GPT-5.2 vs Step 5 Preview
Full profiles: GPT-5.2 (OpenAI) and Step 5 Preview (StepFun): every benchmark, price and rank.
Updated · Data from 10 sources
As of 14 Nov 2026, GPT-5.2 and Step 5 Preview are neck and neck on quality (59.9 vs 54.3 on the BenchLeader Index, inside the ±6.0 margin), and Step 5 Preview is 3.4× cheaper at $1.43 vs $4.81 per 1M tokens blended. Choose GPT-5.2 for coding (63 vs 50), reasoning (60 vs 52), and agents & tools (54 vs 50); choose Step 5 Preview for fast first answers (24.3 s vs 80.1 s) and maths (60 vs 57) and long context (69 vs 66), where it also streams ~2× faster.
Key facts (as of 14 Nov 2026)
- GPT-5.2 scores 59.9 on the BenchLeader Index and Step 5 Preview scores 54.3 (BenchLeader, 14 Nov 2026).
- GPT-5.2 costs $4.81 per 1M tokens blended (list price, 14 Nov 2026).
- Step 5 Preview costs $1.43 per 1M tokens blended (list price, 14 Nov 2026).
- Step 5 Preview streams ~2× faster than GPT-5.2 (Artificial Analysis, 7-day median, 14 Nov 2026).
- Of the 6 benchmarks both models have, Step 5 Preview leads 6 (BenchLeader, 14 Nov 2026).
GPT-5.2 benchmarks 
Step 5 Preview 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-5.2 | 63 vs 50 category score; Vibe Code Bench v1.1 53.5% vs 70.4% | High |
| Agents & tool use | GPT-5.2 | 54 vs 50 category score | Medium |
| Maths | Step 5 Preview | 60 vs 57 category score; ProofBench 42.0% vs 15.0% | Medium |
| Reasoning | GPT-5.2 | 60 vs 52 category score; Humanity's Last Exam (AA) 37.7% vs 46.5% | High |
| Knowledge | GPT-5.2 | 56 vs 54 category score | Medium |
| Long documents | Step 5 Preview | 69 vs 66 category score; AA-LCR v1.1 88.3% vs 82.7% | Medium |
| Chat quality (human preference) | – | Only Step 5 Preview has an Arena rating (63 category score) | Not comparable |
| Lowest cost | Step 5 Preview | $1.43 vs $4.81 per 1M blended; agentic coding session $0.071 vs $0.161 | – |
| Fastest first answer | Step 5 Preview | 24.3 s (Step 5 Preview) vs 80.1 s (GPT-5.2 (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-5.2 (xhigh) and Step 5 Preview are neck and neck on quality (59.9 vs 54.3, within the ±6.0 margin).
- GPT-5.2 (xhigh) is stronger in agents & tools, coding, knowledge, reasoning.
- Step 5 Preview is stronger in long context, maths.
- Step 5 Preview is 3.4× cheaper ($1.43 vs $4.81 per 1M blended).
- Step 5 Preview streams 1.9× faster (133 vs 71 tokens per second).
At a glance
| Spec | GPT-5.2 | Step 5 Preview |
|---|---|---|
| Developer | OpenAI | StepFun |
| Released | 11 Dec 2025 | 18 Sep 2026 |
| Context window (tokens) | 400,000 | 1,000,000 |
| Compared at | GPT-5.2 (xhigh) | Step 5 Preview |
| Open weights | No | Yes |
| Price in / out / cached input per 1M | $1.75 / $14.00 / $0.175 | $1.00 / $2.70 / $0.105 |
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-5.2 (xhigh) | Step 5 Preview | Leader |
|---|---|---|---|
| Reasoning | 60.4 | 51.7 | GPT-5.2 |
| Coding | 63.0 | 49.9 | GPT-5.2 |
| Agents & tools | 53.8 | 49.7 | GPT-5.2 |
| Maths | 56.5 | 60.1 | Step 5 Preview |
| Knowledge | 56.1 | 53.7 | GPT-5.2 |
| Instruction following | 73.7 | – | not comparable |
| Human preference | – | 63.1 | not comparable |
| Multimodal | – | 59.3 | not comparable |
| Long context | 66.0 | 69.0 | Step 5 Preview |
| Composite | – | 38.3 | not comparable |
On one chart
Each domain on the index's own scale, where 50 is an average model. Click a domain to see the benchmarks behind it. Pick up to 5 models from the list.
