Claude Haiku 5.5 Benchmarks
Last updated: October 8, 2026
Anthropic's October 7, 2026 release put Claude Haiku 5.5 on seven benchmarks, each measured against Haiku 4.5, GPT-6 Luna, and Sonnet 5.5.[1] The table also places the small tier against a GPT-6-class model — GPT-6 Luna — and Haiku 5.5 leads on every row where both appear.
Below is the full table, copied from the release page, followed by a row-by-row reading and the things the table does not tell you. An em dash means Anthropic did not list that model on that row.
The full score table
| Benchmark | Claude Haiku 5.5 | Claude Haiku 4.5 | GPT-6 Luna | Claude Sonnet 5.5 |
|---|---|---|---|---|
| GDPval-AA v2.1 (knowledge work) | 1620 | 735 | 1437 | 1840 |
| AA-Briefcase v1.1 | 1578 | 614 | 1336 | 1824 |
| OSWorld 2.1 (computer use, offline subset) | 72.4% | 15.7% | 48.9% | 83.9% |
| Humanity's Last Exam (no tools / with tools) | 45.9% / 57.4% | 10.2% / 18.7% | — | 56.9% / 64.5% |
| Terminal-Bench 4.0 (agentic coding) | 39.2% | 0.0% | 16.4% | 70.6% |
| FrontierCode 1.1 (Main) | 46.4% | — | 42.4% | 52.1% (Xhigh) |
| Chartography | 46.4% | 6.4% | 29.1% | 61.6% |
All figures from Anthropic's official release table, October 7, 2026.[1] An em dash (—) means the model is not listed on that row, not a zero. Anthropic does not publish per-effort-setting scores as numbers on the release page; Sonnet 5.5's FrontierCode entry is the only one carrying an effort annotation (Xhigh).
Reading the table, row by row
- GDPval-AA v2.1 (knowledge work) — 1620 vs 735. Haiku 5.5 more than doubles Haiku 4.5's score and edges past GPT-6 Luna's 1437. Sonnet 5.5's 1840 is the only score above it in this table.
- AA-Briefcase v1.1 — 1578 vs 614. The same shape holds for agentic knowledge work: roughly two and a half times the predecessor's score, above GPT-6 Luna (1336), below Sonnet 5.5 (1824).
- OSWorld 2.1 (computer use, offline subset) — 15.7% to 72.4%. The largest percentage-point jump from Haiku 4.5 to Haiku 5.5 in the table, and it lands the small tier within reach of Sonnet 5.5's 83.9% — ahead of GPT-6 Luna's 48.9%.
- Humanity's Last Exam — 10.2% to 45.9% without tools, 18.7% to 57.4% with them. With tools, Haiku 5.5 closes most of the distance to Sonnet 5.5 (64.5%). GPT-6 Luna is not listed on this row.
- Terminal-Bench 4.0 — 0.0% to 39.2%. The loudest single line in the release. The previous small tier did not register on agentic coding at all; the new one does, while Sonnet 5.5 sits at 70.6% — which is why Anthropic frames Haiku 5.5 as a coding subagent rather than a driver.
- FrontierCode 1.1 (Main) — 46.4% vs GPT-6 Luna's 42.4%. One of the two rows where Luna comes closest. Note the labels: Haiku 5.5's entry carries no effort annotation, Sonnet 5.5's 52.1% is at Xhigh, and Haiku 4.5 is not listed.
- Chartography — 6.4% to 46.4%. Same story as OSWorld: a small-tier score that used to be negligible now sits above GPT-6 Luna (29.1%), with Sonnet 5.5 at 61.6%.
Two patterns survive the row-by-row read. First, Haiku 5.5 beats GPT-6 Luna on all six rows where both are listed — including FrontierCode, where the margin is narrow. Second, the tier separation inside Claude is intact: Sonnet 5.5 leads every shared row, by the widest margins on Terminal-Bench (70.6% vs 39.2%) and OSWorld (83.9% vs 72.4%).
What the table does not tell you
- Per-effort-level scores. The release presents Haiku 5.5's quality plot as a chart, not as numbers. Which exact scores correspond to which effort settings is Not yet confirmed, and we do not estimate them.
- Context window, rate limits, throughput. None of these are stated for Haiku 5.5 on the release page. Not yet confirmed — left blank rather than filled with third-party guesses.
- Independent verification. Every figure here is vendor-run. OSWorld is the offline subset, which narrows what the number covers even within that benchmark.
Vendor-reported customer numbers
The release also carries early-customer figures, published by Anthropic from the companies themselves. They are directionally interesting but vendor-reported: HubSpot reports 92.8% task success for CRM tasks; Asana reports a latency drop of more than 30% with reasoning around 2.5 times faster per agent turn; Box reports an 11-point improvement with roughly half the latency.[1] None of these are independent benchmarks, and Anthropic does not publish the underlying evaluation setups.
Frequently asked questions
Who ran these benchmarks?
Anthropic ran them, and the scores are published on its official Claude Haiku 5.5 release page dated October 7, 2026 — including the Haiku 4.5, GPT-6 Luna, and Sonnet 5.5 columns. No independent reproduction was used for this page; every figure here is vendor-run.
Why does Haiku 4.5 show no FrontierCode score?
Because Anthropic does not list one on that row — the release table shows an em dash. It does not mean 0.0%; an em dash simply means the model is not listed for that benchmark. The page does not say why, so we label it as not listed rather than guessing.
What does the Xhigh label on Sonnet 5.5's FrontierCode score mean?
It means Sonnet 5.5's 52.1% on FrontierCode 1.1 was measured at the Xhigh effort setting. Haiku 5.5 is the first Haiku-tier model with an adjustable effort setting, but the release does not publish per-effort-level scores as numbers, so you cannot yet compare the two models at matched effort on this row.
Is Haiku 5.5 better than GPT-6 Luna?
On this table, yes, everywhere the two are listed together. Anthropic lists both models on six rows — GDPval-AA, AA-Briefcase, OSWorld, Terminal-Bench, FrontierCode, and Chartography — and Haiku 5.5 scores higher on all six. Caveats: the table is vendor-run, and GPT-6 Luna's effort settings are not stated, so read it as one published comparison rather than a settled ranking.
Sources
- Anthropic — Claude Haiku 5.5 (official release page) Published October 7, 2026; the source of every score in the table below and of the vendor-reported customer figures.
Every figure on this page comes from Anthropic's official Claude Haiku 5.5 release page, published October 7, 2026. VerdictApi did not use third-party estimates; anything the release does not state is labeled Not yet confirmed.
Last updated: October 8, 2026