Claude Haiku 5.5 ships: 90% cheaper under 100k, tops GPT-6 Luna
Anthropic's new small model adds an effort dial and big coding gains, with fresh API credits this week—and tighter cyber limits that now block pentesting.
Copy markdown
The 90% price cut, with a 100k catch
Input falls from $1 to $0.10 and output from $5 to $0.50 per million tokens for requests under 100k—where Anthropic says about 90% of Haiku traffic sits. Above 100k it's 50% off, and the real-world average lands roughly 75% cheaper than Haiku 4.5.
It beats GPT-6 Luna, and laps Haiku 4.5
Haiku 5.5 tops Luna on every shared benchmark—OSWorld 72.4% vs 48.9%, Terminal-Bench 39.2% vs 16.4%, GDPval 1620 vs 1437. The jump over its own predecessor is bigger still: Terminal-Bench went from 0% to 39%.
First Haiku with an effort dial
You can now trade cost for intelligence with Low-to-Max effort settings, new to the Haiku tier. It defaults to medium, where Terminal-Bench is about 20%; the headline ~39% needs max effort, so tune per task.
Credits land this week; Sonnet caching gets cheaper too
Monthly API credits roll out this week—$100 on Max 5x, $200 on Max 20x, up to $500 shared across a Team plan. Sonnet 5.5 cache reads also drop from $0.20 to $0.10 per million. Live now on the API plus AWS, Google Cloud, and Azure.
The catch: pentesting is now blocked
Haiku 5.5's cybersecurity safeguards are tighter than 4.5's—penetration testing is blocked under standard safeguards. If security tooling is your workflow, test it before you migrate off 4.5.