Across 40 comparable benchmarks, Claude Haiku 5.5 has better reported results on 30 and worse on 10. 2 further shared results are listed below, each with the reason it is not counted.
Strong evidence40 of 42 shared results countedHow we compare
Price basis: Listed rates vs listed rates. The price section below shows the applicable periods.
| Fact | Claude Haiku 5.5 | GPT-6 Luna |
|---|---|---|
| Context window | 1M tokens | 1.05M tokens |
| Input / 1M | $0.100 | $0.100 |
| Output / 1M | $0.500 | $0.500 |
| Blended (4:1) | $0.115 | $0.115 |
| Status | Available | Available |
| Released | October 7, 2026 | September 22, 2026 |
Blended is 4:1 input to output with 90% cache hits at listed rates — set your own mix.
Editor’s take
LLMLearner Editorial · Figures as of · AI-assisted
Claude Haiku 5.5 and GPT-6 Luna both list at $0.10 input / $0.50 output per million tokens. Haiku 5.5 has the better reported result on 23 of 30 comparable benchmarks, with big leads on agentic coding. GPT-6 Luna wins on maths, tax and spreadsheet modelling.
When two models cost the same, the decision is purely about fit, which makes this my favourite small-model comparison of the month.
Haiku 5.5 is the agent and the coder. Terminal-Bench 4.0 39.2% vs 16.4%. Vibe Code Bench v1.1 90.44% vs 81.65%. APEX-SWE 55.1% vs 38.8%. ProofBench (Lean 4) 89% vs 64%. Legal Research Bench 43.27% vs 30.29%. AA-AnalystAgent 53.75% vs 43.75%. It also leads the Vals Index v2.2, 58.71 vs 52.32.
Luna is the quant. FrontierMath v2 78.95% vs 75.09%, FrontierMath Tier 4 56.1% vs 46.34%, Tax Agent Bench 58.86% vs 31.92%, Excel Modeling 68.52% vs 59.02%, IOI (Vals v2) 55.56% vs 47.28%. Neither is good at Zapier’s AutomationBench (Luna 20.7%, Haiku 12%).
The pricing fine print differs. Both are $0.10 / $0.50, but Haiku’s rate applies to prompts up to 100K tokens (above that it is $0.50 / $2.50), while Luna’s applies up to 272K input tokens. If you routinely send prompts between 100K and 272K tokens, Luna is meaningfully cheaper. Both have a context window of about 1M tokens and 128K max output.
My pick: Haiku 5.5 for coding assistants, review bots and anything that runs tools in a loop. GPT-6 Luna for spreadsheet, finance and maths-heavy pipelines, and for long prompts in the 100K–272K range.
Mean of 24 shared benchmarks grouped by capability on a 0–100 scale. How averages are built
Claude Haiku 5.5 ahead on 10 of 12 counted
Claude Haiku 5.5 ahead on 2 of 2 counted
GPT-6 Luna ahead on 2 of 3 counted
Claude Haiku 5.5 ahead on 4 of 4 counted
Claude Haiku 5.5 ahead on 2 of 2 counted
Claude Haiku 5.5 ahead on 2 of 2 counted
GPT-6 Luna ahead on 4 of 7 counted
No shared comparable results for Research & search, Long context, Cybersecurity.
