GPT-5.4 vs GPT-5.6 Sol
Full profiles: GPT-5.4 (OpenAI) and GPT-5.6 Sol (OpenAI): every benchmark, price and rank.
Updated · Data from 11 sources
As of 25 Oct 2026, GPT-5.4 (xhigh) and GPT-5.6 Sol (max) are neck and neck on quality (64.7 vs 68.0 on the BenchLeader Index, inside the ±4.5 margin of error), and GPT-5.4 (xhigh) costs 1.4× less, which makes it the better buy for most workloads. Pick GPT-5.4 for instruction following (72 vs 71); pick GPT-5.6 Sol for maths (71 vs 65).
GPT-5.4 benchmarks Which to pick for…
The pick's numbers come first. Category scores on the BenchLeader scale (50 is the average model, 15 points one standard deviation). High confidence means a lead of half a standard deviation or more; under one point is called either.
| Use case | Pick | Why | Confidence |
|---|---|---|---|
| Coding | GPT-5.6 Sol | 66 vs 65 category score; SciCode 57.1% vs 56.6% | Medium |
| Agents & tool use | GPT-5.6 Sol | 67 vs 64 category score; τ³-Banking (AA) 44.3% vs 39.6% | Medium |
| Maths | GPT-5.6 Sol | 71 vs 65 category score; FrontierMath Tiers 1–3 89.1% vs 78.6% | Medium |
| Reasoning | GPT-5.6 Sol | 72 vs 69 category score; GPQA Diamond 93.5% vs 93.3% | Medium |
| Knowledge | GPT-5.6 Sol | 65 vs 60 category score; SimpleQA Verified 69.7% vs 45.1% | Medium |
| Long documents | GPT-5.6 Sol | 67 vs 66 category score; AA-LCR 84.0% vs 82.0% | Medium |
| Images and other media | GPT-5.6 Sol | 67 vs 61 category score; MMMU-Pro 83.4% vs 78.4% | Medium |
| Following instructions | GPT-5.4 | 72 vs 71 category score; IFBench 74.0% vs 72.7% | Medium |
| Lowest cost | GPT-5.4 | $5.63 vs $8.00 per 1M blended; agentic coding session $0.210 vs $0.320 | – |
| Fastest first answer | GPT-5.6 Sol | 69.5 s (GPT-5.6 Sol (max)) vs 118.0 s (GPT-5.4 (xhigh)) to the first answer token | – |
| High volume on a budget | Neither flagship | A smaller model is usually the better fit: Cheapest good LLMs, Fastest LLMs | – |
Key differences
- GPT-5.4 (xhigh) and GPT-5.6 Sol (max) are neck and neck on quality (64.7 vs 68.0, within the ±4.5 margin).
- GPT-5.4 (xhigh) is stronger in instruction following.
- GPT-5.6 Sol (max) is stronger in agents & tools, coding, composite, knowledge, long context, maths, multimodal, reasoning.
- GPT-5.4 (xhigh) is 1.4× cheaper ($5.63 vs $8.00 per 1M blended).
At a glance
| Spec | GPT-5.4 | GPT-5.6 Sol |
|---|---|---|
| Developer | OpenAI | OpenAI |
| Released | 5 Mar 2026 | 9 Jul 2026 |
| Context window (tokens) | 1,050,000 | 1,050,000 |
| Compared at | GPT-5.4 (xhigh) | GPT-5.6 Sol (max) |
| Open weights | No | No |
| Price in / out / cached input per 1M | $2.50 / $15.00 / $0.250 | $4.00 / $20.00 / $0.400 |
Category by category
Category scores on the BenchLeader scale: 50 is the average evaluated model, 15 points is one standard deviation. A gap under one point is called level.
| Category | GPT-5.4 (xhigh) | GPT-5.6 Sol (max) | Leader |
|---|---|---|---|
| Reasoning | 69.1 | 72.4 | GPT-5.6 Sol |
| Coding | 64.6 | 65.9 | GPT-5.6 Sol |
| Agents & tools | 63.6 | 66.5 | GPT-5.6 Sol |
| Maths | 65.2 | 71.0 | GPT-5.6 Sol |
| Knowledge | 59.7 | 65.3 | GPT-5.6 Sol |
| Instruction following | 72.3 | 71.1 | GPT-5.4 |
| Multimodal | 61.4 | 66.5 | GPT-5.6 Sol |
| Long context | 66.1 | 67.1 | GPT-5.6 Sol |
| Composite | 73.1 | 82.0 | GPT-5.6 Sol |
Where sources disagree
Across the 55 benchmarks both have, GPT-5.4 leads 6, GPT-5.6 Sol leads 49. Every overall index that rates both puts GPT-5.6 Sol ahead.
