Head-to-head comparison

Claude Opus 4.6 vs. GPT-5.4

Compare published benchmark results from matching versions and cohorts, API pricing, and technical specifications. Any documented test setup differences remain visible.

shared result series: 15
distinct benchmarks: 8
latest retrieval: 2026-09-02

First modelSecond model

Specifications

General specifications

Release date, availability, context window, architecture, and published parameter counts at a glance. Undisclosed details remain marked as unavailable.

Specifications for Claude Opus 4.6 and GPT-5.4
Attribute	Claude Opus 4.6	GPT-5.4
Provider	Anthropic	OpenAI
Released	Feb 2026	Mar 5, 2026
Availability	Active	Active
Model type	Proprietary	Proprietary
Parameters	Not published	Not published
Architecture	Not published	Not published
Context window	1,000,000 tokens	1,050,000 tokens
Knowledge cutoff	May 2025	August 31, 2025
Technical sources	Model, Context	Model

Pricing

API pricing compared

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.

API pricing for Claude Opus 4.6 and GPT-5.4
Price type	Claude Opus 4.6	GPT-5.4
API input	$5	$2.5 ≤272K / $5 >272K
API output	$25	$15 ≤272K / $22.5 >272K
Cache read	$0.5	$0.25 ≤272K / $0.5 >272K
Cache write (5 min.)	$6.25	Not listed
Cache write (1 hr.)	$10	Not listed
Price verified	07/30/2026 Source	07/30/2026 Source

Claude Opus 4.6GPT-5.4

API inputUSD per 1M tokens, base tier

Claude Opus 4.6$5

GPT-5.4$2.5

API outputUSD per 1M tokens, base tier

Claude Opus 4.6$25

GPT-5.4$15

Cache readUSD per 1M tokens, base tier

Claude Opus 4.6$0.5

GPT-5.4$0.25

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.

Provider pricing for Claude Opus 4.6 and GPT-5.4
Model	Provider	Purchase route and tariff	Input per 1M tokens	Output per 1M tokens	Source
Claude Opus 4.6	Anthropic	Direct from developerStandard API	Lowest listed price: $5	Lowest listed price: $25	Source07/30/2026
	Amazon Bedrock	Via OpenRouteramazon-bedrock	Lowest listed price: $5	Lowest listed price: $25	Source09/12/2026
	Anthropic	Via OpenRouteranthropic	Lowest listed price: $5	Lowest listed price: $25	Source09/12/2026
	Azure	Via OpenRouterazure/global	Lowest listed price: $5	Lowest listed price: $25	Source09/12/2026
	Claude Platform on AWS	Via OpenRouterclaude-on-aws	Lowest listed price: $5	Lowest listed price: $25	Source09/12/2026
	Google	Via OpenRoutergoogle-vertex/europe	$5.5	$27.5	Source09/12/2026
	Google	Via OpenRoutergoogle-vertex/global	Lowest listed price: $5	Lowest listed price: $25	Source09/12/2026
GPT-5.4	OpenAI	Direct from developerUp to 272,000 context tokens	$2.5	$15	Source07/30/2026
	OpenAI	Direct from developerAbove 272,000 context tokens	$5	$22.5	Source07/30/2026
	Amazon Bedrock	Via OpenRouteramazon-bedrock/us-east-1	$2.75	$16.5	Source09/12/2026
	Azure	Via OpenRouterazure	$2.5	$15	Source09/12/2026
	Azure	Via OpenRouterazure/eu	$2.75	$16.5	Source09/12/2026
	Azure	Via OpenRouterazure/us	$2.75	$16.5	Source09/12/2026
	OpenAI	Via OpenRouteropenai	$2.5	$15	Source09/12/2026
	OpenAI	Via OpenRouteropenai/fast	$5	$30	Source09/12/2026
	OpenAI	Via OpenRouteropenai/flex	Lowest listed price: $1.25	Lowest listed price: $7.5	Source09/12/2026

Performance

Shared benchmarks

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.

Not enough directly matched categories for a reliable radar

Only 3 of at least 5 categories have comparable, direct model measurements. The benchmark tables below show the available evidence without filling missing axes.

