On October 7, Anthropic released Claude Haiku 5.5, the fourth generation of its small, fast model line — and the most aggressive price cut the company has made on a flagship-adjacent model. Input tokens that cost $1.00 per million under Haiku 4.5 now cost $0.10 for prompts under 100,000 tokens, a 90% reduction. Output pricing fell from $5.00 to $0.50 per million tokens under the same threshold. Anthropic frames the release as proof that “cheap” and “capable” no longer trade off against each other the way they used to.
The Numbers Behind the Headline
Anthropic’s own benchmark table, published alongside the release, puts Haiku 5.5 well ahead of its predecessor and ahead of OpenAI’s GPT-6 Luna on every test where both models have a published score. On OSWorld 2.1 — a benchmark that measures how well a model can actually operate a computer, clicking through real desktop and browser tasks rather than just answering questions — Haiku 5.5 scores 72.4%, compared with 15.7% for Haiku 4.5 and 48.9% for GPT-6 Luna. On Terminal-Bench 4.0, which tests command-line task completion, Haiku 5.5 hits 39.2% against Haiku 4.5’s 0.0% and GPT-6 Luna’s 16.4%. On Humanity’s Last Exam without tool use, Haiku 5.5 scores 45.9%, versus 10.2% for its predecessor.
Those are real gaps, not rounding. But Anthropic is careful not to oversell Haiku’s ceiling: its own flagship, Claude Sonnet 5.5, still beats Haiku 5.5 on every one of those same benchmarks — 83.9% on OSWorld 2.1, 70.6% on Terminal-Bench 4.0, 56.9% on Humanity’s Last Exam. The company’s pitch isn’t “Haiku now matches the flagship.” It’s “Haiku now does flagship-class work on the narrower set of tasks — classification, summarization, structured extraction, customer support, subagent orchestration — that make up the bulk of real production API traffic, at a fraction of the cost.”
What Actually Changed
Three things separate Haiku 5.5 from a routine model refresh. First, it’s the first Haiku-class model to ship with an adjustable effort setting — the same mechanism Anthropic introduced further up its lineup, letting developers trade latency for accuracy on a per-request basis rather than being locked into one fixed behavior. For a high-volume, latency-sensitive workload like live chat triage, that’s a real lever, not a marketing bullet point.
Second, Anthropic says Haiku 5.5 is now its fastest model at standard inference speed — faster than Haiku 4.5, and by extension faster than every other model in the current Claude lineup. Speed matters disproportionately for the use cases Anthropic is explicitly targeting with this release: summarization and context compaction inside longer agent sessions, database and API queries, high-frequency classification, and running as the “subagent” that a more expensive model like Sonnet 5.5 or Opus 5.5 delegates narrow subtasks to mid-task. Anthropic’s pricing page lists exactly that pattern — Haiku 5.5 paired as a cheaper worker model behind Devin Fusion, a third-party coding agent, where it posted a 66.2 FrontierCode score as the supporting model rather than the lead.
Third — and this is the part that matters beyond the spec sheet — Anthropic reports substantially fewer misaligned behaviors in its internal alignment evaluations for Haiku 5.5 compared with Haiku 4.5, alongside tighter, though not maximal, cybersecurity safeguards. Penetration-testing requests remain explicitly blocked. For a company racing to ship a cost-optimized model into high-volume production traffic, publishing alignment movement in the same breath as a price cut is a deliberate signal: the cheap model isn’t the corner they’re cutting.
Why Anthropic Is Competing on Price Right Now
The timing isn’t incidental. Haiku 5.5 landed in the same month Anthropic has been reported to be preparing marketing and model announcements ahead of a potential IPO, and in the same stretch where OpenAI, Google DeepMind, Mistral, and a wave of Chinese labs — Moonshot’s Kimi, Zhipu’s GLM, MiniMax — have all been shipping new models at a pace few customers can realistically track. In that environment, raw benchmark leadership is a moving target measured in weeks. Cost-per-task is stickier: once a company has built a classification pipeline or a customer-support triage layer against a model priced at $0.10 per million input tokens, the switching cost to move to a competitor isn’t just technical, it’s a line item a finance team has to re-approve.
Anthropic’s own documentation makes the cost argument concrete with customer-reported numbers rather than synthetic benchmarks: Box reported Haiku 5.5 scoring 11 points higher than Haiku 4.5 on its internal AI evaluation at roughly half the latency, and HubSpot reported a 92.8% average across three runs on its CRM task suite. Those are the kinds of numbers that move procurement decisions, even if they never show up on a public leaderboard.
Where to Get It
Haiku 5.5 is live now across Anthropic’s API under the model ID claude-haiku-5-5, and through all three major cloud platforms — Amazon Bedrock, Google Cloud’s Vertex AI, and Microsoft Azure — meaning enterprise customers already locked into one of those procurement relationships can switch models without switching vendors. Anthropic’s Python and TypeScript SDKs add beta support for computer-use and browser-use tool calling alongside the release. Claude Max, Team, and Ultra subscribers get a monthly API credit pool that can be spent on Haiku 5.5 or any other model in the current lineup, which is Anthropic’s way of nudging existing subscribers to try the cheaper model on workloads they might previously have thrown at Sonnet by default.
One footnote worth flagging for anyone doing their own cost math: Anthropic’s release notes mention that Haiku 5.5 uses a slightly different tokenizer than its predecessor, and tends to use somewhat more tokens to complete the same task. The headline 90% discount is on the per-token rate, not automatically on the per-task bill — worth a quick before/after measurement on your own workload rather than taking the percentage at face value.
The Bigger Picture
Haiku 5.5 won’t be the model anyone points to as evidence of an intelligence breakthrough — that race, this cycle, belongs to the flagship-tier releases from Anthropic, OpenAI, and Google DeepMind. But it’s arguably the more consequential release for how AI actually gets deployed day to day: the overwhelming majority of real-world API traffic is not frontier reasoning, it’s routine classification, extraction, summarization, and tool orchestration running millions of times a day. A model that does that work competently at one-tenth the previous cost changes the unit economics of every product built on top of it, which is a less exciting story than a new benchmark record, but a far more durable one.
