In this article4 sections
Anthropic released Claude Haiku 5.5 on Wednesday, pricing its smallest model at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. That is down from $1 and $5 for Haiku 4.5, and matches the rate OpenAI set for its rival small model, GPT-6 Luna, when it launched on September 22, according to Decrypt.
The company calls Haiku 5.5 its cheapest, fastest and most capable small model. It is aimed at repetitive, high-volume work — summarising documents, querying databases, classification — plus speed-sensitive jobs such as live customer support and operating a web browser on a user’s behalf. Anthropic estimates the average saving at about 75%, since roughly 90% of requests to the previous Haiku fell under the 100,000-token line and the new model splits text into slightly more tokens.
Also readUN Security Council to hear from OpenAI, Anthropic and DeepSeek on AI risks
Key facts
- Prompts above 100,000 tokens are billed at $0.50 per million input tokens and $2.50 per million output tokens, a 50% cut versus the higher tier of Haiku 4.5.
- On the offline subset of OSWorld 2.1, a computer-operation test, Haiku 5.5 scored 72.4% against GPT-6 Luna’s 48.9%; on Terminal-Bench 4.0, an agentic coding evaluation, it scored 39.2% against 16.4% for Luna and 0% for Haiku 4.5.
- Anthropic halved Sonnet 5.5’s cache-read price to $0.10 per million tokens and is rolling out monthly API credits this week: $100 for Max 5x, $200 for Max 20x, and up to $500 pooled across users on Team plans.
- Haiku 5.5 is the first model in the Haiku family with an adjustable effort setting, letting users trade cost against reasoning capability.
- It ships as claude-haiku-5-5 on the Claude website, Amazon Web Services, Google Cloud and Microsoft Azure.
Where it sits in the Claude 5.5 rollout
Haiku 5.5 arrived 15 days after Opus 5.5 on September 22 and nine days after Sonnet 5.5 on September 28 — the third and last of the Claude 5.5 models Anthropic had promised. Opus 5.5 was the first release since chief executive Dario Amodei published an essay urging the industry to slow gains in AI capabilities.
Anthropic’s own evaluations put Haiku 5.5 well behind Sonnet 5.5 on coding work. Sonnet 5.5 scored 70.6% on Terminal-Bench 4.0, the same test where Haiku landed at 39.2%. Cryptobriefing reported that Anthropic said Sonnet 5.5 and Opus 5.5 remain better suited to complex coding tasks, with Haiku intended for narrower assignments that can also run alongside the larger models as a coding subagent. On GDPval-AA v2.1, which rates models on professional work across 44 occupations on an Elo scale, Haiku 5.5 scored 1620 against 1437 for Luna and 735 for Haiku 4.5.
Also readAnthropic's IPO Draft Shows $42B Loss and $2T Valuation Target
The two reports diverge on the OSWorld 2.1 comparison for Haiku 4.5. Decrypt attributed the offline subset figure to Haiku 5.5 at 72.4% against GPT-6 Luna, while Cryptobriefing reported the same 72.4% for Haiku 5.5 against 15.7% for its predecessor — a predecessor figure and framing that Decrypt’s account of that test does not carry. Cryptobriefing also reported the higher-tier pricing for prompts above 100,000 tokens, a detail Decrypt described only as a 50% cut.
Anthropic is additionally adding beta support for computer and browser use to its Python and TypeScript software development kits, per Cryptobriefing.
Why it matters
The price cut moves the entry point for Anthropic’s smallest model level with OpenAI’s competing small model, which matters most to developers running summarisation, classification and support automation at scale, where token costs compound across millions of requests. It also gives Anthropic a cheaper routing option inside its own stack: the adjusted Sonnet cache-read price and the new monthly API credits lower the cost of building agents that lean on cached context, a pattern the company says accounts for a large share of token consumption on agentic workloads.
What to watch
The monthly API credits are scheduled to roll out this week, so the first signal is whether developers report the credits and the new cache-read pricing landing as described. After that, watch independent evaluations of Haiku 5.5 on agentic coding, where Anthropic’s own numbers show it behind Sonnet 5.5.
Sources: Decrypt, Cryptobriefing
