Claude Haiku 5.5 price, benchmarks and the 100K-token catch
Claude Haiku 5.5 price drops 90% to $0.10 and $0.50 per million tokens, matching GPT-6 Luna. What a job costs, whose scores these are, and when Sonnet wins.
Source-based. Written from the documents, reporting and reviews linked in the text. Nothing here was tested hands-on by The Ruling Desk. How we work
Anthropic released Claude Haiku 5.5 on October 7, 2026, and the Claude Haiku 5.5 price is 90% below Haiku 4.5's for most requests: $0.10 per million input tokens and $0.50 per million output tokens, the same list price as OpenAI's GPT-6 Luna. The catch is a second price tier for prompts over 100,000 tokens. Anthropic also reports big benchmark gains, which come from its own table, while the first independent index shows a smaller but real lead over Luna.
Key takeaways
- Prompts up to 100,000 tokens cost $0.10 input and $0.50 output per million tokens, down from $1 and $5 on Haiku 4.5. Longer prompts cost $0.50 and $2.50, only 50% less, per Anthropic's pricing docs as of October 9, 2026.
- Haiku 5.5 counts the same text as about 30% more tokens, so Anthropic estimates real savings at around 75% on average, not 90%.
- Anthropic reports 72.4% on the offline subset of OSWorld 2.1, against 15.7% for Haiku 4.5 and 48.9% for GPT-6 Luna. Those are Anthropic's figures; Artificial Analysis independently scores Haiku 5.5 at 43 and Luna at 38 on its index at max effort.
- It's the first Haiku with adjustable effort levels, from low to max, with medium as the default.
- Sonnet 5.5's cache reads were halved the same day, to $0.10 per million tokens. Anthropic still recommends Sonnet and Opus for complex agentic coding.
What the Claude Haiku 5.5 price means next to Haiku 4.5 and GPT-6 Luna
Anthropic's pricing page lists two sets of Haiku 5.5 prices, split by prompt length. All figures are US dollars per million tokens, as listed on October 9, 2026; the GPT-6 Luna column comes from OpenAI's model page.
| Item | Haiku 4.5 | Haiku 5.5, prompts up to 100K | Haiku 5.5, prompts over 100K | GPT-6 Luna |
|---|---|---|---|---|
| Input | $1 | $0.10 | $0.50 | $0.10 |
| Output | $5 | $0.50 | $2.50 | $0.50 |
| Cache write (5 minutes) | $1.25 | $0.125 | $0.625 | $0.125 |
| Cache read | $0.10 | $0.01 | $0.05 | $0.01 |
| Batch input / output | $0.50 / $2.50 | $0.05 / $0.25 | $0.25 / $1.25 | $0.05 / $0.25 |
Here's our arithmetic at list prices for a job that sends 10 million input tokens and gets back 2 million output tokens, with every prompt under 100,000 tokens: $20 on Haiku 4.5, $2 on Haiku 5.5, $2 on GPT-6 Luna and $40 on Sonnet 5.5.
That $2 needs one correction. Anthropic's model overview says Haiku 5.5 uses a newer tokenizer that turns the same text into about 30% more tokens than Haiku 4.5 did. If your job grows to 13 million input and 2.6 million output tokens, it costs $2.60, still about 87% less than on Haiku 4.5. Anthropic's announcement puts the average saving lower: "On average, it now costs around 75% less to run."
The 100,000-token catch
The higher tier applies to the whole request once its prompt passes 100,000 tokens, and the pricing page says cache reads and writes count toward that total. Anthropic says 90% of Haiku 4.5 requests fell under the line.
If every prompt in the same job ran over 100,000 tokens, it would cost $10 on Haiku 5.5, half of Haiku 4.5's $20. GPT-6 Luna keeps its base price up to 272,000 input tokens, so in that range it costs a fifth of Haiku 5.5 per token. Above 272,000, OpenAI charges 2x on input and 1.5x on output, which puts the same job at $3.50. The two companies count tokens differently, so per-token comparisons across them are only a starting point.
Claude Haiku 5.5 benchmarks: whose numbers are whose
Every score below comes from Anthropic's launch table. The page doesn't say which effort level produced Haiku 5.5's results or who ran each test, and points to the system card for methods. VentureBeat reports that the Terminal-Bench result was at max effort and that the score is about 20% at medium.
