The model completes the Claude 5.5 family, sitting below Sonnet 5.5 and Opus 5.5. Anthropic lists it through the Claude API, AWS, Google Cloud and Microsoft Azure and Foundry, and pitches it at summarisation, database queries, classification, live customer support and coding subagents. IT Voice published the launch details with the full price sheet.
The price list
| Model and prompt size | Input per 1M tokens | Output per 1M tokens |
|---|---|---|
| Claude Haiku 4.5 | $1.00 | $5.00 |
| Claude Haiku 5.5, prompt up to 100k | $0.10 | $0.50 |
| Claude Haiku 5.5, prompt over 100k | $0.50 | $2.50 |
| GPT-6 Luna | $0.10 | $0.50 |
| Mistral Large 4 | $1.36 | $4.18 |
Two thresholds decide the bill. Up to 100,000 tokens of prompt you pay the low tier. Above it, input moves to $0.50 and output to $2.50 per million. Cache reads run at $0.01 per million under the short tier and $0.05 above it. Anthropic also halved Sonnet 5.5 cache-read pricing to $0.10 per million, which matters for anyone running retrieval loops that resend the same context.
In euros, at roughly $1.10 to the euro, the short-tier rates come out near €0.09 per million input tokens and €0.45 per million output tokens. That conversion is our own arithmetic on today's rate, not an Anthropic figure.
Run it on a mid-sized workload: 10 million input tokens and 2 million output tokens a month, all under the threshold. Haiku 4.5 billed $10 for input and $10 for output, so $20, about €18. Haiku 5.5 bills $1 and $1, so $2, about €1.80. Anthropic's own claim is a 75% reduction on typical workloads, lower than this example because longer prompts, cache behaviour and higher output mixes pull the average up.
Benchmarks against Haiku 4.5
| Benchmark | Haiku 4.5 | Haiku 5.5 |
|---|---|---|
| OSWorld 2.1 (computer use) | 15.7% | 72.4% |
| Terminal-Bench 4.0 (agentic coding) | 0.0% | 39.2% |
| Humanity's Last Exam, no tools | 10.2% | 45.9% |
| Humanity's Last Exam, with tools | 18.7% | 57.4% |
Terminal-Bench is the number to read slowly. A score of 0.0% for Haiku 4.5 usually means the model never completed a task in the harness rather than that it just missed the bar. Haiku 5.5 reaches 39.2% at maximum effort, and Anthropic has not published the same figure at lower effort settings. Roughly six in ten agentic coding tasks still fail.
Where it sits in the field
GPT-6 Luna arrived on 7 October at the same $0.10 and $0.50 rates, so the floor for small frontier models has settled at one price point. Google has not published Gemini 4 Argon pricing. Mistral Large 4 launched on 6 October at $1.36 and $4.18, discounted to $0.68 and $2.09 at launch, which is a different size class and about fourteen times the base Haiku 5.5 input rate. DeepSeek-V4.1-Flash uses variable API rates with a 50% off-peak discount. GLM-5.3 is open weights under MIT, so there the comparison is your own GPU time, which is the kind of thing we measure on the AI Arena rig with local models alongside cloud APIs.
Effort setting and cybersecurity filters
The adjustable effort control is new for the Haiku tier. It lets a developer spend more compute on a hard classification or code task and less on a bulk summarisation run, which changes the cost curve per request rather than per token. Anthropic also built in specific handling for narrow cybersecurity-related requests. The company describes it as a safeguard rather than a refusal layer and has not published the exact trigger conditions, so security teams that red-team models should test it themselves before relying on it.
What European teams should check
The AI Act's Article 50 transparency duties and the EU AI Office's governance powers for general-purpose models became binding in August 2026. A customer-support bot running on Haiku 5.5 is in scope. Users have to be told they are talking to a machine, and the provider has to hold the documentation regulators can request. Cheaper tokens do not change that obligation, and the cost saving does not offset a transparency failure.
Data residency depends on the route you take. There is no separate European launch listed for Haiku 5.5; access goes through the Claude API and the partner clouds. Before signing off a data-flow review, confirm which AWS, Google Cloud or Azure region actually serves the request, because the model's availability follows the regional catalogue rather than the model announcement.
Anthropic also added monthly API credits for Claude Max and Team subscribers. Euro-denominated plan pricing is not part of the launch material, so budget on the token rates above.
Is Haiku 5.5 meant to replace Sonnet 5.5?
No. It sits below Sonnet 5.5 in the same Claude 5.5 family and is aimed at high-volume work and subagent roles, while Sonnet and Opus cover heavier reasoning. The cheaper cache-read pricing on Sonnet 5.5 suggests Anthropic expects both tiers to run in the same pipeline.
When do the higher Haiku 5.5 rates apply?
The price sheet splits at 100,000 tokens of prompt. Up to that point you pay $0.10 per million input and $0.50 per million output. Above it, input goes to $0.50 and output to $2.50 per million tokens.
Can I keep data inside the EU?
Anthropic did not announce a separate EU deployment for this model. Requests are served through the Claude API and the AWS, Google Cloud and Azure catalogues, so residency depends on the region your account is routed to. Check that region against your own GDPR documentation before moving production traffic.
