Claude Haiku 5.5: a $0.10 model that beats GPT-6 Luna on Anthropic's own tests

Claude Haiku 5.5 is Anthropic's newest small model, released on October 7, 2026 under the API name claude-haiku-5-5. Anthropic calls it its cheapest, fastest and most capable small model. The price is the headline: from $0.10 per million input tokens and $0.50 per million output tokens, down from $1.00 and $5.00 for Haiku 4.5. The capability claims are the part that needs checking, because every number below comes from Anthropic itself.

What Anthropic published

The model page and the developer docs agree on the basics. Haiku 5.5 has a 1M token context window, a 128K token maximum output and a June 2026 knowledge cutoff. Anthropic says it will not retire the model sooner than October 7, 2027. Pricing is tiered by prompt size: for prompts up to 100,000 tokens, input is $0.10 and output $0.50 per million tokens; above 100,000 tokens, input is $0.50 and output $2.50. Cache reads cost $0.01 per million tokens at the lower tier.

It is also the first Haiku-class model with an adjustable effort setting, which lets a developer trade speed and cost against depth of reasoning. The docs set the default at medium. Anthropic says the model costs about 75% less on average than Haiku 4.5. That is a smaller drop than the 90% list-price cut, and Anthropic does not explain the gap in the page we read.

It is available on the Claude Platform, Amazon Web Services, Google Cloud and Microsoft Azure. Max and Team subscribers also get monthly API credits, from $100 on Max 5x up to a pooled $500 on Team.

The benchmark table, read carefully

Anthropic compares Haiku 5.5 with Haiku 4.5, Sonnet 5.5 and OpenAI's GPT-6 Luna. The jumps over Haiku 4.5 are enormous, partly because the older model was built for an easier era of tests.

• GDPval-AA v2.1 (Elo): Haiku 5.5 scores 1620, against 735 for Haiku 4.5, 1437 for GPT-6 Luna and 1840 for Sonnet 5.5.
• OSWorld 2.1, offline subset: 72.4%, against 15.7% for Haiku 4.5, 48.9% for Luna and 83.9% for Sonnet 5.5.
• Humanity's Last Exam, no tools: 45.9%, against 10.2% for Haiku 4.5 and 56.9% for Sonnet 5.5.
• Terminal-Bench 4.0: 39.2%, against 0.0% for Haiku 4.5, 16.4% for Luna and 70.6% for Sonnet 5.5.
• FrontierCode 1.1 (Main): 46.4%, against 42.4% for Luna and 52.1% for Sonnet 5.5 at its Xhigh setting.

Two readings follow. First, the Luna comparison favors Haiku on every row Anthropic chose to publish. That is normal for a vendor table and tells you little, since a lab picks the tests it wins. Second, the gap to Sonnet 5.5 is widest exactly where agents do long, multi-step work. A 31-point shortfall on Terminal-Bench 4.0 is the difference between a model that finishes a command-line task and one that stalls halfway.

Our earlier coverage of Sonnet 5.5, which launched on September 28, 2026 at $2 input and $10 output per million tokens, is the right yardstick. Haiku 5.5 costs a twentieth of Sonnet 5.5 on input. For work where a wrong answer is cheap to catch, such as classification, extraction and routing, which is how Anthropic's own docs position it, that trade is attractive.

Safety limits that matter for the race

The more revealing section is about what the model will refuse. Anthropic says Haiku 5.5 shows far fewer misaligned behaviors than Haiku 4.5 and is less willing to cooperate with misuse. Its cybersecurity safeguards are stricter than Haiku 4.5's but looser than Sonnet 5.5's, and they still block penetration testing and attacker-oriented techniques. Its biology safeguards match those on Sonnet 5, Sonnet 5.5 and Opus 5: research questions are allowed, likely-harmful requests are restricted.

That tiering fits a pattern. Anthropic now gates its cyber-capable models by who is asking, as the new three-tier scheme we covered in Anthropic's expanded Cyber Verification Program shows. A very cheap model is also a very cheap tool for misuse at scale, so the safeguards matter more as the price falls. The announcement does not say how Anthropic tested those safeguards, and the page points to a system card for detail. We did not find the system card text this run, so treat the safety claims as unverified until outside evaluators publish.

What customers say, and what is missing

Anthropic includes two customer lines. Asana's Aaron Vinh, a staff software engineer, called the experience "noticeably snappier." Box's head of AI products reported an 11-point gain over Haiku 4.5 at about half the latency, according to Anthropic's page. Both are vendor-selected testimonials, useful for direction and not for measurement.

What is missing is an independent score. No third party had published a Haiku 5.5 evaluation when we checked on October 8, 2026. The one search result that compared Haiku with Luna was an aggregator page that listed older model names and could not be used.

Our take

Haiku 5.5 is a real price cut and a plausible upgrade for high-volume, low-stakes work. It is not evidence that the frontier moved. The frontier moved in late September with Opus 5.5 and Sonnet 5.5, and Haiku inherits some of that progress at a fraction of the cost. We would test it on your own routing and extraction jobs before switching, and we would hold off on repeating the Luna claim until Artificial Analysis or Epoch AI run the same tests. Watch two things: whether independent scores track the vendor table, and whether the cheap-model safeguards hold up once outside red teams get to them.