← The Polylog AI Intelligence Brief
Morning Edition · Saturday, October 10, 2026Published at 2:23 AM EDT · New York
Anthropic prices Claude Haiku 5.5 at one-tenth of Haiku 4.5 and lifts its OSWorld score to 72.4%
Input costs $0.10 per million tokens for prompts up to 100,000 tokens, every rate is five times higher above that, and the model accepts a one-million-token context with adjustable effort.
Anthropic released Claude Haiku 5.5 on October 7 and described it as its fastest, cheapest and most capable small model. List pricing is $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. Above that threshold, every rate rises fivefold, to $0.50 and $2.50. Cache reads cost $0.01 per million tokens. Those rates are a tenth of Haiku 4.5's, and Anthropic says the average effective cost is 75 percent lower. Haiku 5.5 is the first Haiku with adjustable effort, and it takes text and image input with a one-million-token context.
The capability gain is largest on computer use. Haiku 5.5 scores 72.4 percent on OSWorld 2.1, up from 15.7 percent for Haiku 4.5, and 39.2 percent on Terminal-Bench 4.0. Artificial Analysis, an independent aggregator, scores it 43 at maximum effort and 34 at the default medium effort, against 17 for Haiku 4.5. Anthropic acknowledges that it still performs below Sonnet 5.5 on hard coding work. The OSWorld and Terminal-Bench figures come from the vendor. The Artificial Analysis index is the only outside measurement so far.
Haiku 5.5 completes a fast repricing of the whole Claude lineup. On September 22, Anthropic said Opus 5.5 performs at the level of Claude Fable 5.1 on most work and costs 40 percent less to run than Opus 5. On September 28, it said Sonnet 5.5 runs 30 percent faster than Sonnet 5 and costs up to 30 percent less.
Rivals are cutting prices on the same schedule. OpenAI released GPT-6.1 Sol on September 29 at $2 per million input tokens and $10 per million output tokens and said it delivers near-Astra capability at one-fifth of Astra's price. Meta prices Muse Spark 1.1 at $1.25 and $4.25. In a customer story, OpenAI says Asana made its browser agent 76 times cheaper and 5 times faster in tests. The page's headline credits GPT-6.1 Sol and its summary credits GPT-6 Astra in Codex, so the model behind the figure is unclear.
For engineers, the pricing structure matters as much as the headline rate. Agent loops that build up context past 100,000 tokens pay five times the base rate. In practice, that makes context compaction and cache hit rates the two main factors in controlling cost.
What this means
Small-model throughput is where agent fleets spend most of their tokens on subagents, classifiers and tool-call routing. A tenfold cut at that tier moves those workloads to Anthropic and away from three groups: hosted open-weight providers, small-model tiers at other labs, and Chinese open-weight models whose main selling point was price. Anthropic wants volume and lock-in to its agent tooling. The tiered context pricing shows it still guards margin on long-context loads. The consensus reads this as a margin sacrifice. The structure suggests something narrower: cheap short prompts, with long-context agent work still priced at a premium. Google's 2024 Gemini Flash price cuts are the precedent, and competitors cut their small tiers within weeks. The likelier outcome is that OpenAI and Google reprice their small tiers within weeks to match Haiku 5.5's short-context rates. That depends on Anthropic sustaining throughput at these prices without tightening rate limits. The read is wrong if competitors hold list prices through the end of the year, or if Anthropic restricts Haiku 5.5 capacity. For listed companies, cheaper inference widens agent deployment and helps compute suppliers on volume, while independent inference resellers face margin pressure.
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
- OpenAI's unreleased model claims 372 math result families, and specialists say most remain unverified
- Fired OpenAI safety researchers deny leaking reports on Astra's less monitorable reasoning
- Microsoft's Decision-1 scores fixed answer options in one pass and charges nothing for output tokens
- SpaceX seeks $40 billion in debt for Nvidia chips as Firmus pulls its $5 billion Australian listing
- OpenAI says Russian and Iranian networks used ChatGPT to run fake journalists and a Latin American think tank
- Windows 11 ships enforced sandboxes for AI agents as an open-source tool hands agents a decompiler
- OpenAI moves all ChatGPT tiers to GPT-6 and lets the model build interactive interfaces inside its answers
Comments
0No comments yet.