Anthropic shipped Claude Haiku 5.5 on October 7, 2026, and two numbers decide whether it matters to you: the price and the computer-use score. Anthropic’s announcement puts the model at roughly 75% cheaper to run than Haiku 4.5, and reports 72.4% on the OSWorld 2.1 computer-use benchmark (offline subset) against 15.7% for Haiku 4.5 and 48.9% for OpenAI’s GPT-6 Luna. A cheap model that can drive a real computer is a different product from a cheap model that can only summarize text.
Key facts
| Fact | Detail |
|---|---|
| Question answered | Claude Haiku 5.5: Anthropic’s Cheapest Model Is Now Usable for Computer Use |
| Short answer | Anthropic released Claude Haiku 5.5 on October 7, 2026: roughly 75% cheaper to run than Haiku 4.5, matching GPT-6 Luna at $0.10/$0.50 per million tokens up to 100k, and scoring 72.4% on the OSWorld 2.1 compute… |
| Reading time | ~3 min |
| Last updated | 2026-10-08 |
What changed
- Price. Simon Willison reports $0.10/$0.50 per million input/output tokens up to 100,000 tokens - an exact match with GPT-6 Luna. Past 100k tokens the price jumps fivefold to $0.50/$2.50, while Luna only steps up at 272,000 tokens and only to $0.20/$0.75. Haiku 4.5 was $1/$5.
- A less generous tokenizer. The same long prompt consumes around 1.25x more tokens than Haiku 4.5, so part of the price cut is offset. Willison calls it the hidden price increase.
- Adjustable effort. Haiku 5.5 is the first Haiku-class model with effort levels (low, medium, high, xhigh, max) and defaults to medium; reasoning cannot be disabled. In Willison’s SVG test, low effort took 7 seconds and cost 0.0936 cents, while max effort took 5 minutes 9 seconds and cost 3.3826 cents.
- Benchmarks beyond OSWorld. 1620 on GDPval-AA v2.1 (Haiku 4.5: 735), 39.2% on Terminal-Bench 4.0 (Haiku 4.5: 0.0%), and 57.4% on Humanity’s Last Exam with tools.
- Side announcements. Sonnet 5.5’s cache-read price is halved, which Anthropic says makes Sonnet 5.5 about 20% cheaper on most agentic work, and Claude Max and Team subscribers get a new monthly API credit.
Why the cheap tier matters for desktop agents
Agent work is not one job - it is one expensive planning call followed by dozens of small ones: read this dialog, classify this file, decide whether the checkbox is ticked, retry a failed step. Paying frontier prices for all of it is what makes desktop automation feel unaffordable.
Haiku 5.5 pushes in two directions at once. The price makes the boring calls cheap enough to run on every step, and the OSWorld jump means the cheap tier can at least attempt the on-screen loop instead of handing every click back to a bigger model. Anthropic positions it as a subagent alongside Opus 5.5 and Sonnet 5.5, and lists browser use and live support as the speed-sensitive workloads it is built for.
Two practical cautions. The 100k-token boundary is a cliff, not a slope - long transcripts that cross it get five times more expensive per token, so compaction still matters. And a cheaper tokenizer is not a cheaper task: measure cost per finished task, not cost per token.
MOOGH runs that split locally: a desktop agent on Windows 11 takes the plan, asks for your approval, then executes the steps on your own PC - file operations, command execution and routine automation. You can point it at any model you have configured; to try the workflow on your own folders, download the client.
What to do with it
- Route summaries, classification, form filling and retry loops to Haiku 5.5; keep planning and final review on a larger model.
- Try the low and medium effort settings first. They are dramatically cheaper than max and, on the pelican test above, max bought reasoning time rather than a better result.
- Watch the 100k-token boundary in long agent sessions and compact before you cross it.
FAQ
- Can Haiku 5.5 replace a frontier model for agent work? No. It trails Sonnet 5.5 and Opus 5.5 on every benchmark Anthropic published, and the OSWorld gap is still wide (72.4% vs 83.9%). It replaces the routine half of an agent’s workload.
- Is it the cheapest option above 100k tokens? No. Past 100,000 tokens GPT-6 Luna is priced lower, because Haiku 5.5 jumps to $0.50/$2.50 while Luna’s first increase comes at 272,000 tokens.
- Does the subscription API credit apply to the free plan? Anthropic says the monthly API credit is for Claude Max and Team subscribers.
Related reading: