Anthropic released Claude Haiku 5.5 on October 7, and the headline is the price. For prompts up to 100,000 tokens, the model costs $0.10 per million input tokens and $0.50 per million output tokens, one tenth of what Haiku 4.5 charged. Anthropic calls it its cheapest, fastest and most capable small model, aimed at high-volume work such as summaries, classification and live customer support, not at the hardest coding tasks.
What it costs
Tokens are the chunks of text a model reads and writes, and API prices are quoted per million of them. According to Anthropic’s announcement, Haiku 5.5 is priced at $0.10 for input and $0.50 for output on prompts up to 100,000 tokens, and at $0.50 and $2.50 above that. Haiku 4.5 charged $1 and $5, and Sonnet 5.5 charges $2 and $10. Cache reads, the cheap rate for text the model has already seen, cost $0.01 per million tokens on short prompts.
Anthropic says about 90% of requests to the previous Haiku fell under the 100,000-token line. It also says the model costs around 75% less to run on average, a smaller figure than the headline cut because Haiku 5.5 uses a new tokenizer that consumes slightly more tokens per task. As a rough illustration, summarizing 10,000 documents of 5,000 tokens each is 50 million input tokens: about $5 on Haiku 5.5 and $50 on Haiku 4.5, before output costs.
OpenAI’s smallest model, GPT-6 Luna, has the same list price of $0.10 and $0.50 according to OpenAI’s pricing table, so the competition is now about quality and speed, not sticker price.
What the benchmarks say
Every figure below comes from Anthropic’s own table and has not been independently reproduced. On OSWorld 2.1, an offline test of an agent operating a real computer, Haiku 5.5 scores 72.4%, against 15.8% for Haiku 4.5 and 48.9% for GPT-6 Luna. On Terminal-Bench 4.0, which measures command-line work, it reaches 39.2% against 16.4% for Luna. On FrontierCode 1.1 it scores 46.4%, ahead of Luna’s 42.4%.
The same table shows the limits. Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0 and 83.9% on the computer-use test. Anthropic itself says Sonnet 5.5 and Opus 5.5 remain the better choices for complex agentic coding, and that Haiku 5.5 suits narrowly scoped jobs that were previously too expensive to automate, such as compaction, summarization and subagent work.
New features and availability
Haiku 5.5 is the first Haiku model with an adjustable effort setting, so developers can trade intelligence for cost, as with Anthropic’s larger models. It is available through the Claude Platform as claude-haiku-5-5, and through Amazon Web Services, Google Cloud and Microsoft Azure. Anthropic published a migration guide and a system card with the evaluation details.
Safeguards are tiered. Anthropic says the cybersecurity protections are stricter than those on Haiku 4.5 but looser than on its other recent models, and that they still block penetration testing and other attacker-oriented techniques. Biology safeguards match those of Sonnet 5, Sonnet 5.5 and Opus 5.
Other changes in the announcement
Anthropic also halved the cache-read price of Sonnet 5.5, from $0.20 to $0.10 per million tokens, which it says lowers the cost of most agent tasks on that model by around 20%. Max and Team subscribers get a new monthly credit for the Claude Platform: $100 for Max 5x, $200 for Max 20x and up to $500 pooled for Team plans, described in a Help Center article. The Python and TypeScript SDKs also gained beta support for computer use and browser use.
What this means for you
If you pay for an API at volume, run your own test set on Haiku 5.5 before switching, and compare cost per completed task, not cost per token. A cheaper model that needs retries can lose its advantage. If your prompts regularly exceed 100,000 tokens, use the higher tier in your estimate.
If you use Claude only through the chat apps, nothing changes directly. The practical effect shows up in products built on the API, where cheaper models make features like live support, document search and background agents easier to offer. For complex coding, long reasoning or sensitive security work, the larger models are still the better fit.
