Claude Haiku 5.5 vs GPT-6 Luna
Humanity's Last Exam Humanity's Last Exam: Very hard expert questions across many fields, written to stump today’s models
Haiku 5.533.8
GPT-6 Luna28.3
AA Omniscience Index AA Omniscience Index: Scores factual knowledge while penalizing incorrect answers to encourage admitting uncertainty
Haiku 5.54.5
GPT-6 Luna-5.1
SimpleQA Verified v1 SimpleQA Verified v1: Accuracy on short factual questions with verified answers, answered without web access
Haiku 5.523.9
GPT-6 Luna · max41.4
CritPt (physics) CritPt (physics): Research-level physics problems that need many careful reasoning steps
Haiku 5.512.9
GPT-6 Luna10.6
AA Intelligence Index 4.3.2 AA Intelligence Index 4.3.2: Artificial Analysis combines tests of knowledge, reasoning, coding and tool use into one score
Haiku 5.534.5
GPT-6 Luna29.9
AA Omniscience · Accuracy AA Omniscience · Accuracy: Measures correct answers on the AA Omniscience factual knowledge test
Haiku 5.534
GPT-6 Luna43.1
AA Omniscience · Hallucination Rate AA Omniscience · Hallucination Rate: Measures how often attempted answers are incorrect on the AA Omniscience factual knowledge test · lower is better
Haiku 5.544.8
GPT-6 Luna84.7
AA-LCR (long context) 1.1 AA-LCR (long context) 1.1: Questions that need reading and connecting facts across very long documents (around 100K tokens)
Haiku 5.577.3
GPT-6 Luna78.3
GDP.pdf · AA all-pass GDP.pdf · AA all-pass: Models reason over professional PDF documents; an attempt passes only when every atomic criterion passes
Haiku 5.515.2
GPT-6 Luna13.8
MLCR-AA MLCR-AA: Models synthesize long medical records; the headline credits concise answers judged both accurate and complete
Haiku 5.5 · max63.3
GPT-6 Luna · max16.1
ARC-AGI-1 · public evaluation ARC-AGI-1 · public-eval: Models infer transformations from examples and solve unseen grid puzzles; the evaluation split is explicit
Haiku 5.574
GPT-6 Luna70.1
ARC-AGI-1 · semi-private ARC-AGI-1 · semi-private: Models infer transformations from examples and solve unseen grid puzzles; the evaluation split is explicit
Haiku 5.557.3
GPT-6 Luna61
ARC-AGI-2 · public evaluation ARC-AGI-2 · public-eval: Models infer transformations from examples and solve unseen grid puzzles; the evaluation split is explicit
Haiku 5.513.8
GPT-6 Luna19.2
ARC-AGI-2 · semi-private ARC-AGI-2 · semi-private: Models infer transformations from examples and solve unseen grid puzzles; the evaluation split is explicit
Haiku 5.59.4
GPT-6 Luna18.1
ARC-AGI-3 · semi-private ARC-AGI-3 · semi-private: Measures interactive game solving through relative human action efficiency
Haiku 5.5 · max0.5
GPT-6 Luna0.2
Chess Puzzles Chess Puzzles: Models choose the best next move from a novel chess position
Haiku 5.5 · max55
GPT-6 Luna · max36
FrontierMath Tier 4 (v2) FrontierMath Tier 4 (v2): Models solve expert-written advanced mathematics problems on the named tier and edition
Haiku 5.5 · max46.3
GPT-6 Luna · max56.1
FrontierMath Tiers 1–3 (v2) FrontierMath Tiers 1–3 (v2): Models solve expert-written advanced mathematics problems on the named tier and edition
Haiku 5.5 · max75.1
GPT-6 Luna · max78.9
Mystery Game Puzzles Mystery Game Puzzles: Models choose the best move in puzzle positions from an undisclosed game
Haiku 5.5 · max30
GPT-6 Luna · max7
Vals MedCode Vals MedCode: Medical coding accuracy on patient encounters
Haiku 5.551.5
GPT-6 Luna44.7
Vals MedScribe Vals MedScribe: Accuracy of clinical notes produced from consultation transcripts
Haiku 5.589.5
GPT-6 Luna83.7
Vals Legal Research Bench Vals Legal Research Bench: Grounded legal research; a task passes only if all criteria pass
Haiku 5.543.3
GPT-6 Luna30.3
Vals Science Index Vals Science Index: Scientific problem solving across the index’s component evaluations
Haiku 5.538.5
GPT-6 Luna28.4
Arena Text · overall Arena Text · overall: Human pairwise preferences in the stated text vote pool
Haiku 5.5—
GPT-6 Luna · max1,443
Arena Text · Chinese Arena Text · chinese: Human pairwise preferences in the stated text vote pool
Haiku 5.5—
GPT-6 Luna · max1,428.3
Arena Text · English Arena Text · english: Human pairwise preferences in the stated text vote pool
Haiku 5.5—
GPT-6 Luna · max1,393
Arena Text · French Arena Text · french: Human pairwise preferences in the stated text vote pool
Haiku 5.5—
GPT-6 Luna · max1,407.3
Arena Text · German Arena Text · german: Human pairwise preferences in the stated text vote pool
Haiku 5.5—
GPT-6 Luna · max1,387.2
Arena Text · instruction following Arena Text · instruction following: Human pairwise preferences in the stated text vote pool
