Claude Haiku 5.5 lands up to 90% cheaper, with effort controls
The cheapest, fastest Claude yet is live on every cloud. Sonnet 5.5 cache reads drop too, Max and Team get API credits, and GLM 5.3 Flash lands same day.
Copy markdown
The cheapest, fastest Claude yet
Haiku 5.5 is live at $0.10 / $0.50 per million tokens for prompts under 100K — about 90% below Haiku 4.5 at that size, ~75% cheaper on average. It ships the first adjustable effort dial (low to max) on a Haiku model, defaulting to medium. Call it now as model ID claude-haiku-5-5.
It got smarter, not just cheaper
This isn't a price-only refresh. OSWorld 2.1 jumps from 15.7% to 72.4% and Terminal-Bench 4.0 from 0% to 39.2% versus Haiku 4.5. Anthropic still points you to Sonnet or Opus for heavy agentic coding, but for high-volume classification, extraction, and tool calls, this is your new default.
Sonnet 5.5 quietly got ~20% cheaper
Cache reads on Sonnet 5.5 were halved, $0.20 to $0.10 per million tokens, effective immediately — Anthropic says that's roughly 20% off most agentic tasks. Some Azure and Google Cloud customers may see it land a few days later.
Free API credits if you're on Max or Team
Rolling out this week: $100/month in Claude Platform credits for Max 5x, $200 for Max 20x, and up to $500 pooled for Team. They work on any model, so your subscription now quietly subsidizes your API builds.
Computer and browser use hit the SDKs
Haiku 5.5 adds beta computer-use and browser-use support to the Python and TypeScript SDKs. A cheap, fast model that can drive a browser is a weekend-project unlock — think bulk form-filling or screen-reading agents that won't blow your token budget.
GLM 5.3 Flash crashed the same party
Z.ai shipped GLM-5.3-Flash the same day. Artificial Analysis scores it 1647 on GDPval-AA and 1454 on AA-Briefcase — edging Haiku 5.5's 1620 / 1578. The cheap-and-fast tier just became a two-horse race worth A/B testing.