Published scores, with matching modes where available. Gaps show Claude Haiku 5.5 relative to GPT-6 Luna. Reading the results
| Benchmark | Claude Haiku 5.5 | GPT-6 Luna | Gap |
|---|---|---|---|
| APEX-SWERepository Engineering · Pass@1 | 55.1Max · tools | 38.8Max · tools | +16.3 pts#7 vs #35 of 61 |
| Code MigrationMaintenance & Optimization · Hidden-test pass rate (%) | 44.6Max · tools | 42.5Max · tools | +2.0 pts#21 vs #29 of 79 |
| DeepSWERepository Engineering · Pass@1 (DeepSWE v1.1) | 59.3Unspecified Reasoning · tools | 66.6≠Max · tools | −7.3 pts#49 vs #37 of 73 |
| FrontierCode 1.1Code Generation & Editing · Weighted rubric score (%) | 46.4Max · tools· frontiercode-1-1-main | 42.4≠Reported best· frontiercode-1-1-main | +4.0 pts#15 vs #23 of 51 |
| FrontierCode 2.0Repository Engineering · Weighted rubric score | 27.7Max · tools + web | 21.9Max · tools + web | 1.27×#13 vs #21 of 36 |
| IOI (Vals v2)Algorithmic Coding · Subtask points (%) | 47.3Max · tools | 55.6Max · tools | −8.3 pts#30 vs #22 of 45 |
| SWE-Bench Pro V2 HardRepository Engineering · Resolve Rate (%, pass@1) | 90.2Unspecified Reasoning · tools | 80.4Unspecified Reasoning · tools | +9.8 pts#4 vs #16 of 21 |
| Terminal-Bench 2.1Agentic Development · Accuracy | 94.4Unspecified Reasoning · tools | 82.0Unspecified Reasoning · tools | +12.4 pts#1 vs #35 of 68 |
| Terminal-Bench 4.0Agentic Development · Resolution rate (%) | 39.2Max · tools | 16.4≠Reported best | +22.8 pts#28 vs #53 of 90 |
| Terminal-Bench 4.1Agentic Development · Resolution rate (%) | 40.0Max · tools | 15.2Max · tools | +24.9 pts#19 vs #36 of 48 |
| Vibe Code Bench v1.1Code Generation & Editing · Pass rate (%) | 90.4Max · tools | 81.7Max · tools | +8.8 pts#5 vs #29 of 121 |
| WeirdML v3ML Engineering · Official score (0-1; higher is better) | 0.12Extra-High · tools | 0.079Extra-High · tools | 1.49×#13 vs #17 of 26 |
| ARC-AGI-1Abstract Generalization · Score (%) | 89.0Max · no tools | 86.7Max · no tools | +2.3 pts#36 vs #44 of 87 |
| ARC-AGI-2Abstract Generalization · Score (% solved); cost per task (USD)Not counted · Speed and cost telemetry is reported, not scored | 61.3Extra-High · no tools | 59.3≠Max · no tools | 1.03×#34 vs #40 of 80 |
| ARC-AGI-3Abstract Generalization | 2.8Max · no tools· arc3-semi-private-provider-adapter | 0.59Max · no tools· arc3-semi-private-provider-adapter | 4.78×#13 vs #14 of 15 |
| FrontierMath Tier 4 v2Mathematics · Accuracy (verification_code) | 46.3Max · tools | 56.1Max · tools | −9.8 pts#24 vs #21 of 73 |
| FrontierMath v2Mathematics · Accuracy (verification_code) | 75.1Max · tools | 78.9Max · tools | −3.9 pts#23 vs #21 of 85 |
| ProofBench v1.1 (Lean 4)Mathematics · Accuracy (%) | 89.0Max · tools | 64.0Max · tools | +25.0 pts#13 vs #23 of 54 |
| MysteryMechanismScientific Reasoning · Accuracy (%) | 35.1Max · tools | 19.4Max · tools | +15.8 pts#13 vs #28 of 32 |
| HealthBench ProfessionalClinical Workflows · Length-adjusted score (%) | 64.8Max · no tools· length-adjusted | 60.8≠Reported best· length-adjusted | +4.0 pts#6 vs #12 of 19 |
| MedCodeClinical Workflows · Accuracy (%) | 51.5Max · tools | 44.7Max · tools | +6.9 pts#15 vs #39 of 97 |
| MedScribeClinical Workflows · Rubric pass rate (%) | 89.5Max · tools | 83.7Max · tools | +5.8 pts#6 vs #41 of 99 |
