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
A field guide to the Haiku 5.5 Benchmark: what each published task score, price tier and context ceiling covers, and where the numbers get misread.
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.
Seedance 2.0
The Future of AI Video Is Here.
AI Multi-Scene Shorts Generator
Create viral AI Shorts instantly
3D Science Video
Create 3D science videos easily
Suno AI Music Generator
Create Professional Music with AI
Free GPT Image 2.5
Next-generation image creator
What the Haiku 5.5 Benchmark Numbers Are Worth
Launch-day figures, token price tiers and per-effort costs, each one filed next to the configuration and date it was pulled.
- Mixed Scales Sit Side by SideElo ratings, partial-credit scores, coding pass rates and accuracy figures follow different scales, so averaging them produces noise.
- Built for Volume, Not One-OffsExtraction, labelling and summarisation loops that run thousands of times a day are where the low per-token rate pays off.
- Two Footnotes That Reshape the TablesThe budget rate ends at 100K prompt tokens, and the 72.4% computer-use figure is partial credit rather than a completed task.
Three Steps From Haiku 5.5 Tables to a Workload Call
A short path from a published result table to a decision you can defend in a planning meeting.
Haiku 5.5 Benchmark: The Findings That Matter
Spec sheet, token tiers, coding and computer-use outcomes, plus third-party index readings for Claude Haiku 5.5.
Context Ceiling and Output Cap
Text and images go in, text comes out, with adaptive thinking enabled. A 1M-token window and a 128K output ceiling leave Haiku 4.5's 200K and 64K far behind.
Pricing Split at 100K Tokens
Past 100,000 prompt tokens, input, output and cache pricing all multiply by five. OpenAI's standard tier, by contrast, runs to 272,000 input tokens before it moves.
Terminal-Bench 4.0 Task Completion
The 39.2% launch mark clears Luna comfortably yet trails both Sol entries. Runs allow an eight-hour agent timeout, and any 2.1 or 3.0 figures come from an earlier harness.
FrontierCode: Mergeable Patches per Dollar
Scoring weighs correctness, tests, diff scope, style and repo conventions. Max effort reaches 46.4% and medium 41.6% at $0.13 per rollout; climbing from xhigh to max doubles the bill for half a point.
OSWorld 2.1: Credit Versus Completion
The setup: 82 offline tasks, five tries each, 1080p screenshots and a 500-step cap. If a workflow has to finish on its own, plan around the 37.1% strict-completion rate.
Third-Party Index and Tool Gains
Artificial Analysis lists all five effort levels on Intelligence Index v4.3.2 with rounded task costs. HLE moves from 45.9% to 57.4% once tools are added, which credits the surrounding system as much as the model.
Haiku 5.5 Benchmark Questions, Answered
Straight answers on Claude Haiku 5.5 scores, token pricing, context ceilings and what the published tables leave out.
How does Haiku 5.5 stack up against GPT-6 Luna and Sol?
Luna and Haiku trade wins across knowledge work, screenshot handling and small coding jobs. Sol and Sol 6.1 charge more and justify it on coding, and Sonnet 5.5 still tops most of the rows.
Why does the 72.4% computer-use figure not equal a success rate?
Require every checkpoint and the same run drops to 37.1%. An agent may find and edit the right file, then save it in the wrong format or quit one action early, so add an end-state check.
At what point does Haiku 5.5 move to the higher price band?
The cheapest rates apply up to 100,000 prompt tokens; beyond that every category costs five times as much. Anthropic's 90%, 50% and roughly 75% savings claims each measure against a different baseline.
Can Haiku 5.5 be self-hosted, and what GPU memory would that need?
The published specs give no parameter count, no active-parameter figure, no layer count and no training compute. 'Small' describes a product tier rather than an architecture, so any local GPU estimate rests on nothing.
Does raising reasoning effort always lift the scores?
No. Sol 6.1 actually scores higher at medium than at max, and Haiku's 0.2-point lead over Sonnet at max sits inside the noise. More effort buys more tokens, more latency and more chances to wander off task.
Why do identical models post different numbers across tables?
Sonnet shows 70.6% in Anthropic's launch table and 61.8% on the public leaderboard snapshot, where Haiku had no visible row at all. Artificial Analysis has since removed its 'Default Fallback' label, while Anthropic's migration guide still states that no server-side fallback exists.
Try the Haiku 5.5 Benchmark on a Real Task of Yours
Choose a single task, fix the effort level, and measure the output against the quality bar your team already uses.