
None
None
Long Story Video Skill
Turn your idea into 10s–10min video—AI writes scripts, prompts, and generates footage automatically.

Ads Video Skill
Generate professional ads and sales videos—AI auto-generates scripts, prompts, and footage.
3D Science Explainer Video Skill
Convert scientific concepts into stunning 3D explain animations
Feedback
freeTrialImage.bannerPity
freeTrialImage.upgradeUnlock
- ✓freeTrialImage.benefitHd
- ✓freeTrialImage.benefitWatermark
- ✓freeTrialImage.benefitUnlimited
Haiku 5.5 Benchmark
Haiku 5.5 Benchmark figures checked against their sources: launch scores, price bands and published limits, with no hands-on test claimed.
All Tools
Discover our comprehensive AI-powered animation toolkit

Nano Banana2
Best Image Generator

Kling3
Next-Gen AI Video Generator

Veo3
Create Stunning Videos with Veo3.1

Seedance2.0
The Future of AI Video Is Here.
Kling Motion Control
Turn reference images into amazing motion videos in minutes

Suno AI Music Generator
Create Professional Music with AI

Nano Banana
Advanced AI Image Generator

AI Image Editor
AI Photo Editor Online
What the Haiku 5.5 Benchmark Tells You Before You Commit
Every launch score, price band and per-effort cost stays attached to its configuration and the date it was retrieved.
- Launch Scores Live on Different ScalesElo, partial credit, coding pass rates and accuracy each sit on their own scale, so averaging them tells you nothing.
- Built for High-Volume, Low-Cost LoopsDesigned for extraction, classification and summarisation jobs that repeat thousands of times a day.
- Two Footnotes That Rewrite the TablesDiscount rates stop at 100K tokens; the 72.4% computer-use figure is partial credit, not a strict pass rate.
Reading the Haiku 5.5 Benchmark in Three Moves
Follow these three steps to convert published figures into a workload decision you can defend.
Haiku 5.5 Benchmark: Six Findings That Matter
Specifications, price bands, coding outcomes, computer-use scores and third-party index data for Claude Haiku 5.5, all kept with their sources.
A One-Million-Token Window and 128K Output
Text and image input, text output, adaptive thinking on. The million-token context and 128,000-token output ceiling dwarf Haiku 4.5's 200,000 and 64,000.
Pricing Shifts Once Prompts Pass 100K
Past 100,000 prompt tokens, input, output and caching each move to a band five times the cheap one. OpenAI's base Standard band reaches 272,000 input tokens.
Terminal-Bench 4.0: 39.2% at Launch
That figure clearly beats Luna but trails the published Sol rows. Runs get an eight-hour agent timeout, and 2.1 or 3.0 numbers come from an earlier experiment.
FrontierCode: 46.4% at Max Effort
Grading weighs correctness, tests, scope, style and repo conventions. Medium effort lands at 41.6% for $0.13 a rollout; xhigh to max doubles spend for half a point.
OSWorld 2.1 Rewards Steps, Not Finished Work
Eighty-two offline tasks, five attempts each, 1080p screenshots and a 500-step ceiling. For a workflow that must finish, 37.1% strict is the honest number.
Third-Party Index Scores and Tool Gains
Artificial Analysis lists Haiku at all five efforts on Intelligence Index v4.3.2 with rounded task costs. HLE reads 45.9% alone and 57.4% with tools — a lift from the surrounding system.
Haiku 5.5 Benchmark: Questions Answered
Straight answers on Claude Haiku 5.5 scores, token pricing, context ceilings and local deployment.
Where does Haiku 5.5 stand next to GPT-6 Luna and GPT-6 Sol?
On knowledge work, screenshot tasks and narrow coding, Haiku trades blows with Luna. Sol and Sol 6.1 cost more and earn it on coding, while Sonnet 5.5 still leads most rows.
Is the 72.4% computer-use figure a success rate?
No. The same run reports 37.1% when every checkpoint must be met. An agent can find and edit the right file, then save it in the wrong format or stop before the final action. Add an end-state check.
At what point does Haiku 5.5 move to the higher price band?
Only prompts up to 100,000 tokens get the cheapest rates; beyond that every category costs five times more. Anthropic's 90%, 50% and ~75% savings figures compare different baselines.
Could Haiku 5.5 run on local hardware?
No parameter count, active-parameter figure, layer count or training compute appears in the published specs. 'Small' labels a product tier, not an architecture, so local GPU estimates have no basis.
Does more reasoning effort always buy better output?
Not always. Sol 6.1 scores better at medium than at max, and Haiku's 0.2-point edge over Sonnet at max is too small to call a win. Extra effort adds tokens, latency and chances to drift off task.
Why do identical models show different numbers across tables?
Sonnet reads 70.6% in Anthropic's launch table and 61.8% on the public leaderboard snapshot, where Haiku had no visible row. Artificial Analysis dropped its 'Default Fallback' label, yet Anthropic's migration guide still says no server-side fallback exists.
Put the Haiku 5.5 Benchmark to Work on Your Own Tasks
Choose one workload, set the effort level, and measure Haiku 5.5 against the quality floor you already accept.