Not drawn, measured for fewer than two of these models: Instruction following, Multimodal, Human preference.
Model, click to add or removeIndex
Where sources disagree
Across the 6 benchmarks both have, GPT-5.2 leads 0, Step 5 Preview leads 6. Every overall index that rates both puts Step 5 Preview ahead.
| Overall index | GPT-5.2 (xhigh) | Step 5 Preview | Leader |
|---|---|---|---|
| BenchLeader Index | 59.9 | 54.3 | neck and neck |
| Arena Text | 1437† | 1449 | Step 5 Preview |
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. Cost, tokens and time per benchmark task are measured: Artificial Analysis’ averages over its whole Intelligence Index run.
| Measure | GPT-5.2 (xhigh) | Step 5 Preview |
|---|---|---|
| Blended price per 1M tokens | $4.81 | $1.43 |
| Input / output per 1M | $1.75 / $14.00 | $1.00 / $2.70 |
| Cached input per 1M | $0.175 | $0.105 |
| Price per task: chat reply | $0.0049 | $0.0012 |
| Price per task: summarize a 30-page report | $0.029 | $0.014 |
| Price per task: code edit | $0.032 | $0.010 |
| Price per task: agentic coding session | $0.161 | $0.071 |
| Price per task: structured extraction | $0.0063 | $0.0025 |
| Agentic coding session, input mostly cached | $0.090 | $0.030 |
| Cost per benchmark task (AA Index) | – | $2.35 |
| Output tokens per task (AA Index) | – | 113k |
| Of which reasoning | – | 80k |
| Time per task (AA Index) | – | 14 min |
| Output speed | 71 tok/s | 133 tok/s |
| Time to first answer | 80.1 s | 24.3 s |
| Time to a complete answer | 161.0 s | 57.4 s |
Each model is compared at its best-ranked reasoning-effort setting. GPT-5.2 (xhigh) scores 59.9 and takes 80 s to its first answer token; GPT-5.2 scores 58.0 in 2 s, GPT-5.2 (medium) scores 56.2 in 11 s. Step 5 Preview scores 54.3 and takes 24 s to its first answer token. 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
163 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-5.2 (xhigh) | Step 5 Preview |
|---|---|---|---|
| BenchLeader Index | BenchLeader | 59.9 | 54.3 |
| Agents & tools score | BenchLeader | 53.8 | 49.7 |
| Coding score | BenchLeader | 63.0 | 49.9 |
| Instruction following score | BenchLeader | 73.7 | – |
| Knowledge score | BenchLeader | 56.1 | 53.7 |
| Long context score | BenchLeader | 66.0 | 69.0 |
| Maths score | BenchLeader | 56.5 | 60.1 |
| Reasoning score | BenchLeader | 60.4 | 51.7 |
| Composite score | BenchLeader | – | 38.3 |
| Human preference score | BenchLeader | – | 63.1 |
| Multimodal score | BenchLeader | – | 59.3 |
| Blended price $/M | BenchLeader | $4.81 | $1.43 |
| Output speed | BenchLeader | 71 tok/s | 133 tok/s |
| Time to first answer | BenchLeader | 80.1 s | 24.3 s |
| Context window | BenchLeader | 400,000 tokens | 1,000,000 tokens |
| Reasoning | |||
| GPQA Diamond | Epoch AI Benchmarking Hub | 91.4% | – |
| Humanity's Last Exam | Scale AI / CAIS | 27.8%† | – |
| SimpleBench | SimpleBench | 45.8%† | – |
| Arena Hard Prompts | Arena (LMArena) | 1497† | 1477 |
| LiveBench 2.0 Reasoning | LiveBench | – | 83.8% |
| LiveBench 1.0 Reasoning | LiveBench | 83.2%† | – |
| GPQA Diamond (AA) | Artificial Analysis | 90.3% | – |
