Estimate from the version's published specs, benchmarks and reviews, 0-5; scores are only comparable within one criteria year. Scored as launched (Claude Haiku 4.5, 15 Oct 2025): 200K-token context and 64K max output (Willison); list price $1 in / $5 out per 1M tokens = $2 blended at 3:1 (Anthropic); text and image input (DataCamp page was Cloudflare-blocked; Willison and Anthropic's model docs describe image input). Anthropic: SWE-bench Verified 73.3%, OSWorld 50.7%, original Terminal-Bench about 41% (Terminus 2 harness; not Terminal-Bench 4.0, so coding null). Speed null: Anthropic says 'more than twice the speed' of Sonnet 4 and swyx measured ~3.5x faster than Sonnet 4.5, neither is tokens/s. Reasoning null (no launch-version AA index opened). Reviews: Willison, Zvi (AI #138), Latent Space AINews (aggregated hands-on reports), Analytics Vidhya (a few real-world prompts), TechCrunch.[1]
Haiku 4.5 certainly isn't that, it looks like they're continuing to focus squarely on the "great at code" part of the market.
The use case here is that it is fast and cheaper, so if you need things like coding subagents this could be the right tool for you.
Early hands-on reports suggest Haiku 4.5 materially improves iteration speed and UX.
In real-world tests, Haiku 4.5 demonstrated strong coding proficiency, logical reasoning, and the ability to adapt to user feedback, making it suitable for both developers and enterprises.
Haiku scored 73% on SWE-Bench verified and 41% on the command-line-focused Terminal-Bench — below Sonnet 4.5, but on par with Sonnet 4, GPT-5, and Gemini 2.5 in each case.
| Introducing Claude Haiku 4.5 · Oct 15, 2025 | 0.20 |
|---|---|
| Claude Haiku 4.5 · Oct 15, 2025 | 0.60 |
| Anthropic launches new version of scaled-down 'Haiku' model · Oct 15, 2025 | 0.40 |
| AI #138 Part 1: The People Demand Erotic Sycophants · Oct 16, 2025 | 0.50 |
| Claude Haiku 4.5 is Here… and it's BETTER than Sonnet 4.5? · Oct 24, 2025 | 0.60 |
Photos from readers, kept apart from the pin's sourced media. Rate what you ate or saw, and on a restaurant pin tag the dishes right on your photo.
The first entry is always the pin's source. Overall confidence is a weighted average of how firmly each reference supports the start and end times used above; a reference counts half as much for every 180 days older than the newest.
Anthropic's[3][4] post is dated "Oct 15, 2025": "Claude[2] Haiku 4.5, our latest small model, is available today to all users"; the Claude release notes ("October 15, 2025 We've launched Claude Haiku 4.5"), CNBC[5] and TechCrunch[6] agree.
Anthropic's[1][3][4] model table: Claude Haiku 4.5 (claude-haiku-4-5-20251001), "The fastest model with near-frontier intelligence", $1 / $5 per MTok, extended thinking, a 200K-token context window, 64K max output, February 2025 reliable knowledge cutoff, July 2025 training data cutoff, retirement not sooner than 15 October 2026.
Anthropic's[1][4] system card: Haiku 4.5 was trained on public internet data up to February 2025, and its safety testing supports release under the AI Safety Level 2 (ASL-2) standard.
Anthropic's[1][3] Haiku model page: Haiku 4.5 is on Claude[2].ai (web, iOS, Android), the Claude Platform, Amazon Bedrock, Google Cloud's Vertex AI, Microsoft Foundry and Claude Code; pricing starts at $1 per million input and $5 per million output tokens, with up to 90% savings from prompt caching and 50% from batch processing.
CNBC (published 2025-10-15T17:00Z): "Anthropic[1][3][4] on Wednesday announced Claude[2] Haiku 4.5", available to free users and now the cheapest model for paid users; it is better at using computers than Sonnet 4, and CPO Mike Krieger says "It punches way above its weight".
Something missing or wrong? Say it in your own words: a link that backs this pin up, a different start or end date and why, or a fact it lacks or gets wrong. The AI checks it against this pin's sources, searches for better ones, and adds any page that backs you up. The pin's own sources still count most. A picture that shows something else, or shows it badly, is looked at too, and moved down or replaced.
| Benchmark | Claude Sonnet 4.5 | Claude Haiku 4.5 | Claude Sonnet 4 | GPT-5 | Gemini 2.5 Pro |
|---|---|---|---|---|---|
| SWE-bench Verified (agentic coding) | 77.2% | 73.3% | 72.7% | 72.8% GPT-5 (high), 74.5% GPT-5-Codex | 67.2% |
| Terminal-Bench (agentic terminal coding) | 50.0% | 41.0% | 36.4% | 43.8% | 25.3% |
| τ2-bench (agentic tool use) | Retail 86.2%, Airline 70.0%, Telecom 98.0% | Retail 83.2%, Airline 63.6%, Telecom 83.0% | Retail 83.8%, Airline 63.0%, Telecom 49.6% | Retail 81.1%, Airline 62.6%, Telecom 96.7% | — |
| OSWorld (computer use) | 61.4% | 50.7% | 42.2% | — | — |
| AIME 2025 (high school math competition) | 100% (python), 87.0% (no tools) | 96.3% (python), 80.7% (no tools) | 70.5% | 99.6% (python), 94.6% (no tools) | 88.0% |
| GPQA Diamond (graduate-level reasoning) | 83.4% | 73.0% | 76.1% | 85.7% | 86.4% |
| MMMLU (multilingual Q&A) | 89.1% | 83.0% | 86.5% | 89.4% | — |
| MMMU, validation (visual reasoning) | 77.8% | 73.2% | 74.4% | 84.2% | 82.0% |