Claude Opus 4.6GPT-5.4

Humanity's Last ExamOverall. Higher is better. No documented setup difference

Claude Opus 4.634.44% ± 1.86

GPT-5.436.24% ± 1.88

METR Time Horizon50% time horizon. Higher is better. No documented setup difference

Claude Opus 4.612 h (95% confidence interval 5.3 h to 60.6 h)

GPT-5.45.7 h (95% confidence interval 3.1 h to 12.8 h)

MultiNRCOverall. Higher is better. Test setup not documented as identical

Claude Opus 4.648.34% ± 3.02

GPT-5.458.29% ± 2.98

PRBench FinanceOverall. Higher is better. Test setup not documented as identical

Claude Opus 4.653.28% ± 0.18

GPT-5.445.63% ± 0.56

PRBench LegalOverall. Higher is better. Test setup not documented as identical

Claude Opus 4.652.27% ± 0.66

GPT-5.444.35% ± 0.28

SecCodeBench-V2Generation with security hints. Higher is better. No documented setup difference

Claude Opus 4.670.65 %

GPT-5.467.82 %

VISTAOverall. Higher is better. Test setup not documented as identical

Claude Opus 4.645.48% ± 0.44

GPT-5.450.89% ± 1.76

VisualToolBenchOverall. Higher is better. Test setup not documented as identical

Claude Opus 4.627.52% ± 1.45

GPT-5.429.17% ± 0.13

METR Time Horizon80% time horizon. Higher is better. No documented setup difference

Claude Opus 4.61.2 h (95% confidence interval 27 min to 2.8 h)

GPT-5.453.9 min (95% confidence interval 24 min to 1.8 h)

MultiNRCOverall. Higher is better. Test setup not documented as identical

Claude Opus 4.657.06% ± 2.99

GPT-5.458.29% ± 2.98

SecCodeBench-V2Generation without hints. Higher is better. No documented setup difference

