← The Polylog AI Intelligence Brief
Morning Edition · Monday, October 12, 2026Published at 2:16 AM EDT · New York
Anthropic prices Claude Haiku 5.5 at one tenth of its predecessor and gives it a one-million-token context window
At $0.10 per million input tokens and $0.50 per million output tokens for prompts below 100,000 tokens, Haiku 5.5 undercuts rival small models while scoring 72.4 percent on OSWorld in Anthropic's reported results.
Anthropic released Claude Haiku 5.5 on October 7 and describes it as its fastest, cheapest and most capable small model yet. The most notable figure is the price. Prompts up to 100,000 tokens cost $0.10 per million input tokens and $0.50 per million output tokens. Longer prompts cost $0.50 and $2.50. Haiku 4.5 cost $1 and $5, so the short-prompt rate is one tenth of the previous price. Anthropic says the model costs about 75 percent less to run on average, and an analysis of the price table notes that this lower figure reflects the long-context tier and the extra tokens spent on thinking.
The architecture and interface changed too. The context window grew fivefold, to one million tokens, the first time Anthropic's cheapest model has had that capacity. Haiku now uses adaptive thinking by default and, for the first time in the Haiku line, accepts the effort parameter. Anthropic reports 72.4 percent on OSWorld, 39.2 percent on Terminal-Bench at maximum effort and a GDPval-AA score of 1620. Those are vendor numbers. The independent check comes from Artificial Analysis, which places Haiku 5.5 at 43 on its Intelligence Index at maximum effort, against 41 for Gemini 3.8 Flash and 38 for GPT-6 Luna. That lead is small, and the maximum-effort setting uses more tokens, which offsets part of the price advantage.
The release fits a broader pattern of price cuts across the industry. Anthropic shipped Sonnet 5.5 on September 28 and said it runs 30 percent faster and costs up to 30 percent less than Sonnet 5. OpenAI launched GPT-6.1 Sol at DevDay on September 29 at $2 input and $10 output per million tokens, one fifth of GPT-6 Astra's rates, and claimed 75.2 percent on DeepSWE v1.1 against Astra's 74.8 percent. In July, Meta began charging for its Muse Spark API at $1.25 and $4.25.
What this means
The mechanism here is a reset of price per unit of capability at the high-volume end of the market. Classification, extraction, routing and sub-agent calls make up most API token volume, and that is where Haiku 5.5 competes. Startups that resell wrapped small models lose margin first. So do open-weight hosting providers whose main selling point was price. Buyers gain, and agent architectures that fan out many cheap sub-agent calls become economical at one-million-token context. Anthropic can afford the cut because efficiency gains lowered its serving costs, but its stated 75 percent figure is lower than the headline tenfold price cut, which suggests the margin is thinner than the sticker price implies. The market already expected price cuts. What is new is a tenfold cut at the small-model tier paired with a fivefold increase in context length. The precedent is the 2024 round of small-model price cuts, after which Google and OpenAI matched prices within weeks. The likelier outcome is that Google and OpenAI cut their own small-model rates within weeks. This depends on Haiku 5.5's independent scores holding up across more evaluations. If Gemini Flash or GPT-6 Luna keep their prices unchanged and still gain market share, this read is wrong.
Part of a tracked trend
Frontier Model Price War
Frontier API vendors increasingly compete on strategic price-cutting rather than pure capability, repeatedly launching or repricing models below prevailing rates to grab share and compressing industry inference margins; expect recurring below-rival pricing moves.
Townsquare
Start a discussion in Townsquare.
More from this edition
- Ukrainian drones knock out three Yandex data centers in four days, disrupting Russia's largest cloud
- Anthropic removes live internet access from all internal evaluations after Claude models acted on real websites
- Anthropic opens three-tier cyber program giving vetted security teams access to Mythos 5.1 with reduced safeguards
- SWE-Duel benchmark pits coding agents against each other to resist training-data contamination
- New research catalogues prompt-injection routes through agent systems as Nadella says models should be treated as compromised
- Two arXiv papers target the ordering limits of autoregressive decoding, one for speculative speed and one for revisable output
- Claude Science agents assemble a full-sky ultraviolet map, with a third of it predicted, not observed
Comments
0No comments yet.