| APEX-AgentsTool Orchestration · Score | 56.1Max · tools | 44.3Max · tools | 1.27×#26 vs #43 of 61 |
| OSWorld 2.1Desktop Workflows · Partial score | 68.9Max · tools· osworld-full-108 | 50.5Max · tools· osworld-full-108 | 1.36×#13 vs #17 of 18 |
| ChartographyDocuments & Charts · Score | 86.2Max · tools | 29.1≠Reported best | 2.96×#6 vs #27 of 29 |
| PerceptionBenchVisual Understanding · Accuracy | 78.6Max · tools | 62.8Max · tools | +15.8 pts#2 vs #8 of 19 |
| AutomationBenchOffice & Business · Pass RateNot counted · Different or unrecorded test environments — excluded from the tally | 12.0Max · tools· zapier-private-heldout | 38.1Max · tools· legacy-default | —#48 vs #26 of 64 |
| EMB (Excel Modeling)Finance · Accuracy (%) | 59.0Max · tools | 68.5Max · tools | −9.5 pts#39 vs #18 of 78 |
| Finance Agent v2Finance · Score | 54.1Unspecified Reasoning · tools | 49.9≠Max · tools | 1.09×#27 vs #43 of 75 |
| GDPxCross-industry Work · GDPx — Score | 4.9Max · tools· overall/score | 5.1Max · tools· overall/score | 0.97×#13 vs #12 of 20 |
| Legal Research BenchLegal · Score | 43.8Max · tools | 30.3Max · tools | 1.44×#26 vs #55 of 91 |
| Public Benefits Bench v1.1Service Workflows · Criterion pass rate (%) | 63.2Max · tools | 57.6Max · tools | +5.5 pts#25 vs #38 of 50 |
| SAGEService Workflows · Accuracy (%) | 45.8Max · tools | 48.1Max · tools | −2.3 pts#38 vs #28 of 82 |
| Tax Agent BenchFinance · Accuracy (%, rubric × citation quality) | 31.9Unspecified Reasoning · tools | 58.9≠Max · tools | −26.9 pts#38 vs #19 of 74 |
| HLEKnowledge Exams · Accuracy | 57.4Max · tools + web | 49.7≠Max · tools | +7.7 pts#21 vs #52 of 151 |
| AA Decision Index v2.1Capability Indices · Index (0-100) | 76.5Max · no tools | 75.7Max · no tools | 1.01×#1 vs #2 of 5 |
| AA Intelligence Index v5.0.1Capability Indices · Index (0-100) | 36.5Max · task tools | 27.9Max · task tools | 1.31×#20 vs #36 of 62 |
| Agent ArenaPreference Arenas · Net improvement index (percentage points vs. average model) | 0.42Thinking High · tools + web | 1.5≠Max · tools + web | 0.28×#29 vs #26 of 35 |
| Agent Arena IndexPreference Arenas · Agent Arena Index rating (index points) | 1095Thinking High · tools + web | 1104≠Max · tools + web | 0.99×#29 vs #28 of 36 |
| LiveBenchCross-capability Suites · Accuracy | 72.1Extra-High · no tools | 72.0≠Max · no tools | +0.0 pts#49 vs #51 of 106 |
| LiveBench 2.0Cross-capability Suites · Overall score (equal-weight mean of 7 suite scores, 0-100) | 75.6Max · task tools | 69.4Max · task tools | 1.09×#16 vs #26 of 51 |
| Vals Index v2.2Capability Indices · GDP-weighted accuracy (%) | 58.7Max · task tools | 52.3Max · task tools | +6.4 pts#18 vs #28 of 36 |
Repository Engineering · Pass@1
Maintenance & Optimization · Hidden-test pass rate (%)
Repository Engineering · Pass@1 (DeepSWE v1.1)
Code Generation & Editing · Weighted rubric score (%)
Repository Engineering · Weighted rubric score
Algorithmic Coding · Subtask points (%)
Repository Engineering · Resolve Rate (%, pass@1)
Agentic Development · Accuracy
Agentic Development · Resolution rate (%)
Agentic Development · Resolution rate (%)
Code Generation & Editing · Pass rate (%)
ML Engineering · Official score (0-1; higher is better)
Mathematics · Accuracy (verification_code)
Mathematics · Accuracy (verification_code)
Mathematics · Accuracy (%)
Scientific Reasoning · Accuracy (%)
Clinical Workflows · Length-adjusted score (%)
Clinical Workflows · Accuracy (%)
Clinical Workflows · Rubric pass rate (%)
Tool Orchestration · Score
Desktop Workflows · Partial score
Documents & Charts · Score
Visual Understanding · Accuracy
Office & Business · Pass Rate
Not counted · Different or unrecorded test environments — excluded from the tally
Finance · Accuracy (%)
Finance · Score
Cross-industry Work · GDPx — Score
Legal · Score
Service Workflows · Criterion pass rate (%)
Service Workflows · Accuracy (%)
Finance · Accuracy (%, rubric × citation quality)
Knowledge Exams · Accuracy
Capability Indices · Index (0-100)
Capability Indices · Index (0-100)
Preference Arenas · Net improvement index (percentage points vs. average model)
Preference Arenas · Agent Arena Index rating (index points)
Cross-capability Suites · Accuracy
Cross-capability Suites · Overall score (equal-weight mean of 7 suite scores, 0-100)
Capability Indices · GDP-weighted accuracy (%)
A dash means no result is available.
| Benchmark | Claude Haiku 5.5 | GPT-6 Luna |
|---|---|---|
| CritPt | — | 19.4 |
| Riemann-bench | — | 59.2 |
| SciCode | — | 54.6 |
| AA Cyber Index | — | 52.7 |
| CursorBench 4.0 | 48.4 | — |
| FrontierSWE v2 | 43.8 | — |
| ProgramBench | — | 6 |
| SWE-bench Multilingual | 83.7 | — |
| SWE-bench Multimodal | 30.7 | — |
| Vibe Code Bench 1-100 | 20.4 | — |
| Agents' Last Exam v1.0 | — | 25.0 |
| OSWorld 2.0 | — | 52.7 |
| Vals Index | — | 58.5 |
| Vals Index v2.1 | — | 51.2 |
| Hypercritica Model Bench | 67.9 | — |
| EBR-bench | 43.8 | — |
| GDP.pdf | — | 22.8 |
| MMMU-Pro | — | 79.9 |
| ZeroBench Main | — | 33.6 |
| AA-AnalystAgent | 53.8 | — |
| AA-Briefcase | 1578 | — |
| GDPval-AA v2 | 1620 | — |
| Harvey's Legal Agent Benchmark | — | 2.9 |
| GoBench | — | 1120 |
| SimpleBench | 59.2 | — |
| Vending-Bench 2 | 3602 | — |
| BioMysteryBench | — | 61.5 |
| HealthBench | — | 54.5 |
Time-of-day providers use peak-hour rates; fixed-price providers use standard rates. Off-peak discounts remain in the published details.
Drag the handle above the bar for cache hits, the one below it for the input : output mix · 5% steps
Blended price = effective input rate at 90% cache hits × 80% + output rate × 20%
Claude Haiku 5.5 and GPT-6 Luna costabout the sameat this token mix.Using listed rates vs listed rates at this token mix.
Claude Haiku 5.5: Standard pricing. context_length <= 100000 Prompt length tier; text and image inputs are token billed. context_length > 100000
GPT-6 Luna: Standard pricing. context_length <= 272000 OpenAI official pricing (developers.openai.com/api/docs/pricing) verified 2026-10-21; standard tier; <=272K input tokens context_length > 272000 OpenAI official pricing (developers.openai.com/api/docs/pricing) verified 2026-10-21; standard tier; >272K input tokens
| Price / 1M tokens | Claude Haiku 5.5 | GPT-6 Luna | Ratio |
|---|---|---|---|
| Input | $0.10 | $0.10 | 1.00× |
| Output | $0.50 | $0.50 | 1.00× |
| Cached input | $0.01 | $0.01 | 1.00× |
| Cache write | $0.125 | $0.125 | 1.00× |
| Blended | $0.115 | $0.115 | 1.00× |
Estimates use the selected token mix. Cache writes, tool fees and long-context surcharges are excluded.
