Anthropic released Claude Haiku 5.5 this week. The headline is a 90 percent cut to input costs: $0.10 per million tokens for requests under 100,000 tokens, down from $1 per million for Haiku 4.5. Anthropic pegs the average real-world saving at around 75 percent, accounting for a tokenizer update that inflates token counts by roughly 30 percent. The price matters. What matters more for anyone building creative AI workflows is the other number: 72.4 percent on OSWorld 2.1, a benchmark for autonomous computer operation, against Haiku 4.5’s 15.7 percent. The previous version was, by that measure, essentially non-functional for anything requiring autonomous computer control. This one crosses a threshold.
What Changed in Haiku 5.5
Haiku 5.5 is the first Haiku-class model with an adjustable effort setting. The effort parameter runs from low to max — low for classification and routing, higher for multi-step extraction or computer-use steps — letting you calibrate cost against quality per task rather than paying Sonnet-tier compute for everything. Context window is 1 million tokens, matching Sonnet 5.5. Max output is 128K tokens. The model accepts image inputs alongside text. Model ID is claude-haiku-5-5, available now on the Claude Platform, AWS Bedrock, Google Cloud, and Azure.
Anthropic calls it “the cheapest, fastest, and most capable small model we’ve ever released.” On Terminal-Bench 4.0, Haiku 5.5 scores 39.2 percent against Haiku 4.5’s zero. On Humanity’s Last Exam, 45.9 percent against Haiku 4.5’s 10.2 percent. These aren’t incremental gains — the model performs meaningfully better at the multi-step tasks agentic pipelines actually demand. Along with the Haiku launch, Anthropic halved the cache-read price for Sonnet 5.5 to $0.10 per million tokens, making the Sonnet-as-orchestrator tier of a tiered architecture roughly 20 percent cheaper too.
The Computer-Use Number Is the Real Story
OSWorld 2.1 tests autonomous desktop operation: navigating applications, completing multi-step tasks, interacting with interfaces without guidance. Haiku 4.5 scored 15.7 percent — roughly one completed task in six. Haiku 5.5 scores 72.4 percent, completing nearly three in four. That’s not a better version of something that barely worked; it’s a different capability entirely. Sonnet 5.5 scores 83.9 percent on the same benchmark and remains the choice for complex or high-stakes computer-use work. But for the routine navigation tasks in an agentic loop — opening files, triggering exports, moving through menus — Haiku 5.5 now handles that work at a fraction of the cost.
One migration note: Haiku 5.5 requires the new computer_toolset_20260801 toolset for computer-use tasks — the previous toolset returns a 400 error. Other breaking changes from Haiku 4.5 include non-default temperature, top_k, and top_p values now returning 400 errors, no assistant prefill, and responses that may begin with thinking blocks. If you’re migrating an existing Haiku 4.5 integration, Anthropic’s migration documentation has the full list.
What This Changes for Creative Pipelines
DaVinci Resolve 21’s native MCP server already lets Claude organize media, navigate the interface, and trigger actions without scripted automation. Running those navigation steps with Haiku 5.5 rather than Sonnet costs roughly a tenth as much per operation. The same math applies to After Effects script generation, batch video metadata processing, ComfyUI node management, or routing files through a multi-stage export pipeline. These are exactly the tasks Haiku 5.5 is designed for — high volume, speed-sensitive, and tolerant of occasional errors that a Sonnet-tier orchestrator can catch and correct.
The tiered architecture that now makes practical sense: Sonnet 5.5 or Opus 5.5 as the reasoning and decision-making layer, Haiku 5.5 as the execution layer for high-volume steps that need to click buttons and move files. The cost reduction makes building and running that architecture economically viable at a scale where it wasn’t before.
Pricing in Full
Requests under 100,000 tokens: $0.10 per million input, $0.50 per million output, $0.01 per million for cached input reads. Requests over 100,000 tokens jump to $0.50 input and $2.50 output — the threshold is easier to hit than it sounds, given that the new tokenizer uses roughly 30 percent more tokens than Haiku 4.5 on the same input text. The Batch API halves input and output rates for asynchronous work. Haiku 4.5 was $1 input and $5 output with no tiered structure; the effective reduction depends on your actual prompt length distribution, but on most high-volume agentic workloads the saving is substantial.
What Haiku 5.5 Is Not
Sonnet 5.5 still leads on complex judgment calls, extended reasoning, and anything requiring real creative generation — 83.9 percent on OSWorld versus 72.4 percent, 70.6 percent on Terminal-Bench versus 39.2 percent. The effort control in Haiku 5.5 helps calibrate cost against quality per task, but the ceiling is what it is. Haiku 5.5 is the model you use at the edges of your pipeline: the one making ten thousand routine decisions per hour so the more capable model can focus on the ten that actually matter.
Sources:
Anthropic — Introducing Claude Haiku 5.5 (official announcement)
Yahoo Finance / Daniel Howley — Anthropic reveals Haiku 5.5 model as AI pricing war intensifies
DataCamp — Claude Haiku 5.5: Features, Benchmarks, and Pricing
Developers Digest — Claude Haiku 5.5: Pricing, Migration Changes, and When to Use It
tbreak.com — Claude Haiku 5.5 launches with up to 90% lower prices