Haiku 5.5—
GPT-6 Luna · max1,410
Arena Text · Japanese Arena Text · japanese: Human pairwise preferences in the stated text vote pool
Haiku 5.5—
GPT-6 Luna · max1,365.1
Arena Text · Korean Arena Text · korean: Human pairwise preferences in the stated text vote pool
Haiku 5.5—
GPT-6 Luna · max1,367.7
Arena Text · non-English Arena Text · non english: Human pairwise preferences in the stated text vote pool
Haiku 5.5—
GPT-6 Luna · max1,383.3
Arena Text · Polish Arena Text · polish: Human pairwise preferences in the stated text vote pool
Haiku 5.5—
GPT-6 Luna · max1,382.5
Arena Text · Russian Arena Text · russian: Human pairwise preferences in the stated text vote pool
Haiku 5.5—
GPT-6 Luna · max1,393.4
Arena Text · Spanish Arena Text · spanish: Human pairwise preferences in the stated text vote pool
Haiku 5.5—
GPT-6 Luna · max1,393.1
| Benchmark | Haiku 5.5 | GPT-6 Luna |
|---|---|---|
| Humanity's Last Exam | 33.8 | 28.3 |
| AA Omniscience Index | 4.5 | -5.1 |
| SimpleQA Verified v1 | 23.9 | 41.4 · max |
| CritPt (physics) | 12.9 | 10.6 |
| AA Intelligence Index 4.3.2 | 34.5 | 29.9 |
| AA Omniscience · Accuracy | 34 | 43.1 |
| AA Omniscience · Hallucination Rate | 44.8 | 84.7 |
| AA-LCR (long context) 1.1 | 77.3 | 78.3 |
| GDP.pdf · AA all-pass | 15.2 | 13.8 |
| MLCR-AA | 63.3 · max | 16.1 · max |
| ARC-AGI-1 · public evaluation | 74 | 70.1 |
| ARC-AGI-1 · semi-private | 57.3 | 61 |
| ARC-AGI-2 · public evaluation | 13.8 | 19.2 |
| ARC-AGI-2 · semi-private | 9.4 | 18.1 |
| ARC-AGI-3 · semi-private | 0.5 · max | 0.2 |
| Chess Puzzles | 55 · max | 36 · max |
| FrontierMath Tier 4 (v2) | 46.3 · max | 56.1 · max |
| FrontierMath Tiers 1–3 (v2) | 75.1 · max | 78.9 · max |
| Mystery Game Puzzles | 30 · max | 7 · max |
| Vals MedCode | 51.5 | 44.7 |
| Vals MedScribe | 89.5 | 83.7 |
| Vals Legal Research Bench | 43.3 | 30.3 |
| Vals Science Index | 38.5 | 28.4 |
| Arena Text · overall | — | 1,443 · max |
| Arena Text · Chinese | — | 1,428.3 · max |
| Arena Text · English | — | 1,393 · max |
| Arena Text · French | — | 1,407.3 · max |
| Arena Text · German | — | 1,387.2 · max |
| Arena Text · instruction following | — | 1,410 · max |
| Arena Text · Japanese | — | 1,365.1 · max |
| Arena Text · Korean | — | 1,367.7 · max |
| Arena Text · non-English | — | 1,383.3 · max |
| Arena Text · Polish | — | 1,382.5 · max |
| Arena Text · Russian | — | 1,393.4 · max |
| Arena Text · Spanish | — | 1,393.1 · max |
Specs
| Spec | Haiku 5.5 | GPT-6 Luna |
|---|---|---|
| Input, per 1M | $0.10 | $0.10 |
| Output, per 1M | $0.50 | $0.50 |
| Cost per task? | $0.047 | $0.017 |
| Speed, tok/s | 179 | — |
| Context | 1M | 1.05M |
| Weights | Closed | Closed |
| Released | Oct 2026 | Sep 2026 |
| See images | ✓ | ✓ |
| Read PDFs | Partial | Partial |
| Hear audio | — | — |
| Watch video | — | — |
| Call tools | ✓ | ✓ |
| Force a tool call | Partial | Partial |
| Return strict JSON | ✓ | ✓ |
| Search the web | ✓ | ✓ |
| Run code | ✓ | ✓ |
| Use a computer | ? | ✓ |
What breaks when you switch
Haiku 5.5 → GPT-6 Luna
Function calling only with none · Chat Completions›
Chat Completions function calling requires reasoning_effort none; use Responses for reasoning with tools.
More than 272K input tokens›
Prompts above 272K input tokens charge 2× input and cache rates and 1.5× output rates for the entire request.
Remove unsupported sampling parameters›
When reasoning effort is not none, remove temperature, top_p and top_logprobs, plus Chat Completions logprobs or Responses output logprobs inclusion.
Cache writes billed at 1.25x›
Cache writes cost 1.25× the uncached input token rate.
GPT-6 Luna → Haiku 5.5
Context: 1.05M → 1M
Remove sampling parameters›
Unsupported sampling values, any top_k or combined temperature and top_p return HTTP 400; omit all three and use prompting.
Replace manual thinking budgets›
Manual thinking with budget_tokens returns HTTP 400; replace the old thinking configuration.
No assistant prefill›
A final assistant prefill returns HTTP 400 even with thinking off; end messages with a user turn.
Replace old computer tool · Claude API / Google Cloud›
computer_20250124 returns HTTP 400 on the Claude API and Google Cloud; use computer_toolset_20260801 there.
Keep conversations append-only›
Replaying thinking after earlier history changes returns HTTP 400, with older accounts checked only when the binding field is set; keep history append-only.
Bedrock needs another computer version · Bedrock›
Bedrock does not accept computer_20250124; use computer_20251124 with the computer-use-2025-11-24 beta header.