model pricing war has a new floor.

The capability jumps are what justify the Haiku name finally doing real work. On GDPval-AA v2.1, Haiku 5.5 scores 1620 against Haiku 4.5’s 735 and Luna’s 1437. On the OSWorld 2.1 computer-use subset, 72.4% versus 15.7%. On Terminal-Bench 4.0, 39.2% versus a flat 0.0%. The context window expands from 200,000 to 1,000,000 tokens. Knowledge cutoff is June 2026, text output only.

Customer validation lands where it counts for always-on agent work. HubSpot’s Ze’ev Klapow reported 92.8% accuracy across three runs on the company’s simulated CRM task suite, the best score HubSpot has recorded there, with the fastest completion and lowest false-positive rate. Asana’s Aaron Vinh described “over a 30% reduction in latency for task completions and up to 2.5x faster inference per agent turn” powering its AI Teammates product. AlphaSense moved from 0.76 to 0.84 across 400 production queries on a feature it runs about 8 million times a week. Box reported 11 points higher accuracy at roughly half the latency.

Anthropic also cut Sonnet 5.5 cache reads 50%, from $0.20 to $0.10 per million tokens, trimming roughly 20% off typical agentic Sonnet workloads. Startup Fortune’s framing captured the shift: a model at a tenth of a cent per thousand input tokens “changes which problems are even worth running through an LLM in the first place.” The economics of persistent subagents just stopped being the reason not to ship one.

Claude Haiku 5.5 cuts small-model prices 90% below the 100K-token line

Sources