Compare published benchmark results from matching versions and cohorts, API pricing, and technical specifications. Any documented test setup differences remain visible.
shared result series
179
distinct benchmarks
22
latest retrieval
2026-09-04
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-5.4
Coding: 99.7 / 68.2 (6 benchmark families)
Reasoning: 100 / 96.3 (1 benchmark family)
Knowledge: 100 / 98.2 (1 benchmark family)
Finance: 99 / 99.6 (3 benchmark families)
Legal: 100 / 49.5 (2 benchmark families)
Agentic tasks: 100 / 95.6 (1 benchmark family)
Multimodal: 100 / 98.5 (1 benchmark family)
Cybersecurity: 100 / 83.2 (1 benchmark family)
21 matched rows from 19 benchmark families were considered. The radar shows 8 of 10 comparable categories.
GPT-5.6 SolGPT-5.4
Code MigrationOverall. Higher is better. No documented setup difference
GPT-5.6 SolWinner52.916 %
GPT-5.434.984 %
CorpFin v2Overall. Higher is better. No documented setup difference
GPT-5.6 Sol64.375 %
GPT-5.4Winner65.268 %
CyberBenchOverall. Higher is better. No documented setup difference
GPT-5.6 SolWinner88.136 %
GPT-5.472.881 %
GPQA DiamondOverall. Higher is better. No documented setup difference
GPT-5.6 SolWinner95.202 %
GPT-5.491.666 %
Harvey's Legal Agent BenchmarkOverall · Task fully resolved. Higher is better. No documented setup difference
GPT-5.6 SolWinner2.5 %
GPT-5.40 %
IOIOverall. Higher is better. No documented setup difference
GPT-5.6 SolWinner86.667 %
GPT-5.467.834 %
LegalBenchOverall. Higher is better. No documented setup difference
GPT-5.6 SolWinner86.965 %
GPT-5.486.044 %
LiveCodeBench v6Overall. Higher is better. No documented setup difference
GPT-5.6 Sol82.604 %
GPT-5.4Winner84.141 %
LMArena Text, Style ControlOverall. Higher is better. Test setup not documented as identical
GPT-5.6 SolHigher measured value1,482.868
GPT-5.41,465.527
MedCodeOverall. Higher is better. No documented setup difference
GPT-5.6 SolWinner43.974 %
GPT-5.441.292 %
MedScribeOverall. Higher is better. No documented setup difference
GPT-5.6 SolWinner85.233 %
GPT-5.477.549 %
MMLU-ProOverall. Higher is better. No documented setup difference
GPT-5.6 SolWinner89.1 %
GPT-5.487.482 %
Key to the marks:better valuebetter, but not conclusivetie
Benchmark scores for GPT-5.6 Sol and GPT-5.4
Benchmark and task
GPT-5.6 Sol
GPT-5.4
Source
Code MigrationOverallTest details
No documented difference in the test setupVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05
Test setup not documented as identicalNo direct winnerVersion: 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-5.4: Snapshot: gpt-5.4, Harness: Arena text, style controlGPT-5.6 Sol: Style-controlled Arena rating from 23,153 anonymous pairwise votes.GPT-5.4: Style-controlled Arena rating from 63,615 anonymous pairwise votes.
Higher measured value1,482.868
95% confidence interval: ±5.12The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,465.527
95% confidence interval: ±3.807The 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: full finance setMetric: weighted rubric scoreScoring: Higher is betterGPT-5.6 Sol: Snapshot: gpt-5.6-sol-max, Reasoning: max, Attempts: pass@1 GPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Attempts: pass@1600 finance tasks with task-specific rubrics and o4-mini as judge.
Test setup not documented as identicalNo direct winnerVersion: full legal setMetric: weighted rubric scoreScoring: Higher is betterGPT-5.6 Sol: Snapshot: gpt-5.6-sol-max, Reasoning: max, Attempts: pass@1 GPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Attempts: pass@1500 legal tasks with task-specific rubrics and o4-mini as judge.
No documented difference in the test setupVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.6 Sol: Harness: OpenHands GPT-5.4: Snapshot: gpt-5.4-2026-03-05, Harness: OpenHands
No documented difference in the test setupVersion: 1.1Metric: accuracyScoring: Higher is betterGPT-5.6 Sol: Harness: OpenHands GPT-5.4: Snapshot: gpt-5.4-2026-03-05, Harness: OpenHands
Test setup not documented as identicalNo direct winnerVersion: 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-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style controlGPT-5.6 Sol: Style-controlled Arena rating from 23,153 anonymous pairwise votes.GPT-5.4: Style-controlled Arena rating from 60,657 anonymous pairwise votes.
Higher measured value1,482.868
95% confidence interval: ±5.12The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,476.488
95% confidence interval: ±3.881The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
LMArena Text, Style ControlBusiness, management and financeTest details
Test setup not documented as identicalNo direct winnerVersion: 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-5.4: Snapshot: gpt-5.4, Harness: Arena text, style controlGPT-5.6 Sol: Style-controlled Arena rating from 4,509 anonymous pairwise votes.GPT-5.4: Style-controlled Arena rating from 13,060 anonymous pairwise votes.
Higher measured value1,490.042
95% confidence interval: ±9.485The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,474.871
95% confidence interval: ±6.534The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
LMArena Text, Style ControlBusiness, management and financeTest details
Test setup not documented as identicalNo direct winnerVersion: 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-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style controlGPT-5.6 Sol: Style-controlled Arena rating from 4,509 anonymous pairwise votes.GPT-5.4: Style-controlled Arena rating from 12,346 anonymous pairwise votes.
Higher measured value1,490.042
95% confidence interval: ±9.485The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,483.779
95% confidence interval: ±6.654The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
Key to the marks:better valuebetter, but not conclusivetie
Our model tests
Community tests you can inspect
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.
First attempt
Run date
7 Sept 2026
API provider
openrouter
Requested API model
openai/gpt-5.6-sol
Pinned endpoint
OpenAI | openai/gpt-5.6-sol-20260709
Reasoning setting
high
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-5.4
API access was unavailable for this run. The model was not evaluated.
First attempt
Run date
7 Sept 2026
API provider
openrouter
Requested API model
openai/gpt-5.4
Pinned endpoint
OpenAI | openai/gpt-5.4-20260305
Reasoning setting
high
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.
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-5.4 with other leading models using the same data and benchmark logic.