Claude Haiku 5.5 Benchmarks: Artificial Analysis Scores & Batch Metrics

# Claude Haiku 5.5 Benchmarks: Artificial Analysis Scores & Batch Reality Check

People searching haiku 5.5 benchmarks and haiku 5.5 artificial analysis usually want two things: public leaderboard numbers, and whether those scores matter for production CSV classification. This page covers both.

Official product notes: What’s new in Claude Haiku 5.5. Independent leaderboard write-up: Artificial Analysis on Claude Haiku 5.5 (7 Oct 2026).

Headline public scores (as of Oct 2026)

How to read this: Artificial Analysis ranks Haiku 5.5 as a leading small-class model at max effort — ahead of peers like GLM-5.3 Flash (~42), Gemini 3.8 Flash (~41), and GPT-6 Luna (~38), and close to much larger open models such as Kimi K3 (~44). It is not Sonnet-class intelligence.

Scores move when harnesses, effort settings, or safety filters change. Always check the live AA model page and Anthropic’s latest system card before locking a procurement decision.

Effort settings change the score (and the token bill)

Haiku 5.5 is the first Haiku with Anthropic effort + adaptive thinking. AA’s article highlights:

• Max Intelligence Index 43, but ~162k output tokens per Index task (~3× GPT-6 Luna max).
• Moving xhigh → max buys ~+2 Index points for ~1.8× tokens.
• At high effort, Index 38 with ~55k tokens/task — similar intelligence to Luna max, still slightly heavier on tokens.

For batch labeling, default-on thinking can inflate latency and credits even when the final label is one word. Prefer low/medium effort (or thinking disabled where allowed) for pure classification templates, and reserve max effort for hard extraction / agentic steps.

Pricing context that couples to benchmarks

AA notes Haiku 5.5 list pricing (Anthropic) roughly:

• $0.10 / $0.50 per 1M input/output tokens for prompts ≤100k
• $0.50 / $2.50 when prompts go above 100k (5× step-up)
• Cache reads / short cache writes are much cheaper than full input

Combined with the ~30% tokenizer inflation vs Haiku 4.5, sticker price alone understates effective $/1k CSV rows. See our Haiku 5.5 pricing guide.

Why chat/coding benchmarks are a weak proxy for batch CSV

Public indices optimize for:

• Multi-step coding / terminal / knowledge-work agent tasks
• Long reasoning traces under max effort
• Arena-style chat quality

BatchHaiku workloads usually care about:

A model that wins Terminal-Bench can still over-refuse ticket text or emit multi-sentence answers when you asked for a single label — both show up as production failures, not Intelligence Index points.

Haiku 5.5 specifics that skew harnesses

• Adaptive thinking may add tokens and latency vs a “text-only” baseline.
• Newer tokenizer increases token counts ≈ 30% vs Haiku 4.5 for the same string.
• Sampling params like temperature are rejected — keep harnesses Anthropic-5.5-compatible.
• AA flagged AutomationBench over-refusal during pre-release; treat that sub-score cautiously until re-run.

Suggested methodology on BatchHaiku

1. Upload a labeled sample CSV (20–50 rows).
2. Run the matching template (sentiment, ticket sorting, RAG filter, extraction).
3. Export results and score offline against gold labels (F1 + invalid-rate).
4. Compare credit burn and p95 latency at your chosen effort / thinking settings.
5. Scale to full corpora only after error budget passes.

Open the batch tool to collect live timings and credit burn on Cloudflare AI–backed Haiku inference.

Related

• Haiku 5.5 pricing & cost
• Haiku 5.5 vs Sonnet 5.5
• What is Claude Haiku 5.5?

BatchHaiku is an independent third-party tool, not affiliated with Anthropic or Artificial Analysis. Benchmark figures above are cited from public third-party / vendor reports and may change.