Reported cost per task, paired with the score from the same run. Settings and sources may differ; costs do not affect the verdict.
Claude Haiku 5.5 relative to GPT-6 Luna
Summary: 12 of 14 shared benchmarks have matching recorded settings and sources. The average is a geometric mean of 12 positive-cost ratios, one equal weight per benchmark. This benchmark set can include related tests and indexes; it is not a workload cost estimate.
Claude Haiku 5.5 13.2× more expensive
Claude Haiku 5.5 8.9× more expensive
Claude Haiku 5.5 7.2× more expensive
Claude Haiku 5.5 7× more expensive
Claude Haiku 5.5 7× more expensive
Claude Haiku 5.5 4.3× more expensive
Claude Haiku 5.5 4× more expensive
Claude Haiku 5.5 2.8× more expensive
Claude Haiku 5.5 13.2× more expensive
Claude Haiku 5.5 8.9× more expensive
Claude Haiku 5.5 7.2× more expensive
Claude Haiku 5.5 7× more expensive
Claude Haiku 5.5 7× more expensive
Claude Haiku 5.5 4.3× more expensive
Claude Haiku 5.5 4× more expensive
Claude Haiku 5.5 2.8× more expensive
Effort, protocol, aggregation and source match; other run conditions may differ.
Claude Haiku 5.5
GPT-6 Luna
Effort, protocol, aggregation and source match; other run conditions may differ.
Claude Haiku 5.5
GPT-6 Luna
Settings or sources differ or are not recorded; interpret the comparison with care.
Claude Haiku 5.5
GPT-6 Luna
Settings or sources differ or are not recorded; interpret the comparison with care.
Claude Haiku 5.5
GPT-6 Luna
Effort, protocol, aggregation and source match; other run conditions may differ.
Claude Haiku 5.5
GPT-6 Luna
Effort, protocol, aggregation and source match; other run conditions may differ.
Claude Haiku 5.5
GPT-6 Luna
Effort, protocol, aggregation and source match; other run conditions may differ.
Claude Haiku 5.5
GPT-6 Luna
Effort, protocol, aggregation and source match; other run conditions may differ.
Claude Haiku 5.5
GPT-6 Luna
Effort, protocol, aggregation and source match; other run conditions may differ.
Claude Haiku 5.5
GPT-6 Luna
Effort, protocol, aggregation and source match; other run conditions may differ.
Claude Haiku 5.5
GPT-6 Luna
Effort, protocol, aggregation and source match; other run conditions may differ.
Claude Haiku 5.5
GPT-6 Luna
Effort, protocol, aggregation and source match; other run conditions may differ.
Claude Haiku 5.5
GPT-6 Luna
Effort, protocol, aggregation and source match; other run conditions may differ.
Claude Haiku 5.5
GPT-6 Luna
Effort, protocol, aggregation and source match; other run conditions may differ.