| Overall index | GPT-5.4 (xhigh) | GPT-5.6 Sol (max) | Leader |
|---|---|---|---|
| BenchLeader Index | 64.7 | 68.0 | neck and neck |
| AA Intelligence Index v4.3.2 | 39.0 | 47.0 | GPT-5.6 Sol |
| Arena Text | 1475† | 1485 | GPT-5.6 Sol |
Each publisher runs its own tasks, harnesses and effort settings, so their overall verdicts can differ; BenchLeader reconciles them on one scale. † measured at another effort setting of the same model (the publisher did not test the one compared here).
Price, speed and latency
List prices blended at three input tokens per output token. Per-task prices use the token counts on the methodology page; suite cost is what Artificial Analysis paid to run its whole Intelligence Index.
| Measure | GPT-5.4 (xhigh) | GPT-5.6 Sol (max) |
|---|---|---|
| Blended price per 1M tokens | $5.63 | $8.00 |
| Input / output per 1M | $2.50 / $15.00 | $4.00 / $20.00 |
| Cached input per 1M | $0.250 | $0.400 |
| Price per task: chat reply | $0.0055 | $0.0076 |
| Price per task: summarize a 30-page report | $0.039 | $0.060 |
| Price per task: code edit | $0.037 | $0.054 |
| Price per task: agentic coding session | $0.210 | $0.320 |
| Price per task: structured extraction | $0.0080 | $0.012 |
| Cost to run a full benchmark suite | – | $1.99 |
| Output speed | 96 tok/s | 89 tok/s |
| Time to first answer | 118.0 s | 69.5 s |
| Time to a complete answer | 123.2 s | 75.1 s |
Each model is compared at its best-ranked reasoning-effort setting. GPT-5.4 (xhigh) scores 64.7 and takes 118 s to its first answer token; GPT-5.4 (low) scores 59.7 in 2 s, GPT-5.4 (high) scores 58.9 in 1 s. GPT-5.6 Sol (max) scores 68.0 and takes 69 s to its first answer token; GPT-5.6 Sol (xhigh) scores 66.6 in 17 s, GPT-5.6 Sol (high) scores 66.3 in 7 s. Higher effort means more thinking before the answer: better scores, longer waits and more tokens billed. How reasoning effort changes scores and cost.
Every benchmark
165 benchmarks, grouped by category; each name links to the board and its leaders. † measured at another effort setting of the same model, shown for reference and not counted as a win.
| Metric | Publisher | GPT-5.4 (xhigh) | GPT-5.6 Sol (max) |
|---|---|---|---|
| BenchLeader Index | BenchLeader | 64.7 | 68.0 |
| Agents & tools score | BenchLeader | 63.6 | 66.5 |
| Coding score | BenchLeader | 64.6 | 65.9 |
| Composite score | BenchLeader | 73.1 | 82.0 |
| Instruction following score | BenchLeader | 72.3 | 71.1 |
| Knowledge score | BenchLeader | 59.7 | 65.3 |
| Long context score | BenchLeader | 66.1 | 67.1 |
| Maths score | BenchLeader | 65.2 | 71.0 |
| Multimodal score | BenchLeader | 61.4 | 66.5 |
| Reasoning score | BenchLeader | 69.1 | 72.4 |
| Blended price $/M | BenchLeader | $5.63 | $8.00 |
| Output speed | BenchLeader | 96 tok/s | 89 tok/s |
| Time to first answer | BenchLeader | 118.0 s | 69.5 s |
| Context window | BenchLeader | 1,050,000 tokens | 1,050,000 tokens |
