CLAUDE HAIKU 5.5 IS ANTHROPIC'S FASTEST MODEL YET

Claude Haiku 5.5, released by Anthropic on October 7, 2026, is priced at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens and $0.50 and $2.50 above that, matching OpenAI's GPT 6 Luna on the lower tier. Anthropic says it costs around 75 percent less to run than Haiku 4.5 on average, is its fastest model to date, and is the first Haiku with an adjustable effort setting, with medium as the default. It scored 72.4 percent on the offline subset of OSWorld 2.1 against 15.7 for Haiku 4.5 and 39.2 percent on Terminal Bench 4.0 at maximum effort, with medium near 20 percent. Simon Willison measured about 1.25 times as many tokens on the same long prompt as Haiku 4.5 because of the newer tokenizer, and Haiku 5.5 is the fourth Claude price cut since September 1, 2026.
Key Points
- Claude Haiku 5.5 launched October 7, 2026, at $0.10 per million input tokens and $0.50 per million output for prompts up to 100,000 tokens, matching OpenAI's GPT 6 Luna, and rises to $0.50 and $2.50 above that line.
- Anthropic's average saving over Haiku 4.5 is 75 percent rather than the 90 percent sticker cut, because the newer tokenizer counts about 1.25 times as many tokens on the same prompt, as Simon Willison measured on launch day.
- The default effort is medium, which scores near 20 percent on Terminal Bench 4.0 against the 39.2 percent headline at maximum effort, and Sonnet 5.5 still leads at 70.6 percent.
Anthropic cut the sticker on its smallest model by 90 percent on October 7, 2026, and matched OpenAI's GPT 6 Luna to the cent, 15 days after Luna launched on September 22. Anthropic's own figure for what a real workload saves is 75 percent, not 90. The gap between those numbers is the story, because three things in the fine print, a tokenizer that counts about 1.25 times as many tokens on the same prompt, a price cliff at 100,000 tokens, and a medium default effort that roughly halves the headline coding score, decide what Claude Haiku 5.5 costs on a random Tuesday.
The thesis: the cheapest small model Anthropic has ever released is also the only current Claude priced by how long your prompt is, after four Claude price cuts in 36 days.
90 Percent Off Stops at 100,000 Tokens
Claude Haiku 5.5 lists at $0.10 per million input tokens and $0.50 per million output for any request whose prompt is 100,000 tokens or under, against $1 and $5 on Haiku 4.5. Over 100,000 tokens the rate jumps to $0.50 in and $2.50 out, five times the lower tier and half of Haiku 4.5.
The pricing page states the catch. A request's prompt length counts every input token, cache reads and cache writes included, and a request over the line pays the higher prices even when part of its prompt is a cache hit. Every other Claude model from 4.6 onward bills the full 1 million token window at one flat rate, a structure that held even when Fable 5.1 cut its cache reads 75 percent on September 1; the pricing page lists Haiku 5.5 as the single exception. Carl Franzen at VentureBeat reported Anthropic's estimate that about 90 percent of Haiku 4.5 requests fall under 100,000 tokens.
OpenAI's Luna matches all four of the lower tier rates, input, output, cache write and cache read, so the lock in score below the line is zero: one model string, same sticker. Simon Willison's read on October 7: "Above 100,000 tokens, Luna looks like a much better deal."
Simon Willison Found the Hidden Price Increase
Claude models from 4.7 onward run on a newer tokenizer that Anthropic's pricing documentation says produces approximately 30 percent more tokens for the same text, and Haiku 5.5 is the first Haiku built on it. Haiku 4.5, released almost a year earlier, predates the switch, so a per token price cut lands on a bigger token count.
Willison measured it on launch day on his blog: "My Claude Token Counter tool shows that the same long prompt uses around 1.25x as many tokens with Haiku 5.5 compared to Haiku 4.5, so there's a hidden price increase there." Franzen's reporting confirms the 75 percent average already bakes in the updated tokenizer and the mix of request sizes.
Willison also ran one pelican prompt at every effort level. Low effort cost 0.0936 cents and took 7 seconds. Max effort took 5 minutes 9 seconds and cost 3.3826 cents, roughly 36 times the price on the same prompt.
