Primary use stated or inferred cautiously from official documentation.
A
ANTHROPIC / claude-haiku-5-5
Claude Haiku 5.5
Anthropic model released on October 7, 2026 for frequent, cost- and latency-sensitive tasks. It accepts text and images and returns text, including code; it does not generate images, audio or video. Its one-million-token window and adaptive thinking expand the Haiku family without making it a universal replacement for Sonnet or Opus on complex work.
01 / SUMMARY
The essential
Maximum input capacity when published by the source.
Documented maximum generation limit.
Verified availability channels.
Primary functional family and verified specialties.
Declared terms for API access, model weights, or self-hosting.
02 / TECHNICAL SHEET
Limits and integration
| API ID | claude-haiku-5-5 |
|---|---|
| Model type | Text and reasoning · Code |
| Access model | Paid proprietary |
| License | Proprietary |
| Deployment | Hosted API |
| Release | 07 OCT 2026 |
| Knowledge cutoff | JUN 2026 |
| Entrance | Text · Image |
| Exit | Text |
| Context window | 1,000,000 tokens |
| maximum output | 128,000 tokens |
| Reasoning | Adaptive thinking · default medium effort |
| Published tools | Use of tools · Computer use · Browser use |
| Structured Outings | Not published |
| Batch processing | Compatible · 50% discount |
| Prompt cache | Supported |
| Fine-tuning | Not published |
| Verified platforms | Claude API · Amazon Bedrock · Google Cloud · Microsoft Foundry · Claude Platform on AWS |
Classification, data extraction, summaries, conversation compaction and routing. It can act as a subagent for bounded tasks and use tools, a browser or a computer. The default effort is medium; changing effort affects cost and speed. Browser use is available on Claude API and Google Cloud, not necessarily with an identical contract across all clouds.
Standard maximum output is 128,000 tokens, including thinking; the 300,000-token Batch limit requires a specific beta. The new tokenizer counts approximately 30% more tokens for the same text than Haiku 4.5. Omit non-default temperature, top_p and top_k, manual thinking budgets and assistant prefill. Thinking blocks can only be reused in the account that generated them or linked accounts. Validate permissions, refusals and computer actions.
03 / CAPABILITIES
What it can do
Orientation
Classification, data extraction, summaries, conversation compaction and routing. It can act as a subagent for bounded tasks and use tools, a browser or a computer. The default effort is medium; changing effort affects cost and speed. Browser use is available on Claude API and Google Cloud, not necessarily with an identical contract across all clouds.
Context
1.000.000 tokens · 128.000 tokens
Tools and integration
Use of tools · Computer use · Browser use
Access
Claude API and associated clouds
Input modalities
TextImage
04 / PRICES
Documented cost
| Concept | Worth | Unit/condition |
|---|---|---|
| Standard input | 0.10 / 0.50 USD / 1 M tokens | Standard API rate |
| Cached input | 0.01 / 0.05 USD / 1 M tokens | Reading reused prefixes |
| Cache write or storage | 0.125 / 0.625 USD (5 min); 0.20 / 1.00 USD (1 h) / 1 M tokens | The condition varies by provider |
| Standard output | 0.50 / 2.50 USD / 1 M tokens | May include reasoning tokens |
| Batch input | 0.05 / 0.25 USD / 1 M tokens | Asynchronous processing |
| Batch output | 0.25 / 1.25 USD / 1 M tokens | Asynchronous processing |
Each pair of rates corresponds to prompts up to 100,000 tokens / over 100,000 tokens, including output and caching. Input: USD 0.10 / 0.50; output: USD 0.50 / 2.50 per million tokens. Batch discounts input and output by 50%. Caching distinguishes reads from five-minute or one-hour writes. Numeric fields represent only the short-prompt tier, not a long-context budget. Tools and regional conditions may add charges.
i Prices change and may depend on level, region or context length. Check the source before making a decision.
05 / EVALUATIONS
How to read the results
“
Eight results published by Anthropic, not reproduced by Inferama. GDPval-AA v2.1 and AA-Briefcase v1.1 are ratings, not percentages. OSWorld 2.1 uses the offline subset; Humanity’s Last Exam is recorded separately with and without tools. We retain Terminal-Bench and FrontierCode versions and Chartography’s no-tools condition. Compare configurations, effort, tools and cost per task; these figures do not guarantee your application’s results.
| Benchmark | Result | Metric | Source |
|---|---|---|---|
| GDPval-AA v2.1 | 1.620 | Rating | View source ↗ |
| AA-Briefcase v1.1 | 1.578 | Rating | View source ↗ |
| OSWorld 2.1 (offline subset) | 72,4 % | Accuracy | View source ↗ |
| Humanity’s Last Exam (no tools) | 45,9 % | Accuracy | View source ↗ |
| Humanity’s Last Exam (with tools) | 57,4 % | Accuracy | View source ↗ |
| Terminal-Bench 4.0 | 39,2 % | Accuracy | View source ↗ |
| FrontierCode 1.1 (Main) | 46,4 % | Accuracy | View source ↗ |
| Chartography (no tools) | 46,4 % | Accuracy | View source ↗ |
Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.
06 / VERSIONS
Chronology
Claude Haiku 5.5
Version added to Inferama's verified catalog.
07 / SOURCES