| Reasoning | |||
| GPQA Diamond | Epoch AI Benchmarking Hub | 93.3% | 93.5% |
| Humanity's Last Exam | Scale AI / CAIS | 36.2% | – |
| SimpleBench | SimpleBench | – | 64.8%† |
| Arena Hard Prompts | Arena (LMArena) | 1487† | 1512† |
| LiveBench 2.0 Reasoning | LiveBench | – | 91.7%† |
| LiveBench 1.0 Reasoning | LiveBench | 88.1% | 91.7% |
| GPQA Diamond (AA) | Artificial Analysis | 92.0% | 94.1% |
| Humanity's Last Exam (AA) | Artificial Analysis | 43.7% | 49.5% |
| GPQA Diamond (Vals) | Vals AI | 91.7% | 95.2% |
| EnigmaEval | Scale AI SEAL | – | 37.1%† |
| Kagi LLM Benchmark | Kagi LLM Benchmark | 63.8%† | 67.0%† |
| ARC-AGI-1 | ARC Prize | 93.7% | 96.5% |
| ARC-AGI-2 | ARC Prize | 74.0% | 92.5% |
| ARC-AGI-3 | ARC Prize | 0.2%† | 7.8% |
| CritPt | Artificial Analysis | 23.4% | 32.3% |
| MysteryMechanism | Vals AI | – | 33.3% |
| Chess Puzzles | Epoch AI Benchmarking Hub | 48.0% | 59.0% |
| Mystery Game Puzzles | Epoch AI Benchmarking Hub | 37.0% | 58.0% |
| CL-bench | Epoch AI Benchmarking Hub | 27.9% | – |
| CL-bench Life | Epoch AI Benchmarking Hub | 21.7% | – |
| LMCA | Epoch AI Benchmarking Hub | 52.0% | 58.4% |
| DTBench | Epoch AI Benchmarking Hub | 94.4% | 95.5% |
| ForecastBench | ForecastBench | 63.1%† | 65.5% |
| BTF-3 | Epoch AI Benchmarking Hub | – | 13.7% |
| Coding | |||
| SWE-bench Verified (Epoch) | Epoch AI Benchmarking Hub | 76.9%† | – |
Show all 165 benchmarks
| Metric | Publisher | GPT-5.4 (xhigh) | GPT-5.6 Sol (max) |
|---|---|---|---|
| Coding | |||
| SciCode | SciCode | 56.6% | 57.1% |
| WeirdML | WeirdML | 77.7% | 87.0% |
| FrontierCode | Cognition | – | 47.5%† |
| GSO-Bench | GSO-Bench | 31.4% | 76.5%† |
| Arena Coding | Arena (LMArena) | 1511† | 1534† |
| Arena WebDev | Arena (LMArena) | 1401† | 1615† |
| LiveBench 2.0 Coding | LiveBench | – | 89.7%† |
| LiveBench 1.0 Coding | LiveBench | 77.5% | 83.9% |
| SciCode (AA) | Artificial Analysis | – | 57.1% |
| LiveCodeBench | Vals AI | 84.1% | 82.6% |
| IOI | Vals AI | – | 91.2% |
| SWE-bench (Vals) | Vals AI | 78.2% | 96.2% |
| SWE-Bench Pro | Scale AI SEAL | 59.1% | – |
| Code Migration | Vals AI | 35.0% | 52.9% |
| ProgramBench | Vals AI | 0.0% | 1.5% |
| Vibe Code Bench 1-100 | Vals AI | – | 20.0% |
| Vibe Code Bench v1.1 | Vals AI | 67.4% | 80.5% |
| SWE-Bench Pro (private) | Scale AI SEAL | 59.1% | – |
| SWE Atlas: Codebase QnA | Scale AI SEAL | 40.8% | 46.0%† |
| SWE Atlas: Refactoring | Scale AI SEAL | 44.3% | – |
| SWE Atlas: Test Writing | Scale AI SEAL | 44.4% | 45.9%† |
| MirrorCode | Epoch AI Benchmarking Hub | 15.6%† | 20.0%† |
| Surface Evolver Bench | Epoch AI Benchmarking Hub | – | 93.1%† |
| DeepSWE v1.1 | Epoch AI Benchmarking Hub | 51.8% | 72.7% |
| CursorBench | Cursor | – | 41.7% |
| ALE-Bench | Epoch AI Benchmarking Hub | 1607† | 2176.9 |
| AlgoTune | AlgoTune | 1.9† | – |
| FrontierSWE | Epoch AI Benchmarking Hub | – | 32.2% |
| Terminal-Bench 4.0 (AA) | Artificial Analysis | – | 39.9% |
