Compare published benchmark results from matching versions and cohorts, API pricing, and technical specifications. Any documented test setup differences remain visible.
shared result series
222
distinct benchmarks
37
latest retrieval
2026-09-29
Change model selection
Specifications
General specifications
Release date, availability, context window, architecture, and published parameter counts at a glance. Undisclosed details remain marked as unavailable.
The table compares official standard API rates per 1M tokens, including published cache pricing. Batch, fast, priority, regional surcharges, tool calls, and cloud platform pricing are excluded.
Key to the marks:better valuebetter, but not conclusivetie
Pricing by provider
Direct developer pricing and provider endpoint tariffs routed through OpenRouter are listed separately. Prices are in USD per 1M tokens. Context-dependent direct tiers plus regional, Flex, Priority, and other OpenRouter tariffs remain individually identifiable.
Only published results with the same benchmark version, task, metric, and comparison cohort are paired. Different reasoning levels, agents, harnesses, or output limits appear directly in the table.
Capability profile from matched benchmarks
Each axis averages directly matched benchmark families. A value of 100 means the stronger value within this pair, not a universal quality score.
GPT-5.6 Sol
GPT-5.6 Luna
Coding: 98.6 / 74.5 (8 benchmark families)
Reasoning: 100 / 76.4 (3 benchmark families)
Mathematics: 100 / 85.3 (2 benchmark families)
Knowledge: 100 / 96.7 (2 benchmark families)
Finance: 99 / 96.9 (7 benchmark families)
Legal: 99 / 94.7 (4 benchmark families)
Agentic tasks: 97.2 / 64.4 (3 benchmark families)
Multimodal: 100 / 95.7 (1 benchmark family)
216 matched rows from 35 benchmark families were considered. The radar shows 8 of 15 comparable categories.
GPT-5.6 SolGPT-5.6 Luna
ARC-AGI-2Overall. Higher is better. No documented setup difference
GPT-5.6 SolWinner85.4 %
GPT-5.6 Luna29.3 %
ARC-AGI-3RHAE overall score. Higher is better. Test setup not documented as identical
GPT-5.6 SolHigher measured value2.15 %
GPT-5.6 Luna0.18 %
BioMysteryBenchOverall. Higher is better. No documented setup difference
GPT-5.6 SolWinner71.111 %
GPT-5.6 Luna61.481 %
BrowseCompOverall. Higher is better. No documented setup difference
GPT-5.6 SolWinner90.4 %
GPT-5.6 Luna83.3 %
Code MigrationOverall. Higher is better. No documented setup difference
GPT-5.6 SolWinner52.916 %
GPT-5.6 Luna44.55 %
CorpFin v2Overall. Higher is better. No documented setup difference
GPT-5.6 SolWinner64.375 %
GPT-5.6 Luna64.219 %
CyberBench v1.1Overall. Higher is better. No documented setup difference
GPT-5.6 SolWinner76.31 %
GPT-5.6 Luna73.631 %
EMBOverall. Higher is better. No documented setup difference
GPT-5.6 SolWinner72.339 %
GPT-5.6 Luna67.116 %
Finance Agent (v2)Overall. Higher is better. No documented setup difference
GPT-5.6 Sol53.756 %
GPT-5.6 LunaWinner55.044 %
GPQA DiamondOverall. Higher is better. No documented setup difference
GPT-5.6 SolWinner95.202 %
GPT-5.6 Luna91.666 %
Harvey's Legal Agent BenchmarkOverall · Task fully resolved. Higher is better. No documented setup difference
GPT-5.6 SolWinner2.5 %
GPT-5.6 Luna1.25 %
IOIOverall. Higher is better. No documented setup difference
GPT-5.6 SolWinner91.167 %
GPT-5.6 Luna61.778 %
Key to the marks:better valuebetter, but not conclusivetie
Benchmark scores for GPT-5.6 Sol and GPT-5.6 Luna
Benchmark and task
GPT-5.6 Sol
GPT-5.6 Luna
Source
ARC-AGI-2OverallTest details
No documented difference in the test setupVersion: 2Metric: exact task accuracyScoring: Higher is betterSettings: Reasoning: high, Attempts: two exact output guesses per taskThe same verified test with five reasoning levels per model.
No documented difference in the test setupVersion: 1,266 tasksMetric: accuracyScoring: Higher is betterCross-vendor comparison table with browsing tools. Exact agent systems differ.
Winner90.4 %
Editorial selection from the vendor table, not a complete extract of the comparison cohort.
83.3 %
Editorial selection from the vendor table, not a complete extract of the comparison cohort.
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.6 Sol: Snapshot: gpt-5.6-sol-xhigh, Reasoning: xhigh, Harness: Arena text, style control GPT-5.6 Luna: Snapshot: gpt-5.6-luna-xhigh, Reasoning: xhigh, Harness: Arena text, style controlGPT-5.6 Sol: Style-controlled Arena rating from 34,258 anonymous pairwise votes.GPT-5.6 Luna: Style-controlled Arena rating from 35,823 anonymous pairwise votes.
