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
Every published Haiku 5.5 Benchmark figure in one place — task scores, token pricing, context caps and the caveats most write-ups leave out.
All Tools
Discover our comprehensive AI-powered animation toolkit
MiniMax H3
MiniMax H3 AI Video Generator
Seedance 2.5
The Future of AI Video Is Here.
Seedance 2.0
The Future of AI Video Is Here.
Veo3.1
Create Stunning Videos with Veo3.1
FLUX 3 Video Generator
Kling 3.0
Next-Gen AI Video Generator
Grok Video Generator
Create Videos from Text or Images with AI
MiniMax H3 video generator
MiniMax H3 AI Video Generator
Inside the Haiku 5.5 Benchmark Numbers
Every reported score, token tier and per-run cost stays paired with the setup and the date it was pulled from.
- How Each Reported Figure Is MeasuredElo, partial-credit, coding and accuracy scores sit on different scales, so averaging them together is meaningless.
- Economics of a Small Model at VolumeDesigned for extraction, classification and summarization jobs that run thousands of times.
- Two Footnotes That Rewrite the TablesDiscount pricing ends at 100K tokens, and the 72.4% computer-use figure is partial credit, not completion.
How to Put the Haiku 5.5 Benchmark to Work
Three practical steps that convert published figures into a decision about your own workload.
What Stands Out in the Haiku 5.5 Benchmark
Specs, token tiers, coding scores, computer-use results and third-party index readings for Claude Haiku 5.5, gathered in one view.
Context Window and Output Ceiling
Text-and-image input, text output, adaptive thinking enabled. A 1M-token context and a 128,000-token output cap leave Haiku 4.5's 200,000 and 64,000 far behind.
Two Pricing Tiers by Prompt Length
Pass 100,000 prompt tokens and every category — input, output, caching — moves to a tier five times the cheap rate. OpenAI's Standard band runs to 272,000 input tokens.
Terminal-Bench 4.0: Finished Tasks
The 39.2% launch result clears Luna by a wide margin yet sits below the published Sol rows. Runs allow an eight-hour agent timeout, and any 2.1 or 3.0 figures come from an earlier experiment.
FrontierCode: Mergeability vs. Effort Cost
Scoring weighs correctness, tests, scope, style and repo conventions. Max effort returns 46.4% and medium 41.6% at $0.13 per rollout, yet moving xhigh to max doubles the bill for half a point.
OSWorld 2.1: Partial Credit vs. Strict Pass
The run covers 82 offline tasks, five attempts each, 1080p screenshots and a 500-step cap. When a workflow has to finish, the 37.1% strict rate is the number that counts.
Third-Party Index Scores and Tool Gains
Artificial Analysis lists all five effort levels on Intelligence Index v4.3.2 with rounded task costs. HLE shows 45.9% unaided and 57.4% with tools — a lift that comes from the surrounding system.
Haiku 5.5 Benchmark: Common Questions
Answers on Claude Haiku 5.5 scores, token pricing, self-hosting limits and reasoning-effort trade-offs.
Where does Haiku 5.5 sit next to GPT-6 Luna and Sol?
Across knowledge work, screenshot jobs and narrow coding, Haiku and Luna trade wins. Sol and Sol 6.1 charge more and justify it on coding, yet Sonnet 5.5 keeps the lead on most rows.
What makes the 72.4% computer-use number fall short of a success rate?
The identical run yields just 37.1% once every checkpoint is required. A model may find and edit the correct file, then write it in the wrong format or halt before the final step. Always add an end-state check.
Where is the cutoff for Haiku 5.5's cheaper pricing?
The lowest rates apply only to prompts of 100,000 tokens or fewer; beyond that, each category costs five times as much. Anthropic's 90%, 50% and roughly 75% savings figures are measured against different baselines.
Could Haiku 5.5 be self-hosted, and what GPU memory would that take?
The published specs list no parameter count, active parameters, layer count or training compute. 'Small' describes a product tier rather than an architecture, so any self-hosting GPU estimate rests on nothing.
Do higher reasoning-effort settings always pay off?
No. Sol 6.1 actually scores higher at medium than at max, and Haiku's 0.2-point lead over Sonnet at max is too narrow to claim a win. More effort also means more tokens, more latency and more room to wander off task.
Why does one model post different numbers in different tables?
Sonnet shows 70.6% in Anthropic's launch table but 61.8% on the public leaderboard snapshot, where Haiku has no visible row. Artificial Analysis removed its 'Default Fallback' tag, while Anthropic's migration guide still states that no server-side fallback exists.
Run the Haiku 5.5 Benchmark Against Your Own Workload
Choose a single task, set the effort level, and measure Haiku 5.5 against the quality bar your team already uses.