| Terminal-Bench 2.1 (AA) | Artificial Analysis | 78.3% | 88.0% |
| Startrise Frontend | Startrise | – | 66.2%† |
| Agents & tools | |||
| Terminal-Bench | Terminal-Bench | 81.8%† | 37.3% |
| OSWorld-Verified 2.0 | OSWorld | – | 27.3% |
| APEX-Agents | Mercor | 52.4%† | – |
| Arena Agent | Arena (LMArena) | – | 6.1† |
| LiveBench Long Agentic | LiveBench | – | 48.5%† |
| LiveBench Browser | LiveBench | – | 88.0%† |
| LiveBench Agentic Coding | LiveBench | 53.8% | 56.2% |
| Terminal-Bench Hard | Artificial Analysis | 57.6% | 65.9% |
| τ²-Bench Telecom (AA) | Artificial Analysis | 87.1% | 85.1% |
| Terminal-Bench 2.1 (Vals) | Vals AI | – | 85.8% |
| MCP Atlas | Scale AI SEAL | 70.6% | – |
| HiL-Bench | Scale AI SEAL | 9.7%† | 32.3%† |
| τ³-bench Banking | τ²-bench | 39.4% | 46.9%† |
| τ³-Banking (AA) | Artificial Analysis | 39.6% | 44.3% |
| ITBench SRE (AA) | Artificial Analysis | – | 56.2% |
| Analyst Agent (AA) | Artificial Analysis | – | 47.5% |
| APEX-Agents (AA) | Artificial Analysis | 33.3% | – |
| CUA-bench | Vals AI | – | 8.3% |
| CyberBench | Vals AI | – | 76.3% |
| Finance Agent v2 | Vals AI | – | 53.8% |
| Harvey's Legal Agent Benchmark v1 | Vals AI | 0.0% | 2.5% |
| Legal Research Bench | Vals AI | – | 48.1% |
| SkillsBench | Vals AI | 51.7% | 54.1% |
| SREBench | Vals AI | – | 30.5% |
| Tax Agent Bench | Vals AI | – | 31.1% |
| Terminal-Bench 2.0 (Vals) | Vals AI | 58.4%† | – |
| Terminal-Bench 4.0 (Vals) | Vals AI | – | 37.9% |
| Terminal-Bench Science | Vals AI | – | 20.0% |
| Time Horizon Index: KSP | Vals AI | – | 23.8% |
| Web Search Index | Vals AI | – | 43.6% |
| Arena Search | Arena (LMArena) | – | 1258† |
| BALROG | BALROG | – | 60.0% |
| DeepResearch Bench | DeepResearch Bench | 35.1%† | – |
| PostTrainBench | Epoch AI Benchmarking Hub | 19.0%† | 36.2% |
| METR Time Horizons | METR | 74.3% | – |
| GBAEval | Epoch AI Benchmarking Hub | 45.1%† | 52.6%† |
| Vending-Bench 2 | Andon Labs | 6144.2† | 9619.4† |
| AutomationBench | Artificial Analysis | – | 60.1% |
| GDP.pdf (AA) | Artificial Analysis | – | 27.2% |
| EnterpriseOps-Gym | Artificial Analysis | – | 42.9% |
| AA-Briefcase v1.1 | Artificial Analysis | – | 1479 |
| Terminal-Bench-Science 0.1 (AA) | Artificial Analysis | – | 22.4% |
| DeepSearchQA | 63.7%† | – | |
| FACTS Search | – | 88.5%† | |
| Maths | |||
| FrontierMath Tiers 1–3 | Epoch AI Benchmarking Hub | 78.6% | 89.1% |
| FrontierMath Tier 4 | Epoch AI Benchmarking Hub | 49.0% | 82.9% |
| OTIS Mock AIME | Epoch AI Benchmarking Hub | 95.3% | 100.0% |
| ProofBench | Vals AI | 56.0% | 83.0% |
| LiveBench Mathematics | LiveBench | 94.2% | 96.2% |
| AIME (Vals) | Vals AI | 96.7% | – |
| AIME 2026 | MathArena | 99.2% | – |
| HMMT February 2026 | MathArena | 97.7% | – |
| USAMO 2026 | MathArena | 95.2% | – |
| MathArena Apex | MathArena | 54.2% | – |
| Arena Maths | Arena (LMArena) | 1460† | 1496† |
| Knowledge | |||