Medium Is the Default, and It Roughly Halves the Coding Score
Haiku 5.5 ships with effort set to medium on the Claude API, with five stops from low to max. Anthropic's effort chart, as Franzen read it for VentureBeat, places the 39.2 percent Terminal Bench 4.0 score at maximum effort, with medium landing near 20 percent.
Opus 5 shipped with the same dial on July 24, 2026, at three settings, low, medium and high, and Haiku 5.5 is the first Haiku class model to get one. Anthropic's table: 72.4 percent on the offline subset of OSWorld 2.1 against 15.7 for Haiku 4.5 and 48.9 for GPT 6 Luna, 45.9 percent on Humanity's Last Exam with no tools against 10.2, and a GDPval AA v2.1 score of 1620 against 735 for Haiku 4.5 and 1437 for Luna. Sonnet 5.5 still leads every row, at 83.9 on OSWorld and 70.6 on Terminal Bench, and Anthropic's page sends complex agentic coding to Sonnet 5.5 and Opus 5.5.
Every figure above is vendor reported; the customer quotes name a basis. Aaron Vinh, a staff software engineer at Asana, said on the launch page, "We're very impressed with Claude Haiku 5.5, particularly its speed," and put the gain at over a 30 percent reduction in latency for task completions and up to 2.5 times faster inference per agent turn against the model Asana uses today.
Count the Price Cuts Since September 1
Haiku 5.5 is the fourth price cut on the Claude lineup since September 1, 2026. Fable 5.1 cut cache reads 75 percent to $0.25 per million on September 1, Opus 5.5 arrived September 22 at $4 and $20 per million, 20 percent under Opus 5, and on October 7 Sonnet 5.5 cache reads were halved to $0.10 alongside the Haiku launch, which Anthropic says makes Sonnet about 20 percent cheaper on most agentic work.
OpenAI set the pace. Its small tier launched July 9 as GPT 5.6 Luna at $1 in and $6 out; GPT 6 Luna, launched September 22, lists at $0.10 and $0.50, and Haiku 5.5 matched them 15 days later. It is also the third Claude 5.5 model in 15 days, after Opus 5.5 on September 22 and Sonnet 5.5 on September 28.
Max 5x subscribers get $100 in API credits a month, Max 20x get $200, and Team plans get up to $500 pooled across users. Willison notes the credits do not roll over.
Try Haiku 5.5 for anything that stays under 100,000 tokens at the default effort, and count a week of real tokens against the Haiku 4.5 bill before believing 90. Skip max effort for volume work; the 5 minute 9 second pelican is the whole argument. The invoice to watch is the first one after a cached conversation crosses 100,000 tokens, because Anthropic's pricing page says that request pays the higher rate on every token in it, cache hits included.
Frequently Asked Questions
When was Claude Haiku 5.5 released?
Anthropic released Claude Haiku 5.5 on October 7, 2026, 15 days after Opus 5.5 on September 22 and nine days after Sonnet 5.5 on September 28.
How much does Claude Haiku 5.5 cost?
It costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, and $0.50 and $2.50 per million for prompts over 100,000 tokens. Anthropic says it runs around 75 percent cheaper than Haiku 4.5 on average.
Is Claude Haiku 5.5 cheaper than GPT 6 Luna?
Below 100,000 tokens the two list identical rates for input, output, cache writes and cache reads. Above 100,000 tokens Haiku 5.5 charges five times its lower tier, and Simon Willison judged Luna the better deal there.
Does Claude Haiku 5.5 have an effort setting?
Yes, it is the first Haiku class model with an adjustable effort setting, with five levels from low to max and medium as the default on the Claude API.
How does Claude Haiku 5.5 compare to Haiku 4.5 on benchmarks?
Anthropic reports 72.4 percent on the offline subset of OSWorld 2.1 against 15.7 for Haiku 4.5, 39.2 percent on Terminal Bench 4.0 against 0.0, and 45.9 percent on Humanity's Last Exam without tools against 10.2.
What is the Claude Haiku 5.5 model ID?
The model ID is claude-haiku-5-5 on the Claude API, Google Cloud and Microsoft Foundry, and anthropic.claude-haiku-5-5 on Amazon Bedrock.
Topics: claude-haiku-5-5, ai-models, benchmarks, openai, anthropic, claude