Across 24 comparable benchmarks, GPT-5.4 has better reported results on 21 and worse on 2, with 1 tied. 1 further shared result is listed below, each with the reason it is not counted.
Strong evidence24 of 25 shared results countedHow we compare
Price basis: Listed rates vs standard rates. The price section below shows the applicable periods.
| Fact | GPT-5.4 | GPT-5.4 mini |
|---|---|---|
| Context window | 1M tokens | 400K tokens |
| Input / 1M | $2.50 | $0.750 |
| Output / 1M | $15.00 | $4.50 |
| Blended (4:1) | $3.38 | $1.01 |
| Status | Available | Available |
| Released | March 5, 2026 | March 17, 2026 |
Blended is 4:1 input to output with 90% cache hits at listed rates — set your own mix.
Mean of 18 shared benchmarks grouped by capability on a 0–100 scale. How averages are built
GPT-5.4 ahead on 5 of 5 counted
GPT-5.4 ahead on 1 of 1 counted
GPT-5.4 ahead on 3 of 3 counted
GPT-5.4 ahead on 2 of 2 counted
GPT-5.4 ahead on 3 of 4 counted
GPT-5.4 ahead on 1 of 1 counted
GPT-5.4 ahead on 1 of 1 counted
GPT-5.4 ahead on 2 of 4 counted
No shared comparable results for Research & search, Cybersecurity.
Published scores, with matching modes where available. Gaps show GPT-5.4 relative to GPT-5.4 mini. Reading the results
| Benchmark | GPT-5.4 | GPT-5.4 mini | Gap |
|---|---|---|---|
| Code MigrationMaintenance & Optimization · Hidden-test pass rate (%) | 35.0Extra-High · tools | 12.9Extra-High · tools | +22.0 pts#38 vs #62 of 79 |
| SWE-Bench Pro - PublicRepository Engineering · Accuracy | 59.1Extra-High · tools | 54.4Extra-High · tools | +4.7 pts#19 vs #36 of 66 |
| Terminal Bench 2.0Agentic Development · Accuracy | 75.1Extra-High · tools | 60.0Extra-High · tools | +15.1 pts#4 vs #17 of 46 |
| Terminal Bench HardAgentic Development · Accuracy | 57.6Extra-High · tools | 52.3Extra-High · tools | +5.3 pts#6 vs #11 of 157 |
| Vibe Code Bench v1.1Code Generation & Editing · Pass rate (%) | 48.5Extra-High · tools | 48.0Extra-High · tools | +0.5 pts#66 vs #67 of 121 |
| ARC-AGI-1Abstract Generalization · Score (%) | 93.7Normal · no tools | 63.7≠Extra-High · tools | +30.0 pts#26 vs #56 of 87 |
| ARC-AGI-2Abstract Generalization · Score (% solved); cost per task (USD)Not counted · Speed and cost telemetry is reported, not scored | 74.0Extra-High · no tools | 18.9≠Extra-High · tools | 3.91×#25 vs #53 of 80 |
| FrontierMath - Tier 4Mathematics · Accuracy | 27.1Extra-High · no tools | 2.1≠Thinking High · no tools | +25.0 pts#5 vs #31 of 50 |
| FrontierMath Tier 4 v2Mathematics · Accuracy (verification_code) | 49.0Extra-High · tools | 9.8Extra-High · tools | +39.2 pts#23 vs #62 of 73 |
| FrontierMath v2Mathematics · Accuracy (verification_code) | 78.6Extra-High · tools | 51.2Extra-High · tools | +27.4 pts#22 vs #55 of 85 |
| CritPtScientific Reasoning · Score | 23.4Extra-High · no tools | 10.0Extra-High · no tools | 2.34×#23 vs #55 of 134 |
| GPQA DiamondScientific Reasoning · Accuracy | 89.9Thinking High · no tools | 64.1≠Normal · no tools | +25.8 pts#53 vs #160 of 200 |
| Claw BenchTool Orchestration · Accuracy | 92.7Thinking · tools | 75.3Thinking · tools | +17.4 pts#3 vs #25 of 29 |
| MCP-AtlasTool Orchestration · Pass rate / claim coverage | 70.6Extra-High · tools | 56.7Extra-High · tools | +13.9 pts#35 vs #49 of 55 |
| OSWorld-VerifiedDesktop Workflows · Accuracy | 75.0Extra-High · tools | 72.1Extra-High · tools | +2.9 pts#18 vs #24 of 36 |