| SimpleQA Verified | Epoch AI Benchmarking Hub | 45.1% | 69.7% |
| LiveBench 2.0 Data Analysis | LiveBench | – | 79.8%† |
| LiveBench Research | LiveBench | – | 54.5%† |
| LiveBench 1.0 Data Analysis | LiveBench | 79.3% | 79.8% |
| AA-Omniscience | Artificial Analysis | 5.8 | 22.0 |
| MMLU-Pro | Vals AI | 87.5% | 89.1% |
| LegalBench | Vals AI | 86.0% | 87.0% |
| CorpFin | Vals AI | 65.3% | 64.4% |
| TaxEval | Vals AI | 74.0% | 74.8% |
| MedQA | Vals AI | 96.1% | – |
| PRBench Finance | Scale AI SEAL | 45.6%† | 50.5% |
| PRBench Legal | Scale AI SEAL | 44.4%† | 50.5% |
| MultiNRC | Scale AI SEAL | 58.3%† | – |
| BioMysteryBench | Vals AI | – | 71.1% |
| CaseLaw v2 | Vals AI | 63.8% | – |
| Excel Modeling Benchmark | Vals AI | – | 72.3% |
| MedCode | Vals AI | 41.3% | 44.0% |
| MedScribe | Vals AI | 77.5% | 85.2% |
| Public Benefits Bench v1.1 | Vals AI | – | 66.5% |
| DrugDiscoveryBench | Scale AI SEAL | – | 56.5% |
| EBR-bench | Epoch AI Benchmarking Hub | 25.4% | 44.8% |
| GDP.pdf | Epoch AI Benchmarking Hub | – | 30.7% |
| AA-Omniscience: accuracy | Artificial Analysis | 50.9% | 59.4% |
| AA-Omniscience: non-hallucination | Artificial Analysis | 8.3% | 7.8% |
| Instruction following | |||
| LiveBench Language | LiveBench | 82.6% | 87.7% |
| LiveBench Instruction Following | LiveBench | 70.2% | 71.8% |
| IFBench | Artificial Analysis | 74.0% | 72.7% |
| Arena Instruction Following | Arena (LMArena) | 1460† | 1491† |
| Human preference | |||
| Arena Text | Arena (LMArena) | 1465† | 1485† |
| EQ-Bench 4 | EQ-Bench | 1260† | 1240† |
| Arena Creative Writing | Arena (LMArena) | 1435† | 1470† |
| Arena Multi-turn | Arena (LMArena) | 1480† | 1490† |
| Arena Longer Queries | Arena (LMArena) | 1477† | 1499† |
| Multimodal | |||
| Arena Vision | Arena (LMArena) | 1292† | 1281† |
| LiveBench Design | LiveBench | – | 81.6%† |
| MMMU-Pro | Artificial Analysis | 78.4% | 83.4% |
| VISTA | Scale AI SEAL | 50.9%† | – |
| MMMU-Pro (Vals) | Vals AI | 87.5% | 88.8% |
| MortgageTax | Vals AI | 68.3% | 67.3% |
| SAGE | Vals AI | 43.3% | 52.6% |
| VTB | Scale AI SEAL | 29.2%† | – |
| Blueprint-Bench 2 | Epoch AI Benchmarking Hub | 27.1%† | 33.6%† |
| Long context | |||
| AA-LCR | Artificial Analysis | 82.0% | 84.0% |
| Arena Document | Arena (LMArena) | 1470† | 1489† |
| MLCR | Artificial Analysis | – | 30.0% |
| Composite | |||
| Epoch Capabilities Index | Epoch AI Benchmarking Hub | 157.3† | 162.8† |
| LiveBench 2.0 | LiveBench | – | 76.3%† |
| LiveBench 1.0 | LiveBench | 78.0% | 81.0% |
| AA Intelligence Index v4.3.2 | Artificial Analysis | 39.0 | 47.0 |
| Vals RSI Index | Vals AI | 15.5% | 23.9% |
| Vals Science Index | Vals AI | – | 41.5% |
| Vals Multimodal Index | Vals AI | – | 72.6% |
| Safety & honesty | |||
| FORTRESS | Scale AI SEAL | – | 20.1%† |
| MASK | Scale AI SEAL | 89.7%† | – |
Data as of 2026-10-25. Best configuration of each model; every score links to its source on the model pages.
