benchmarks
Claude Haiku 4.5 is Anthropic's compact model for coding, conversational applications and tool-using agents, with support for extended thinking and computer use. Its documented specifications include a 200K-token context window and standard API pricing of $1 per million input tokens and $5 per million output tokens.
Benchmark results
Every figure links to where it was published. Compare like with like — harnesses and prompt setups differ between vendors.
| Benchmark | Score | Source |
|---|---|---|
| SWE-bench VerifiedResolving real GitHub issues | 73.3%15 Oct 2025 | View |
| GPQA DiamondPhD-level science questions | 71.2%22 Oct 2025 | View |
| OTIS Mock AIME 2024-25Competition mathematics | 66.7%22 Oct 2025 | View |
| Terminal-BenchLong-horizon command-line agent tasks | 35.5%15 Oct 2025 | View |
| FrontierMath | 5.9%22 Oct 2025 | View |
| SimpleQA VerifiedShort factual questions (hallucination resistance) | 13.2%10 Aug 2026 | View |
| MATH Level 5 | 96.4%22 Oct 2025 | View |
| LMArena TextHuman preference Elo from blind chat votes (not a percentage) | 1397 Elo22 Oct 2026 | View |
| LMArena WebDevHuman preference Elo for building web apps (not a percentage) | 1330 Elo23 Oct 2026 | View |
Weighted mean of 7 verified benchmark results: SWE-bench Verified 73.3%, GPQA Diamond 71.2%, OTIS Mock AIME 2024-25 66.7%, Terminal-Bench 35.5%, FrontierMath 5.9%, SimpleQA Verified 13.2%, MATH Level 5 96.4%.
Capabilities
- Code generation, debugging and software-engineering tasks.
- Extended thinking for tasks requiring additional reasoning.
- Tool use for agentic workflows and external integrations.
- Computer use through an integrated execution environment.
- Text and image input with text output.
What improved
- Adds extended thinking to the Haiku model line.
- Adds computer-use capabilities to the Haiku model line.
- Raises coding performance to approximately Claude Sonnet 4's level in Anthropic's published evaluation.
Known flaws & limitations
- Scores below Claude Sonnet 4.5 on Anthropic's published SWE-bench Verified evaluation.
- The 200K-token context window limits the material that can fit into one request.
- Does not generate native image or audio output.
- Generated code and factual answers can contain errors and require independent validation.
Risks, safety & security issues
- Untrusted documents, websites and tool results can carry prompt-injection instructions that attempt to redirect agents.
- Computer-use and tool permissions can turn incorrect model decisions into unintended external actions.
- Incorrect answers or generated code can cause downstream harm when used without independent checks.
Good fit for
- Latency-sensitive conversational assistants and customer-support workflows.
- Coding assistance with automated tests and human review.
- High-volume document classification, extraction and summarization.
- Scoped agent tasks with restricted tools and clear completion criteria.
Poor fit for
- High-stakes decisions without qualified human review.
- Unrestricted autonomous operation with sensitive data or irreversible permissions.
- Workloads requiring more than 200K tokens in a single context.
- Native image or audio generation.
References
Profile last reviewed 10/27/2026 and re-checked automatically every few hours against the vendor's published documentation, papers and public leaderboards. Scores come from the cited benchmark results above — confirm limits and pricing against the provider's own documentation before building on them.