Model changes · Speed, results and division of work
Haiku 5.5: cheaper small tasks alongside Sonnet and Opus
From a cycling pelican to repeated lookups: connect an independent score, effort-level results and two pricing tiers to see which work the new Haiku could take on.
Anthropic released Haiku 5.5 on October 7. This reading connects official documentation, Artificial Analysis measurements and Simon Willison's public tests. The illustrations are his results, not a new BitShovel benchmark.
The same drawing prompt can take very different times
Simon Willison tested five effort levels with a cycling-pelican prompt. He reports 7 seconds and under 0.1 US cents at low, versus 5 minutes 9 seconds and about 3.38 cents at Max. The images show the old and new results; the timings show that a cheaper model can still think for a long time.
The author noticed that Max recognized this familiar benchmark and tried a different prompt too. Treat this as an inspectable example, not a general verdict on drawing or work. Together with the earlier Sonnet tests, it makes model choice and effort choice worth considering separately.
Sources and further reading
- Simon Willison · Drawing results and effort levelsPublic developer test · October 7Original new/old Haiku images, costs and waiting times within this test's scope.
Its role: taking on frequent, smaller tasks
Anthropic positions Haiku 5.5 around summaries, classification, extraction and subagent work: small operations repeated many times. Sonnet and Opus remain options for more complex coding and complete tasks. Existing Claude users can consider moving selected stages instead of replacing an entire workflow.
On October 8, Artificial Analysis's model page lists Max at 43 on its v4.3.2 Intelligence Index and about 243 output tokens per second, alongside high output volume. Throughput after generation begins does not measure time to a complete answer. Its price-range ranking is not an overall ranking of every model.
Sources and further reading
- Anthropic · Haiku 5.5 releaseOfficial positioning and customer accountsTask positioning, customer use, average cost claims and the Sonnet cache-price change.
- Artificial Analysis · Haiku 5.5Independent evaluation · read October 8The current model page's Max configuration, v4.3.2 index and output speed; price-group rank is not an overall model rank.
The price changes above 100,000 prompt tokens
The official documentation gives the standard API rates below. A 1M-token context window does not mean the lowest rate applies throughout: prompts above 100k use a different tier. The newer tokenizer can also count more tokens for the same text—approximately 30% more according to the documentation—so compare complete request costs.
| Model and condition | Input | Output |
|---|---|---|
| Haiku 4.5 | $1.00 | $5.00 |
| Haiku 5.5 · prompt ≤ 100k | $0.10 | $0.50 |
| Haiku 5.5 · prompt > 100k | $0.50 | $2.50 |
Sources and further reading
- Anthropic · Haiku 5.5 releaseOfficial positioning and customer accountsTask positioning, customer use, average cost claims and the Sonnet cache-price change.
- Claude Platform · Model and pricingOfficial documentation · checked October 8The 100k-token price boundary, context capacity and tokenizer changes.
Let one model assemble the work and another retrieve a fact
Anthropic quotes Rogo describing a larger model preparing a deck while a Haiku subagent retrieves segment revenue from an annual report. This is a vendor-published customer account, not our replication, but it illustrates a specific division of work.
For everyday source work, our suggested trial is structured extraction followed by passing both the extract and original to the writing model. Check omissions and provenance before expanding usage. For repeated judgment and long-task coordination, continue with the existing Opus creation and revision examples.
Sources and further reading
- Anthropic · Haiku 5.5 releaseOfficial positioning and customer accountsTask positioning, customer use, average cost claims and the Sonnet cache-price change.
Try it inside a tool you already use
Claude users can start in their existing app. The official product page lists Haiku 5.5 for Free, Pro, Max, Team and Enterprise on web, iOS and Android, and in Claude Code. Compare a short, familiar source for speed and omissions before moving more work.
GitHub is gradually adding Haiku 5.5 for Copilot Pro, Pro+, Max, Business and Enterprise. Look in supported model pickers, including VS Code, CLI, web and mobile; organization policies also apply. Usage-based product billing means these API rates are not a subscription's total price.
Sonnet 5.5 users also receive a change: Anthropic cuts cache reads from $0.20 to $0.10 per million tokens. This applies to cache reads; total savings depend on the request mix. It does not require switching models.
Sources and further reading
- Anthropic · Haiku 5.5 releaseOfficial positioning and customer accountsTask positioning, customer use, average cost claims and the Sonnet cache-price change.
- GitHub · Haiku 5.5 in CopilotOfficial product integration · October 7Eligible plans, model selection and gradual rollout.
- Anthropic · Haiku in ClaudeOfficial product page · checked October 8Plans and clients offering Haiku 5.5, including Claude Code.
See how these changes connect
Move from smaller tasks to everyday work and compare Sonnet results and effort levels.
Topic · UpdatedA complete work needs further revisions: see another layer of work handled by Opus.
