On October 7, 2026, Anthropic released Claude Haiku 5.5. The company calls it its fastest and cheapest small model yet, and names high-volume, latency-sensitive work such as summarization, context compression, database queries, and classification as its main uses. Subagents in Claude Code, real-time customer support, and browser use are also listed as use cases.
Two numbers stand out in this announcement. On the official evaluations, several of the 8 benchmark scores rose sharply from Haiku 4.5, and API pricing starts at $0.10 per million input tokens when the prompt is 100K tokens or less. Still, choosing a model means looking at task type and input length together, not at a single ranking line.
We pulled together the key figures from the performance table and the accuracy-versus-cost charts by effort level in Anthropic's official announcement, along with the launch video and performance materials from its official X account.
8 Benchmarks, With Strengths That Vary by Task
Below are the results Anthropic published, in the order Haiku 5.5, Haiku 4.5, GPT-6 Luna, and Sonnet 5.5 as a reference model. Because these are results the company published itself, they are hard to treat as independent third-party reproductions. It is also not appropriate to add up scores measured under different conditions into one overall ranking.
| Official benchmark | Haiku 5.5 | Haiku 4.5 | GPT-6 Luna | Sonnet 5.5 | Condition |
|---|---|---|---|---|---|
| GDPval-AA v2.1 | 1620 | 735 | 1437 | 1840 | Knowledge work |
| AA-Briefcase v1.1 | 1578 | 614 | 1336 | 1824 | Long-horizon knowledge work |
| OSWorld 2.1 | 72.4% | 15.7% | 48.9% | 83.9% | offline subset |
| Humanity’s Last Exam | 45.9% | 10.2% | — | 56.9% | No tools |
| Humanity’s Last Exam | 57.4% | 18.7% | — | 64.5% | With tools |
| Terminal-Bench 4.0 | 39.2% | 0.0% | 16.4% | 70.6% | Agentic coding |
| FrontierCode 1.1 Main | 46.4% | — | 42.4% | 52.1% | Sonnet at Xhigh |
| Chartography | 46.4% | 6.4% | 29.1% | 61.6% | No tools |
The strengths these evaluations show do not all point the same way. On the OSWorld offline subset, Haiku 5.5 scored 72.4%, ahead of Haiku 4.5 at 15.7% and GPT-6 Luna at 48.9%. It also scored much higher than the previous Haiku on GDPval-AA and AA-Briefcase. Sonnet 5.5, on the other hand, scored higher on every item the table has in common, and on Terminal-Bench 4.0 in particular it reached 70.6% against Haiku 5.5's 39.2%.
So a summary like "the small model beat the higher-tier models across the board" would be wrong. The official guidance also recommends Sonnet 5.5 or Opus 5.5 for complex agentic coding like Terminal-Bench, and says Haiku 5.5 delivers value when deployed on narrow, repetitive tasks.
This is the launch post from Claude's official account, including the announcement video.
Effort Control and the 100K-Token Pricing Tier
Haiku 5.5 is the first Haiku model to offer an adjustable effort setting. You can tune it toward lower cost or toward using more intelligence, and Anthropic published accuracy-versus-cost charts by effort level for OSWorld, GDPval-AA, and Humanity’s Last Exam. The balance between performance and cost changes with the setting, so you cannot assume one configuration is best for every call.
You can see the announcement visuals in the official performance chart post and the effort-level chart post.
| USD per 1M tokens | Haiku 5.5 Up to 100K / Over 100K | Haiku 4.5 | Sonnet 5.5 |
|---|---|---|---|
| Input | 0.10 / 0.50 | 1.00 | 2.00 |
| Output | 0.50 / 2.50 | 5.00 | 10.00 |
| Cache read | 0.01 / 0.05 | 0.10 | 0.10 |
| Cache write | 0.125 / 0.625 | 1.25 | 2.50 |
Compared with Haiku 4.5, Haiku 5.5's standard rates are 90% lower for requests up to 100K tokens and 50% lower above 100K. Anthropic's figure of "about 75% lower running cost on average" is not a discount applied uniformly to every request. It is a workload-average estimate that reflects the distribution of earlier Haiku requests, about 90% of which were 100K or less, and the slight increase in tokens needed for the same task under the new tokenizer.
A separate Batch API offers savings of up to 50% depending on conditions. When forecasting usage, it is more accurate to calculate with each request's input length, cache hits, output volume, and effort setting.
Customer Figures Are Each Company's Early Tests
The customer examples in Anthropic's announcement are useful, but they are not shared external evaluations. They are results each customer measured on its own product with its own evaluation method.
- Asana: Compared with its current model, it reported task completion latency down by more than 30% and up to 2.5x faster reasoning per agent turn.
- HubSpot: It recorded a 3-run average of 92.8% on its own CRM evaluation.
- AlphaSense: It scored 0.84 against Haiku 4.5's 0.76 on 400 queries, which it described as a statistically significant improvement. The feature is called about 8 million times a week.
- Box: It reported a score 11 points higher than Haiku 4.5 and roughly half the latency. The quoted description does not state the scoring scale.
- Rogo: It gave a qualitative assessment that the model suits short, high-volume tasks such as quick lookups, summaries, and subagents.
- Cognition/Devin Fusion: It said pairing Haiku 5.5 as a sidekick kept FrontierCode at 66.2 while reducing cost and latency.
These results can be read as early signals about real product workloads, but they are no guarantee that the same metrics and conditions will reproduce at every organization. HubSpot's score, Box's 11 points, and Devin Fusion's 66.2 in particular should be kept in the context of each company's own published evaluation.
Which Tasks Fit, and What to Verify Further
Haiku 5.5 is worth considering when you run many small, fast calls, such as summarization, classification, routing, context compression, real-time responses, browser tasks, and subagents for coding led by a higher-tier model. Its standard API rates are lower than processing the same work with Haiku 4.5, and the improvement on some public benchmarks was large too.
For complex multi-step coding and agent work, Sonnet 5.5 or Opus 5.5 may be the better choice. Raising effort changes accuracy and cost together, so it is practical to first run a small evaluation of your own that reflects real input lengths and success criteria.
On safety, Anthropic said most alignment evaluations improved over Haiku 4.5 and that misaligned behavior and a tendency to assist with misuse decreased. Its cyber safeguards are more restrictive than Haiku 4.5's and allow somewhat more defensive work than the latest higher-tier models, but it blocks techniques that are easy to abuse for attacks, such as penetration testing. It says the biology safeguards are the same as for Sonnet 5, Sonnet 5.5, and Opus 5. On the other hand, a problem such as "over-refusing legitimate requests" and any official refusal rate are not confirmed in this announcement, so there is no basis for asserting either.
It is available on Claude.ai Free, Pro, Max, Team, and Enterprise on web and iOS/Android, and in Claude Code and the Claude Platform/API under the model ID claude-haiku-5-5. The cloud providers named are Amazon Bedrock, Google Cloud, and Microsoft Foundry. Regional and deployment details should be checked with each provider.
For examples of using Claude with work tools, see Using Claude with Google Docs, Sheets, and Slides, and for the agent-building tool workflow, continue with the ElevenAgents Architect overview.
The official evaluation table and the pricing and availability details for Claude Haiku 5.5 are on Anthropic's announcement page.