| PinchBench v2Tool Orchestration · Average score (%) | 75.7Reported best | 79.2Reported best | −3.5 pts#21 vs #17 of 51 |
| Context ArenaMemory & Persistence · Accuracy (8 needles, 4K-128K context) | 86.2Extra-High · no tools | 50.9Extra-High · no tools | +35.2 pts#13 vs #50 of 61 |
| MMMU-ProVisual Understanding · Accuracy | 82.1Thinking · tools | 73.3≠Extra-High · no tools | +8.8 pts#21 vs #77 of 136 |
| Creative WritingWriting | 1841Normal · no tools | 1665Normal · no tools | 1.11×#20 vs #40 of 121 |
| Harvey's Legal Agent BenchmarkLegal · Accuracy (%) | 0.0Extra-High · tools | 0.0Extra-High · tools | =#44 vs #44 of 52 |
| SAGEService Workflows · Accuracy (%) | 43.3Extra-High · tools | 50.8Extra-High · tools | −7.5 pts#49 vs #15 of 82 |
| τ³-BankingService Workflows · Score | 39.4Extra-High · tools | 25.6Extra-High · tools | 1.54×#26 vs #47 of 120 |
| HLEKnowledge Exams · Accuracy | 52.1Extra-High · tools | 41.5Extra-High · tools | +10.6 pts#41 vs #76 of 151 |
| ECICapability Indices · ECI score (capability index, higher is better) | 157Reported best | 149Reported best | 1.05×#16 vs #55 of 168 |
| LiveBenchCross-capability Suites · Accuracy | 78.0Deep Thinking · no tools | 66.4Deep Thinking · no tools | +11.6 pts#18 vs #69 of 106 |
Maintenance & Optimization · Hidden-test pass rate (%)
Repository Engineering · Accuracy
Agentic Development · Accuracy
Agentic Development · Accuracy
Code Generation & Editing · Pass rate (%)
Mathematics · Accuracy
Mathematics · Accuracy (verification_code)
Mathematics · Accuracy (verification_code)
Scientific Reasoning · Score
Scientific Reasoning · Accuracy
Tool Orchestration · Accuracy
Tool Orchestration · Pass rate / claim coverage
Desktop Workflows · Accuracy
Tool Orchestration · Average score (%)
Memory & Persistence · Accuracy (8 needles, 4K-128K context)
Visual Understanding · Accuracy
Writing
Legal · Accuracy (%)
Service Workflows · Accuracy (%)
Service Workflows · Score
A dash means no result is available.
| Benchmark | GPT-5.4 | GPT-5.4 mini |
|---|---|---|
| AIME 2026 | 99.2 | — |
| FrontierMath | 47.6 | — |
| HMMT Feb 2026 | 97.7 | — |
| MathArena Apex | 54.2 | — |
| Riemann-bench | 41.6 | — |
| SciCode | — | 52.1 |
| APEX-SWE | 36.5 | — |
| DeepSWE | 51.8 | — |
| FrontierCode 1.1 | — | 43.0 |
| FrontierSWE v1 | 46.0 | — |
| GSO | 31.4 | — |
| PostTrain Bench | 19.0 | — |
| SWE-Atlas-QnA | 40.8 | — |
| SWE-Bench Pro - Commercial | 43.4 | — |
| SWE-sweep | — | 0.49 |
| Terminal-Bench 4.0 | — | 2.5 |
| Terminal-Bench 4.1 | — | 1.8 |
| WeirdML v2 | 77.7 | — |
| Agents' Last Exam v1.0 | 20.5 | — |
| APEX-Agents | 52.4 | — |
| Pinch Bench | 90.5 | — |
| Tool Decathlon | — | 42.9 |
| METR Time Horizons v1.1 | 342 | — |
| Text Arena (Coding) | 1457 | — |
| Vals Index | — | 39.6 |
| Vals Index v2.1 | — | 33.2 |
| EBR-bench | 25.4 | — |
| GDP.pdf | — | 13.4 |
| AA-AnalystAgent | — | 15.0 |
| AutomationBench | 11.6 | — |
| EMB (Excel Modeling) | — | 45.4 |
| Finance Agent v2 | — | 45.4 |
| Harvey Lab-AA | — | 60.8 |
| Legal Research Bench | — | 12.5 |
| τ²-Bench - Telecom | 64.3 | — |
| ARC-AGI-3 | 0.21 | — |
| Vending-Bench 2 | 6144 | — |
| BrowseComp | 82.7 | — |
| MedCode | 41.3 | — |
| MedScribe | 77.5 | — |
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%
GPT-5.4 costs233%morethan GPT-5.4 mini per million tokens — 3.3× the price.Using listed rates vs standard rates at this token mix.
