Claude Haiku 5.5: Anthropic Cuts Prices by Up to 90 Percent

Set Trending Topics as a preferred source on Google.

The A.I. price war has reached the small models: Anthropic has just introduced Claude Haiku 5.5, the company’s cheapest and fastest small model yet. For most requests, it costs only a tenth of its predecessor, Haiku 4.5, while promising to do far more. That puts Anthropic level on price with OpenAI’s small GPT-6 Luna model, just weeks before its planned I.P.O.

A Tenth of the Price, With Catches

Through the API, Haiku 5.5 costs $0.10 (about 9 euro cents) per million input tokens and $0.50 (about 45 euro cents) per million output tokens. Haiku 4.5 cost $1 and $5, respectively. The low rate applies only to requests of up to 100,000 tokens, however. Above that, the price rises to $0.50 for input and $2.50 for output per million tokens. According to Anthropic, about 90 percent of previous Haiku requests fall below that threshold.

A second catch is in the fine print: A new tokenizer counts the same text as roughly 30 percent more tokens. On balance, Anthropic therefore promises costs about 75 percent lower for typical workloads.

The key facts:

• Context: up to one million tokens, with up to 128,000 output tokens.
• Effort: It is the first Haiku model with adjustable effort, with medium as the default.
• Availability: through the Claude API as well as Amazon Bedrock, Google Cloud and Microsoft Foundry.
• Safety: stricter cybersecurity rules than its predecessor, with penetration testing blocked by default.

A Big Leap on Agent Tasks

According to benchmarks published by Anthropic, the gap to its predecessor is huge. On OSWorld, a test in which models operate a computer, Haiku 5.5 scores 72.4 percent, compared with 15.7 percent for Haiku 4.5 and 48.9 percent for GPT-6 Luna. On the coding test Terminal-Bench, Haiku 5.5 reaches 39.2 percent versus 16.4 percent for Luna, though only at maximum effort. At the default setting, it scores about 20 percent. The larger Sonnet 5.5 remains ahead on every test. The figures have not yet been independently verified.

Along with the launch, Anthropic is also cutting the price of cached input for Sonnet 5.5 from $0.20 to $0.10 per million tokens. Subscribers to the Max and Team plans also get monthly API credits, ranging from $100 (about €89) for Max 5x to $500 (about €445) for a team.

A Price War Before the I.P.O.

Competition among small models is especially fierce. According to VentureBeat, GPT-6 Luna costs exactly the same as Haiku 5.5, Google’s Gemini 3.5 Flash-Lite comes in at $0.30 for input and $2.50 for output, and SpaceXAI’s Grok 4.3 at $1.25 and $2.50. Then there are Chinese open-weight models such as DeepSeek, which are often even cheaper.

The launch comes at a delicate moment for Anthropic. The company is reportedly preparing an I.P.O. for mid-November, and on Wall Street, many are waiting for the listing. Haiku 5.5 is the third new model in the 5.5 generation within a few weeks, after Opus 5.5 and Sonnet 5.5. Whether the price cut pays off depends on how much additional usage the cheaper model brings in.