Winner1,483.171
95% confidence interval: ±4.527The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,453.977
95% confidence interval: ±4.487The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
Test setup not documented as identicalNo direct winnerVersion: arxiv:2609.25804v2Metric: both-orders-correct, engineering/research macro averageScoring: Higher is betterGPT-5.6 Sol: Reasoning: xhigh, Output limit: 65,536 tokens, Attempts: 502 questions, two option orders each GPT-5.6 Luna: Reasoning: default, Output limit: 65,536 tokens, Attempts: 502 questions, two option orders eachSame prompt for all models. Each of 502 questions is shown twice with the options reversed. Average of engineering and research success rates, requiring both orders to be correct. 65,536-token output budget; unparseable answers count as wrong.
Higher measured value59.7 %
0 of 1,004 responses were unparseable and counted as wrong. Not a Gradually test.
49 %
0 of 1,004 responses were unparseable and counted as wrong. Not a Gradually test.
No documented difference in the test setupVersion: 2Metric: exact task accuracyScoring: Higher is betterSettings: Reasoning: low, Attempts: two exact output guesses per taskThe same verified test with five reasoning levels per model.
Test setup not documented as identicalNo direct winnerVersion: 1,266 tasksMetric: accuracyScoring: Higher is betterGPT-5.6 Sol: Reasoning: ultraCross-vendor comparison table with browsing tools. Exact agent systems differ.
Higher measured value92.2 %
Editorial selection from the vendor table, not a complete extract of the comparison cohort.
83.3 %
Editorial selection from the vendor table, not a complete extract of the comparison cohort.
No documented difference in the test setupVersion: 1Metric: criteria pass rateScoring: Higher is betterSettings: Reasoning: max, Output limit: 128,000 tokens
Legal Research BenchAdministrative / Regulatory · All-passTest details
No documented difference in the test setupVersion: 1Metric: all-pass rateScoring: Higher is betterSettings: Reasoning: max, Output limit: 128,000 tokens
LMArena Text, Style ControlBusiness, management and financeTest details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.6 Sol: Snapshot: gpt-5.6-sol-xhigh, Reasoning: xhigh, Harness: Arena text, style control GPT-5.6 Luna: Snapshot: gpt-5.6-luna-xhigh, Reasoning: xhigh, Harness: Arena text, style controlGPT-5.6 Sol: Style-controlled Arena rating from 6,528 anonymous pairwise votes.GPT-5.6 Luna: Style-controlled Arena rating from 6,871 anonymous pairwise votes.
Winner1,485.385
95% confidence interval: ±8.098The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,463.458
95% confidence interval: ±7.984The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
No documented difference in the test setupVersion: 1Metric: accuracyScoring: Higher is betterSettings: Reasoning: max, Harness: OpenHands, Output limit: 128,000 tokens
No documented difference in the test setupVersion: 1.2Metric: benchmark scoreScoring: Higher is betterSettings: Reasoning: max, Output limit: 128,000 tokens
No documented difference in the test setupVersion: 1.2Metric: benchmark scoreScoring: Higher is betterSettings: Reasoning: max, Output limit: 128,000 tokens
No documented difference in the test setupVersion: 2Metric: exact task accuracyScoring: Higher is betterSettings: Reasoning: max, Attempts: two exact output guesses per taskThe same verified test with five reasoning levels per model.
No documented difference in the test setupVersion: ARC-AGI-3 2026 verifiedMetric: RHAE scoreScoring: Higher is betterSettings: Reasoning: max, Agent: ARC Prize official harness, Harness: ARC Prize official harnessOfficial ARC Prize harness on unseen interactive environments. Reasoning levels remain separate.
Key to the marks:better valuebetter, but not conclusivetie
Our model tests
Our tests
Compare how the models respond to the same prompt. Each test shows the first attempt, with no subsequent fixes to the generated code. These results do not contribute to an overall score.
Pelican on a bicycle
A community classic for free-form SVG drawing.
GPT-5.6 Sol
API access was unavailable for this run. The model was not evaluated.
Reasoning (requested): high
First attempt
Run date
7 Sept 2026
API provider
openrouter
Requested API model
openai/gpt-5.6-sol
Reasoning setting
high
Pinned endpoint
OpenAI | openai/gpt-5.6-sol-20260709
Temperature
Provider default
Token limit including reasoning
8,192 tokens
Test protocol
community-visual-tests-v1
Technical checks
Not run
Not yet visually reviewed. A successful recording does not confirm correct physics or full compliance with the prompt.
GPT-5.6 Luna
A community classic for free-form SVG drawing.
Reasoning (requested): high
First attempt
Run date
7 Sept 2026
API provider
openrouter
Requested API model
openai/gpt-5.6-luna
API model reported in response
openai/gpt-5.6-luna
Reasoning setting
high
Pinned endpoint
OpenAI | openai/gpt-5.6-luna-20260709
Temperature
Provider default
Token limit including reasoning
8,192 tokens
Test protocol
community-visual-tests-v1
Technical checks
Passed, not a quality rating
Response time
49.9 s
Visually reviewed. This individual test does not establish overall model quality.
Your workload, budget, and required context length matter most. A coding benchmark says little about visual understanding or agent performance.
The standard price table uses direct developer rates. The provider comparison labels OpenRouter endpoints, regions, and special tariffs separately.
Every result links to its measurement source and retrieval date. Labels distinguish vendor reports, official benchmark leaderboards, and independent evaluations. Disputed or archived results are excluded.
More matchups
Related LLM comparisons
Compare GPT-5.6 Sol and GPT-5.6 Luna with other leading models using the same data and benchmark logic.