What changed
- GPT-5.6 Sol (max): 1 new result (AA-Briefcase v1.1).
- GPT-5.4 (xhigh): 2 new results (Harvey's Legal Agent Benchmark v1, Vals RSI Index).
- GPT-5.6 Sol (max): 3 new results (Harvey's Legal Agent Benchmark v1, Vals RSI Index, Vals Science Index).
- GPT-5.6 Sol (max): 1 score revision.
- GPT-5.6 Sol (max): 1 new result (ForecastBench).
- GPT-5.4 (xhigh): 2 new results (USAMO 2026, τ³-bench Banking).
- GPT-5.6 Sol (max): 5 new results (AutomationBench, GDP.pdf (AA), MLCR…).
- GPT-5.6 Sol (max): 1 score revision.
- GPT-5.4 (xhigh): 1 score revision.
- GPT-5.6 Sol (max): 1 score revision.
Related comparisons
- GPT-5.4 vs Claude Opus 5.5
- GPT-5.4 vs Gemini 4 Argon
- GPT-5.4 vs Muse Wonder Full
- GPT-5.6 Sol vs Claude Opus 5.5
- GPT-5.6 Sol vs Gemini 4 Argon
- GPT-5.6 Sol vs Muse Wonder Full
- GPT-5.6 Sol vs Claude Fable 5.1
- GPT-5.4 Pro vs GPT-5.6 Sol
Rankings: Best LLM overall · Best LLM for agents · Best cheap LLMs
GPT-5.4 vs GPT-5.6 Sol: questions
- Is GPT-5.4 better than GPT-5.6 Sol?
- GPT-5.4 (xhigh) and GPT-5.6 Sol (max) are neck and neck on quality (64.7 vs 68.0, within the ±4.5 margin). The BenchLeader Index combines every independent quality benchmark; the two are within the margin of error as of 2026-10-25, so check the category scores for your use.
- Which is cheaper, GPT-5.4 or GPT-5.6 Sol, and will that change?
- GPT-5.4 is cheaper today: $5.63 against $8.00 per million tokens, blended at three input tokens per output token. No price change has been announced for either.
- Is GPT-5.4 better than GPT-5.6 Sol for coding?
- GPT-5.6 Sol scores higher in coding (66 vs 65 on the category index, where 50 is average). On LiveCodeBench, the top shared coding benchmark, it is 84.1% for GPT-5.4 against 82.6% for GPT-5.6 Sol.
- Is GPT-5.4 better than GPT-5.6 Sol for agents and tool use?
- GPT-5.6 Sol scores higher in agentic tasks (67 vs 64 on the category index, where 50 is average). On τ³-Banking (AA), the top shared agents & tools benchmark, it is 39.6% for GPT-5.4 against 44.3% for GPT-5.6 Sol.
- Which is better for long documents?
- GPT-5.6 Sol scores higher on long-context benchmarks (67 vs 66). Both accept 1,050,000 tokens of context.
- Which has lower latency?
- GPT-5.6 Sol (max) answers first: 69.5 s against 118.0 s to the first answer token, as Artificial Analysis measures it. Reasoning effort drives most of this: a higher setting thinks longer before it answers.
- Which is faster, GPT-5.4 or GPT-5.6 Sol?
- GPT-5.4 streams faster: 96 against 89 output tokens per second.
- Do independent benchmarks agree on GPT-5.4 vs GPT-5.6 Sol?
- Across the 55 benchmarks both have, GPT-5.4 leads 6, GPT-5.6 Sol leads 49. Every overall index that rates both puts GPT-5.6 Sol ahead.
- When was GPT-5.6 Sol released?
- GPT-5.6 Sol was released on 9 Jul 2026.