Claude Haiku 4.5

Overview

Claude Haiku 4.5, released on October 15, 2025, is Anthropic's fastest and most cost-efficient model. It remains the current Haiku-tier offering as of July 2026 — Anthropic shipped 4.6 revisions to the Sonnet and Opus lines but never released a Haiku 4.6.

Haiku 4.5 delivers coding performance comparable to Claude Sonnet 4 — a model considered state of the art shortly before its release — at one-third the cost and more than twice the speed. On certain tasks, notably computer use, it surpasses Sonnet 4 outright.

The practical result is that near-frontier capability becomes viable in places where frontier models are not: real-time chat, high-volume classification, sub-agent fan-out, and cost-sensitive deployments.

Capabilities

• Cost efficiency: $1/$5 per million tokens — the cheapest model in Anthropic's current lineup.
• Speed: More than twice the speed of Claude Sonnet 4, per Anthropic. A customer quoted in the launch post reports it "up to 4-5 times faster than Sonnet 4.5 at a fraction of the cost."
• Coding: 73.3% on SWE-bench Verified, close to models several times its price.
• Computer use: Surpasses Claude Sonnet 4 at browser navigation and tool use, which is why it powers latency-sensitive surfaces like Claude for Chrome.
• Extended thinking: Supports manual extended thinking with a configurable token budget, lifting Terminal-Bench from 40.21% to 41.75% with a 32K budget.
• Alignment: Statistically significantly lower misaligned-behavior rates than Sonnet 4.5 and Opus 4.1 in automated assessments.

Technical Specifications

• API model ID: claude-haiku-4-5 (alias) → claude-haiku-4-5-20251001 (pinned snapshot)
• Context window: 200K tokens
• Max output: 64K tokens per request
• Thinking: Manual extended thinking — thinking: {"type": "enabled", "budget_tokens": N}. Minimum 1024, must be strictly less than max_tokens.
• Effort parameter: Not supported. Passing output_config.effort errors on Haiku 4.5.
• Reliable knowledge cutoff: February 2025
• Training data cutoff: July 2025
• Comparative latency: Fastest in the Claude lineup
• Safety level: ASL-2
• Architecture: Transformer-based, developed under Anthropic's Constitutional AI approach

Supported features

Extended thinking, vision, prompt caching (minimum cacheable prefix 4096 tokens), Message Batches, structured outputs, tool use, the memory tool, bash and text-editor tools, code execution, and tool search.

Not supported: adaptive thinking, the effort parameter, fast mode, task budgets, and mid-conversation system messages.

API Behavior

Haiku 4.5 predates the API changes that landed with Opus 4.7 and later, so it behaves like the older generation:

• Extended thinking, not adaptive. Use thinking: {"type": "enabled", "budget_tokens": N}. A budget_tokens greater than or equal to max_tokens returns a 400.
• Sampling parameters work. temperature, top_p, and top_k are all accepted — unlike on Fable 5, Opus 4.8, 4.7, and Sonnet 5, where they return a 400.
• effort errors. The output_config.effort parameter is not available on this model.
• Separate rate-limit pool. Haiku 4.5 has its own rate limits, distinct from Haiku 3 and 3.5. If you are ramping traffic while migrating, check your tier's Haiku 4.5 limits.

Use Cases

• Real-time applications: Chat assistants, customer-service agents, and any surface where response latency is user-visible.
• Sub-agent fan-out: Cheap parallel workers driven by an Opus- or Sonnet-tier orchestrator.
• High-volume classification and extraction: Tasks where per-request cost dominates and frontier reasoning is unnecessary.
• Computer use at scale: Browser navigation and tool use, where it outperforms Claude Sonnet 4.
• Pair programming: Fast inline completions and small refactors.
• Cost-sensitive deployments: Startups, educational platforms, and internal tooling with tight per-seat budgets.

Performance / Benchmarks

Anthropic publishes Haiku 4.5's benchmark comparison as an image ("Comparison table of frontier models across popular benchmarks"). Only three of its scores appear anywhere as text, and all three live in the announcement's Methodology footnotes. The Claude Haiku 4.5 System Card is a safety document and contains no capability benchmark table. This section is therefore short by necessity.