| Benchmark (Anthropic's table) | Haiku 5.5 | Haiku 4.5 | GPT-6 Luna | Sonnet 5.5 |
|---|---|---|---|---|
| GDPval-AA v2.1 (knowledge work, Elo) | 1620 | 735 | 1437 | 1840 |
| AA-Briefcase v1.1 (knowledge work) | 1578 | 614 | 1336 | 1824 |
| OSWorld 2.1, offline subset (computer use) | 72.4% | 15.7% | 48.9% | 83.9% |
| Humanity's Last Exam, no tools | 45.9% | 10.2% | Not listed | 56.9% |
| Humanity's Last Exam, with tools | 57.4% | 18.7% | Not listed | 64.5% |
| Terminal-Bench 4.0 (agentic coding) | 39.2% | 0.0% | 16.4% | 70.6% |
| FrontierCode 1.1 (agentic coding) | 46.4% | Not listed | 42.4% | 52.1% (xhigh effort) |
| Chartography, no tools (visual reasoning) | 46.4% | 6.4% | 29.1% | 61.6% |
On Anthropic's numbers, Haiku 5.5 beats GPT-6 Luna on every result listed for both and trails Sonnet 5.5 on all of them. The widest gap to Sonnet is agentic coding in a terminal: 39.2% against 70.6%.
Sonnet 5.5's OSWorld score here, 83.9%, differs from the 80.1% in its own launch table. This one is labeled as the offline subset, so our reading is that it's a different slice of the test, though Anthropic doesn't say so.
What independent testing shows so far
Artificial Analysis, which runs its own evaluations, scores Haiku 5.5 at 43 on its Intelligence Index v4.3.2 at max effort and 34 at the default medium effort, as of October 9, 2026. Its GPT-6 Luna page shows 38 at max and 30 at medium. That backs the direction of Anthropic's claims, with a smaller margin than the launch table suggests. Sonnet 5.5 scores 56 on the same index.
Speed is the clearer win: Artificial Analysis measured Haiku 5.5 at 166 to 240 output tokens per second across effort levels, against 115 to 128 for the Luna variants it timed. The cheapest run per task still belongs to Luna, though: $0.0045 per index task at low effort, against $0.02 for Haiku 5.5 at low.
Haiku 5.5 vs Sonnet 5.5: when to pick which
Anthropic pitches Haiku 5.5 for high-volume, cost-sensitive work: summaries, classification, database queries, subagents working under Sonnet or Opus, live customer support and browser use. Its own page says Sonnet 5.5 and Opus 5.5 remain better at complex agentic coding.
Sonnet 5.5 also got cheaper on October 7. Its cache reads fell from $0.20 to $0.10 per million tokens, which Anthropic says cuts its cost on most agentic tasks by about 20%. That doesn't change its $2 and $10 base price, 20 times Haiku 5.5's short-prompt rate. See how Opus 5.5 compares if you need the top tier.
Haiku 5.5 makes more sense when:
- Your prompts stay under 100,000 tokens, which is where the 90% cut applies.
- Speed matters, as in chat support or an agent's quick side tasks.
- The task is narrow: routing, extraction or tagging, where Sonnet's extra reasoning goes unused.
Pick Sonnet 5.5 for long coding sessions in a terminal, where Anthropic's own table shows the biggest gap. If you're choosing between budget models, our GPT-6 Luna breakdown covers OpenAI's side.
What changes if you switch from Haiku 4.5
The model ID is claude-haiku-5-5 on the Claude API, Google Cloud, Microsoft Foundry and Claude Platform on AWS, and anthropic.claude-haiku-5-5 on Amazon Bedrock. It has a 1 million token context window and up to 128,000 output tokens, and Anthropic commits to keeping it available until at least October 7, 2027.
The migration guide lists the code changes. Thinking budgets set with budget_tokens return an error; you steer thinking with the effort setting instead. Non-default temperature, top_p and top_k values are rejected, and so is prefilling the assistant's reply. Priority Tier isn't supported on Haiku 5.5, and token counts and max_tokens limits need rechecking because of the new tokenizer.
Bottom line
Claude Haiku 5.5 is about 90% cheaper per token than Haiku 4.5 on prompts up to 100,000 tokens, ties GPT-6 Luna's list price there, and costs five times Luna's per-token rate between 100,000 and 272,000 tokens. Its benchmark lead over Luna is Anthropic's own claim, and Artificial Analysis confirms a smaller one, plus a clear speed edge. If you run Haiku 4.5, switch after recounting tokens and fixing the breaking changes. If your work is hard coding, Sonnet 5.5 is still the pick. More coverage lives in our AI section.
FAQ
How much does Claude Haiku 5.5 cost?
The Claude Haiku 5.5 price on the Claude API is $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 for longer prompts, as of October 9, 2026. Cache reads cost $0.01 or $0.05 per million, and the Batch API halves input and output prices.
Is Claude Haiku 5.5 better than GPT-6 Luna?
On Anthropic's own table, Haiku 5.5 leads Luna on all six results listed for both. On Artificial Analysis' independent index, it scores 43 to Luna's 38 at max effort and runs faster, but Luna's cheapest setting costs less per task.
What are effort levels in Claude Haiku 5.5?
Effort is a setting that tells the model how much to think before it answers. Haiku 5.5 offers low, medium, high, xhigh and max, with medium as the default on the API. Higher effort usually scores better and uses more tokens.