Almost exactly a year ago, Anthropic released Claude Haiku 4.5, a compact, affordable, and fast alternative to Sonnet and Opus. Now there’s finally a fully updated lineup in the Claude series: Opus 5.5 for ambitious tasks, Sonnet 5.5 for slightly less complex work, and Haiku 5.5 as a fast executor, for both humans and other agents.
Haiku 5.5 is significantly cheaper than 4.5: each million input tokens costs $0.10, and output tokens are 50 cents per million. That’s on par with the pricing of GPT-6 Luna, Haiku’s OpenAI equivalent. Mistral Small 4, released the day before yesterday, is slightly more expensive at $0.15 per million input tokens and 60 cents for outputs.
That does call for some nuance, however. Anthropic claims that Haiku 5.5 is actually 90 percent cheaper than Haiku 4.5, but this is only true for requests up to 100,000 tokens. Anyone exceeding that limit ends up with an option that is 50 percent cheaper. In principle, the performance improvements should be enough to make this price advantage merely a bonus.
Price war
The price war extends beyond the introduction of the compact Claude Haiku model. Sonnet 5.5 becomes significantly cheaper thanks to a 50 percent discount on cache reads. In other words, if a token has already been used recently, Sonnet can retrieve this information at a lower cost. For long-running agentic tasks, this accounts for the lion’s share of tokens, since each token is reprocessed after a tool call, output, or new prompt.
More importantly, with Haiku and the more affordable Sonnet, Anthropic has a “neater” pricing structure. Anyone looking to “burn” tokens can always run Opus 5.5 at maximum effort, with the only drawback being slower output compared to lower effort levels or smaller models. Anyone with a token or usage limit will be more than happy to use Sonnet or Haiku whenever possible.
Subagents, agents that perform tasks on behalf of other agents or do so in parallel for an end user, give Haiku 5.5 a much stronger case for existence compared to Opus and Sonnet than was the case a year ago. At that time, such delegation and orchestration were not yet as mature as they are now.
The reality of agreements
Many organizations have now opted for Claude. Although individual users and API users can, in fact, always switch models, a Claude Team or Enterprise customer is bound by Anthropic’s model selection. This means that Haiku 5.5 fills a gap that was gaping wide throughout 2026. The reality of contractual agreements is that users are now building on Claude more quickly and thereby allowing themselves to be lured into a lock-in. This can be many times more cost-effective than using only the API. Nevertheless, this is Anthropic’s strategy: to turn its own product into a platform that goes beyond just the models. Haiku 5.5 is an important component of that.
Read also: With Mistral Large 4, Europe is (somewhat) competitive again