Claude Haiku 5.5
GPT-6 Luna
| Benchmark | Claude Haiku 5.5 | GPT-6 Luna |
|---|---|---|
| Agents' Last Exam v1.0 | Not reported | 3 more settings |
| Code Migration | Not reported | |
| CursorBench 4.0 | $0.32 / task Average · Task score: 42.3 Thinking High (With Tools) · cursorbench-4-0-cursor-agent $0.08 / task Average · Task score: 30.9 Thinking Low (With Tools) · cursorbench-4-0-cursor-agent 2 more settings$0.17 / task Average · Task score: 36.9 Thinking Medium (With Tools) · cursorbench-4-0-cursor-agent $0.56 / task Average · Task score: 44.3 Extra-High (With Tools) · cursorbench-4-0-cursor-agent | Not reported |
| EMB (Excel Modeling) | Not reported | |
| Finance Agent v2 | Not reported | |
| FrontierCode 1.1 | $1.3284 / task Average · Weighted rubric score: 46.4 Max (With Tools) · frontiercode-1-1-main $0.256 / task Average · Weighted rubric score: 41.9 Thinking High (With Tools) · frontiercode-1-1-main $0.0596 / task Average · Weighted rubric score: 34.8 Thinking Low (With Tools) · frontiercode-1-1-main 2 more settings$0.13 / task Average · Weighted rubric score: 41.6 Thinking Medium (With Tools) · frontiercode-1-1-main $0.6297 / task Average · Weighted rubric score: 45.8 Extra-High (With Tools) · frontiercode-1-1-main | Not reported |
| FrontierSWE v2 | Not reported | |
| IOI (Vals v2) | Not reported | |
| Legal Research Bench | Not reported | |
| MedCode | Not reported | |
| MedScribe | Not reported | |
| MysteryMechanism | Not reported | |
| ProofBench v1.1 (Lean 4) | Not reported | |
| Public Benefits Bench v1.1 | Not reported | |
| SAGE | Not reported | |
| Tax Agent Bench | $1.1642 / task Average · Accuracy (%, rubric × citation quality): 31.9 Unspecified Reasoning (With Tools) | Not reported |
| Terminal-Bench 2.1 | Not reported | |
| Vals Index v2.1 | Not reported | |
| Vibe Code Bench 1-100 | Not reported | |
| Vibe Code Bench v1.1 | Not reported |
| Attribute | Claude Haiku 5.5 | GPT-6 Luna |
|---|---|---|
| Availability | Available | Available |
| Provider | Anthropic | OpenAI |
| Model type | Reasoning model | Reasoning model |
| First release | October 7, 2026 | September 22, 2026 |
| Context window | 1M tokens | 1.05M tokens |
| Max output | 128K tokens | 128K tokens |
| Architecture | Undisclosed | Undisclosed |
| Documented capabilities | Reasoning | Reasoning, Multilingual, Coding, Function calling, Tool use, Structured output, Computer use, Web search, Long context |
| Knowledge cutoff | 2026-06 | 2026-05-18 |
| Input modalities | Text, Image | Text, Image |
| Output modalities | Text | Text |
| Weights | Closed / API only | Closed / API only |
| Commercial use | Proprietary | Proprietary |
Head-to-head comparisons by others, picked and summarised by LLMLearner Editorial. Each links the original.
Artificial Analysis
Favours Claude Haiku 5.5
Claude Haiku 5.5 and GPT-6 Luna share the same base rates — $0.10 / $0.50 per million input/output tokens for shorter prompts (both charge more for very long prompts: Haiku above 100K tokens, Luna above 272K) — so Artificial Analysis's comparison comes down to what each does with them.
At max effort Haiku 5.5 scores 43 on the Intelligence Index against 38 for Luna, leads Terminal-Bench 4.0 by a wide margin (32.8% vs 12.6%) and generates about 240 tokens per second against 137. Luna spends fewer tokens, so its max-effort index task costs $0.07 against $0.21.
Matched on cost the picture tightens: Haiku 5.5 at high effort costs about $0.08 per task and scores 38, the same as Luna at max.
Roboflow Vision Evals
Mixed result
Across Roboflow's Vision Evals (detection, counting, identification, OCR and visual reasoning, pooled at low effort), Claude Haiku 5.5 averages 69.6% and GPT-6 Luna 69.7% — ranks 20 and 19 of 61.
Haiku 5.5 is ahead on visual reasoning (68.9% vs 64.2%) and OCR (61.2% vs 59.8%); Luna on identification (87.5% vs 83.3%) and counting at low effort. Luna is a little cheaper ($0.0004 vs $0.0006 per sample) and faster (10.2 s vs 11.0 s).
These are reported-score comparisons, not controlled head-to-head tests.
What people found comparing these models, or using either one. Comments also appear on the models' own pages.
Used Claude Haiku 5.5 or GPT-6 Luna? Sign in with Google to share what worked, what did not, and which API or setup you tested.
No community comments on either model yet.
Alternatives with the most benchmark overlap for this matchup.