GPT-5.4: 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; start date not published, verification date used
GPT-5.4 mini: Standard pricing. OpenAI official pricing (developers.openai.com/api/docs/pricing) verified 2026-10-21; standard tier
| Price / 1M tokens | GPT-5.4 | GPT-5.4 mini | Ratio |
|---|---|---|---|
| Input | $2.50 | $0.75 | 3.33× |
| Output | $15.00 | $4.50 | 3.33× |
| Cached input | $0.25 | $0.075 | 3.33× |
| Cache write | — | — | — |
| Blended | $3.38 | $1.01 | 3.33× |
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.
GPT-5.4 relative to GPT-5.4 mini
Summary: 2 of 2 shared benchmarks have matching recorded settings and sources. The average is a geometric mean of 2 positive-cost ratios, one equal weight per benchmark. This benchmark set can include related tests and indexes; it is not a workload cost estimate.
Effort, protocol, aggregation and source match; other run conditions may differ.
GPT-5.4
GPT-5.4 mini
Effort, protocol, aggregation and source match; other run conditions may differ.
GPT-5.4
GPT-5.4 mini
| Benchmark | GPT-5.4 | GPT-5.4 mini |
|---|---|---|
| Agents' Last Exam v1.0 | Not reported | |
| AIME 2026 | Not reported | |
| AutomationBench | $0.07 / task Average · Pass Rate: 0.15 Reported best (effort unspecified) · zapier-private-heldout $0.62 / task Average · Pass Rate: 10.1 Thinking High (With Tools) · zapier-private-heldout 2 more settings$0.36 / task Average · Pass Rate: 5.3 Thinking Medium (With Tools) · zapier-private-heldout | Not reported |
| DeepSWE | Not reported | |
| FrontierCode 1.1 | Not reported | $0.8043 / task Average · Weighted rubric score: 37.2 Thinking High (With Tools) · frontiercode-1-1-extended $0.9537 / task Average · Weighted rubric score: 23.1 Thinking High (With Tools) · frontiercode-1-1-main $0.3288 / task Average · Weighted rubric score: 20.6 Thinking Low (With Tools) · frontiercode-1-1-extended 5 more settings$0.3642 / task Average · Weighted rubric score: 9.3 Thinking Low (With Tools) · frontiercode-1-1-main $0.6716 / task Average · Weighted rubric score: 35.1 Thinking Medium (With Tools) · frontiercode-1-1-extended $0.7833 / task Average · Weighted rubric score: 20.8 Thinking Medium (With Tools) · frontiercode-1-1-main $1.5201 / task Average · Weighted rubric score: 27.0 Extra-High (With Tools) · frontiercode-1-1-main $1.2621 / task Average · Weighted rubric score: 43.0 Extra-High (With Tools) · frontiercode-1-1-extended |
| SWE-sweep | Not reported | |
| Terminal-Bench 4.0 | Not reported | $1.44 / task Average · Resolution rate (%): 2.5 Unspecified Reasoning (With Tools) · vals-mini-swe |
| Terminal-Bench 4.1 | Not reported | |
| Vals Index v2.1 | Not reported |
| Attribute | GPT-5.4 | GPT-5.4 mini |
|---|---|---|
| Availability | Available | Available |
| Provider | OpenAI | OpenAI |
| Model type | Multimodal model | Reasoning model |
| First release | March 5, 2026 | March 17, 2026 |
| Context window | 1M tokens | 400K tokens |
| Max output | 128K tokens | 128K tokens |
| Architecture | Undisclosed | Undisclosed |
| Documented capabilities | Reasoning, Multilingual | Reasoning, Multilingual |
| Knowledge cutoff | 2025-08 | 2025-08-31 |
| Input modalities | Text, Image | Text, Image |
| Output modalities | Text | Text |
| Weights | Closed / API only | Closed / API only |
| Commercial use | Proprietary | Proprietary |
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 GPT-5.4 or GPT-5.4 mini? 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.