The qualifiers are load-bearing, and Anthropic states them verbatim:

• SWE-bench Verified — "All Claude results were reported using a simple scaffold with two tools—bash and file editing via string replacements. We report 73.3%, which was averaged over 50 trials, no test-time compute, 128K thinking budget, and default sampling parameters (temperature, top_p) on the full 500-problem SWE-bench Verified dataset." The run also used a prompt addition: "You should use tools as much as possible, ideally more than 100 times. You should also implement your own tests first before attempting the problem." An unqualified "73.3% on SWE-bench Verified" overstates what an out-of-the-box, no-thinking Haiku 4.5 call will do.
• Terminal-Bench — "All scores reported use the default agent framework (Terminus 2), with XML parser, averaging 11 runs (6 without thinking (40.21% score), 5 with 32K thinking budget (41.75% score)) with n-attempts=1."

Methodology published, scores not

For four further benchmarks, Anthropic describes the methodology in text but leaves the score inside the image:

• τ2-bench — "averaging over 10 runs using extended thinking (128k thinking budget) and default sampling parameters," plus prompt addenda to the Airline and Telecom agent policies.
• AIME — "the average over 10 independent runs that each calculate pass@1 over 16 trials with default sampling parameters ... and 128K thinking budget."
• OSWorld — "the official OSWorld-Verified framework with 100 max steps, averaged across 4 runs with 128K total thinking budget and 2K thinking budget per-step configured."
• MMMLU — "the average of 10 runs over 14 non-English languages with a 128K thinking budget."

We do not reproduce numbers for these, because Anthropic does not publish them as text anywhere we can cite.

The two most-quoted numbers are not Anthropic's

Both come from partner testimonials in the launch post, and both are frequently misattributed:

• "90% of Sonnet 4.5's performance" is measured on Augment's agentic coding evaluation, not Anthropic's: "In Augment's agentic coding evaluation, it achieves 90% of Sonnet 4.5's performance" (Guy Gur-Ari, Co-Founder).
• "4-5× faster than Sonnet 4.5" is a customer quote (Andrew Filev, CEO): "Now it runs up to 4-5 times faster than Sonnet 4.5 at a fraction of the cost." No benchmark or workload qualifier is attached to it.

Anthropic's own comparative claim is coarser: similar coding performance to Claude Sonnet 4 at one-third the cost and more than twice the speed, with computer-use performance exceeding Sonnet 4.

Artificial Analysis

Artificial Analysis lists Haiku 4.5 as two separate entries, both labelled Intelligence Index v4.1:

• Claude 4.5 Haiku (Non-reasoning) — "scores 24 (estimated) on the Artificial Analysis Intelligence Index." The (estimated) flag is Artificial Analysis's own; this is not a measured run.
• Claude 4.5 Haiku (Reasoning) — "scores 30 on the Artificial Analysis Intelligence Index."

Index versions are not comparable, so neither figure can be set against a score from an earlier index revision.

How to read these numbers

The three scores in the table are vendor-reported, and each was produced with a thinking budget enabled — 128K for SWE-bench Verified, 32K for the higher Terminal-Bench figure. Haiku 4.5 supports neither adaptive thinking nor the effort parameter, so reproducing them means setting budget_tokens explicitly. Treat the figures as an upper bound and validate on your own evaluation set.

Limitations

• Smaller context window: 200K tokens, versus 1M on Opus 4.8, Sonnet 5, and Fable 5.
• Lower max output: 64K tokens per request, versus 128K on the larger models.
• Older knowledge: Reliable cutoff of February 2025 — nearly a year behind the January 2026 cutoff of the current Opus, Sonnet, and Fable models. For recent developments, pair it with web search or tool use.
• No adaptive thinking or effort control: Reasoning depth is tuned with a raw token budget rather than the modern effort parameter.
• Not a frontier model: For the hardest reasoning, long-horizon agentic runs, or deep research, step up to Sonnet 5 or Opus 4.8.
• No internet access: Cannot browse unless integrated through tool use or the server-side web search tool.

Pricing & Access

API Pricing

Available on the Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud, and Microsoft Foundry:

• Input: $1 / MTok
• Output: $5 / MTok
• Prompt caching: up to 90% savings on cached input
• Batch API: 50% discount

Consumer Plans

Available across Claude's Free, Pro, and Max plans as the fast-response tier, and in Claude for Chrome.

Ecosystem & Tools

Official SDKs

• Python SDK — full-featured with async support
• TypeScript SDK — type-safe for Node.js and browsers
• REST API — direct HTTP access from any language

SDKs also ship for Java, Go, Ruby, C#, and PHP, along with the ant command-line interface.

Developer Tools

• Anthropic Console — playground and prompt development
• Claude.ai — browser-based chat
• Claude for Chrome — browser agent built on Haiku's computer-use speed

Community & Resources

• Introducing Claude Haiku 4.5 — official announcement
• Claude Haiku product page
• Models Overview
• Pricing Page