Claude Opus 4.651.05 %

GPT-5.450.97 %

VISTAOverall. Higher is better. Test setup not documented as identical

Claude Opus 4.646.07% ± 0.26

GPT-5.450.89% ± 1.76

Key to the marks:better valuebetter, but not conclusivetie

Benchmark scores for Claude Opus 4.6 and GPT-5.4
Benchmark and task	Claude Opus 4.6	GPT-5.4	Source
Humanity's Last ExamOverall Test details No documented difference in the test setupNo clear leadVersion: final 2,500-question public setMetric: accuracyScoring: Higher is betterClaude Opus 4.6: Snapshot: claude-opus-4-6-thinking-max, Attempts: pass@1 GPT-5.4: Snapshot: gpt-5.4-2026-03-05, Attempts: pass@1All 2,500 public questions, temperature 0, and an automatic answer extractor.	34.44% ± 1.86 95% confidence interval: ±1.86	No clear lead36.24% ± 1.88 95% confidence interval: ±1.88	Scale AI Source details Data as of: 09/02/2026
METR Time Horizon50% time horizon Test details No documented difference in the test setupVersion: TH 1.1Metric: 50% time horizonScoring: Higher is betterClaude Opus 4.6: Snapshot: claude-opus-4-6-2026-02-05, Agent: ReAct, Harness: ReAct, Attempts: 6 independent runs per task GPT-5.4: Snapshot: gpt-5.4-2026-03-05, Agent: ReAct, Harness: ReAct, Attempts: 6 independent runs per taskReAct agent with six independent runs per task. Time is estimated human completion duration.	Winner12 h (95% confidence interval 5.3 h to 60.6 h)	5.7 h (95% confidence interval 3.1 h to 12.8 h)	METR Source details Data as of: 05/08/2026
MultiNRCOverall Test details Test setup not documented as identicalNo direct winnerVersion: French, Spanish, and Chinese releaseMetric: accuracyScoring: Higher is betterClaude Opus 4.6: Reasoning: none, Attempts: pass@1 GPT-5.4: Snapshot: gpt-5.4-xhigh, Reasoning: xhigh, Attempts: pass@1Native, non-translated reasoning tasks with automatic evaluation.	48.34% ± 3.02 95% confidence interval: ±3.02	Higher measured value58.29% ± 2.98 95% confidence interval: ±2.98	Scale AI Source details Data as of: 08/26/2026
PRBench FinanceOverall Test details Test setup not documented as identicalNo direct winnerVersion: full finance setMetric: weighted rubric scoreScoring: Higher is betterClaude Opus 4.6: Reasoning: none, 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.	Higher measured value53.28% ± 0.18 95% confidence interval: ±0.18	45.63% ± 0.56 95% confidence interval: ±0.56	Scale AI Source details Data as of: 08/26/2026
PRBench LegalOverall Test details Test setup not documented as identicalNo direct winnerVersion: full legal setMetric: weighted rubric scoreScoring: Higher is betterClaude Opus 4.6: Reasoning: none, 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.	Higher measured value52.27% ± 0.66 95% confidence interval: ±0.66	44.35% ± 0.28 95% confidence interval: ±0.28	Scale AI Source details Data as of: 08/26/2026
SecCodeBench-V2Generation with security hints Test details No documented difference in the test setupVersion: 2.2.0Metric: security-aware generation pass@1Scoring: Higher is better98 project cases. Functional tests run before dynamic exploit checks and security scoring.	Winner70.65 %	67.82 %	SecCodeBench Source details Data as of: 02/16/2026
VISTAOverall Test details Test setup not documented as identicalNo direct winnerVersion: 758 private prompt-image pairsMetric: rubric scoreScoring: Higher is betterClaude Opus 4.6: Reasoning: none, Attempts: pass@1 GPT-5.4: Snapshot: gpt-5.4-xhigh, Reasoning: xhigh, Attempts: pass@1758 single-turn image-text tasks with structured yes-no rubrics.	45.48% ± 0.44 95% confidence interval: ±0.44	Higher measured value50.89% ± 1.76 95% confidence interval: ±1.76	Scale AI Source details Data as of: 08/26/2026
VisualToolBenchOverall Test details Test setup not documented as identicalNo direct winnerVersion: 1,204-task releaseMetric: average pass rateScoring: Higher is betterClaude Opus 4.6: Reasoning: thinking, Attempts: pass@1 GPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Attempts: pass@11,204 tasks, six tools, and at most 20 tool calls per task.	27.52% ± 1.45 95% confidence interval: ±1.45	Higher measured value29.17% ± 0.13 95% confidence interval: ±0.13	Scale AI Source details Data as of: 08/26/2026
METR Time Horizon80% time horizon Test details No documented difference in the test setupVersion: TH 1.1Metric: 80% time horizonScoring: Higher is betterClaude Opus 4.6: Snapshot: claude-opus-4-6-2026-02-05, Agent: ReAct, Harness: ReAct, Attempts: 6 independent runs per task GPT-5.4: Snapshot: gpt-5.4-2026-03-05, Agent: ReAct, Harness: ReAct, Attempts: 6 independent runs per taskReAct agent with six independent runs per task. Time is estimated human completion duration.	Winner1.2 h (95% confidence interval 27 min to 2.8 h)	53.9 min (95% confidence interval 24 min to 1.8 h)	METR Source details Data as of: 05/08/2026
MultiNRCOverall Test details Test setup not documented as identicalNo direct winnerVersion: French, Spanish, and Chinese releaseMetric: accuracyScoring: Higher is betterClaude Opus 4.6: Reasoning: max, Attempts: pass@1 GPT-5.4: Snapshot: gpt-5.4-xhigh, Reasoning: xhigh, Attempts: pass@1Native, non-translated reasoning tasks with automatic evaluation.	57.06% ± 2.99 95% confidence interval: ±2.99	Higher measured value58.29% ± 2.98 95% confidence interval: ±2.98	Scale AI Source details Data as of: 08/26/2026
SecCodeBench-V2Generation without hints Test details No documented difference in the test setupVersion: 2.2.0Metric: native generation pass@1Scoring: Higher is better98 project cases. Functional tests run before dynamic exploit checks and security scoring.	Winner51.05 %	50.97 %	SecCodeBench Source details Data as of: 02/16/2026
VISTAOverall Test details Test setup not documented as identicalNo direct winnerVersion: 758 private prompt-image pairsMetric: rubric scoreScoring: Higher is betterClaude Opus 4.6: Reasoning: max, Attempts: pass@1 GPT-5.4: Snapshot: gpt-5.4-xhigh, Reasoning: xhigh, Attempts: pass@1758 single-turn image-text tasks with structured yes-no rubrics.	46.07% ± 0.26 95% confidence interval: ±0.26	Higher measured value50.89% ± 1.76 95% confidence interval: ±1.76	Scale AI Source details Data as of: 08/26/2026
SecCodeBench-V2Repair with security hints Test details No documented difference in the test setupVersion: 2.2.0Metric: security-aware repair pass@1Scoring: Higher is better98 project cases. Functional tests run before dynamic exploit checks and security scoring.	Winner81.73 %	80.04 %	SecCodeBench Source details Data as of: 02/16/2026
SecCodeBench-V2Repair without hints Test details No documented difference in the test setupVersion: 2.2.0Metric: native repair pass@1Scoring: Higher is better98 project cases. Functional tests run before dynamic exploit checks and security scoring.	Winner73.1 %	61.41 %	SecCodeBench Source details Data as of: 02/16/2026
SecCodeBench-V2Weighted total Test details No documented difference in the test setupVersion: 2.2.0Metric: weighted pass@1Scoring: Higher is better98 project cases. Functional tests run before dynamic exploit checks and security scoring.	Winner64.9 %	59.74 %	SecCodeBench Source details Data as of: 02/16/2026
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.

Claude Opus 4.6

Not tested yet

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.

Prompt and test conditions

Generate an SVG of a pelican riding a bicycle

Token limit including reasoning: 8,192 tokens

Task origin (Simon Willison)

How to read this comparison

Winning one benchmark is not an overall verdict

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

Compare Claude Opus 4.6 and GPT-5.4 with other leading models using the same data and benchmark logic.

View the complete LLM comparison