| Humanity's Last Exam (AA) | Artificial Analysis | 37.7% | 46.5% |
| HLE-Diamond (AA) | Artificial Analysis | – | 24.3% |
| GPQA Diamond (Vals) | Vals AI | 91.7% | – |
| Kagi LLM Benchmark | Kagi LLM Benchmark | 73.3%† | – |
| ARC-AGI-1 | ARC Prize | 86.2% | – |
| ARC-AGI-2 | ARC Prize | 52.9% | – |
| CritPt | Artificial Analysis | 11.6% | 20.9% |
| CritPt v2 hard subset (AA) | Artificial Analysis | – | 32.5% |
| MysteryMechanism | Vals AI | – | 18.5% |
| SciPredict | Scale AI SEAL | 20.6%† | – |
| Chess Puzzles | Epoch AI Benchmarking Hub | 48.0% | – |
| Mystery Game Puzzles | Epoch AI Benchmarking Hub | 22.0%† | – |
| CL-bench | Epoch AI Benchmarking Hub | 18.2%† | – |
| LMCA | Epoch AI Benchmarking Hub | 43.9% | – |
| DTBench | Epoch AI Benchmarking Hub | 90.9% | – |
| ForecastBench | ForecastBench | 63.8%† | – |
| Coding | |||
| SWE-bench Verified (Epoch) | Epoch AI Benchmarking Hub | 73.8%† | – |
| SciCode | SciCode | – | 58.9% |
Show all 163 benchmarks
| Metric | Publisher | GPT-5.2 (xhigh) | Step 5 Preview |
|---|---|---|---|
| Coding | |||
| WeirdML | WeirdML | 72.2% | – |
| GSO-Bench | GSO-Bench | 27.4%† | – |
| Arena Coding | Arena (LMArena) | 1515† | 1505 |
| Arena WebDev | Arena (LMArena) | 1421† | 1573† |
| LiveBench 2.0 Coding | LiveBench | – | 67.6% |
| LiveBench 1.0 Coding | LiveBench | 76.1%† | – |
| LiveCodeBench | Vals AI | 85.4% | – |
| IOI | Vals AI | – | 38.1% |
| SWE-bench (Vals) | Vals AI | 75.8% | – |
| SWE-Bench Pro | Scale AI SEAL | 29.9%† | – |
| SWE-bench Verified (bash only) | SWE-bench | 69.0%† | – |
| SWE-bench Verified (any scaffold) | SWE-bench | 72.8%† | – |
| Code Migration | Vals AI | – | 39.0% |
| Code Migration v1.1 | Vals AI | – | 47.0% |
| SWE-bench Multimodal | Vals AI | – | 40.8% |
| Vibe Code Bench v1.1 | Vals AI | 53.5% | 70.4% |
| SWE-Bench Pro (private) | Scale AI SEAL | 29.9%† | – |
| MirrorCode | Epoch AI Benchmarking Hub | 32.2% | – |
| ALE-Bench | Epoch AI Benchmarking Hub | 1249.8† | – |
| AlgoTune | AlgoTune | 2.0† | – |
| Terminal-Bench 4.0 (AA) | Artificial Analysis | – | 33.3% |
| Design Arena Website | Design Arena | 1186 | – |
| FrontierCode 1.1 (AA) | Artificial Analysis | – | 38.4% |
| FrontierCode 2.0 (AA) | Artificial Analysis | – | 12.0% |
| LiveCodeBench (AA) | Artificial Analysis | 88.9% | – |
| Agents & tools | |||
| Terminal-Bench | Terminal-Bench | 64.9%† | – |
| GDPval | OpenAI | 49.7%† | – |
| Remote Labor Index | Scale AI / CAIS | 2.1%† | – |
| Arena Agent | Arena (LMArena) | – | -3.7 |
| LiveBench Long Agentic | LiveBench | – | 19.4% |
| LiveBench Browser | LiveBench | – | 24.8% |
| LiveBench Agentic Coding | LiveBench | 50.3%† | – |
| Terminal-Bench Hard | Artificial Analysis | 47.0% | – |
| τ²-Bench Telecom (AA) | Artificial Analysis | 84.8% | – |
| MCP Atlas | Scale AI SEAL | 67.6% | – |
| BFCL Overall | Berkeley Function Calling Leaderboard | 55.9%† | – |
| τ²-bench | τ²-bench | 84.8%† | – |
| τ³-bench Banking | τ²-bench | 32.2%† | – |
| GDPval-AA v2.1, normalised score | Artificial Analysis | – | 54.3% |
| GDPval-AA v2.1 | Artificial Analysis | – | 1591 |
| τ³-Banking (AA) | Artificial Analysis | 11.1%† | – |
| ITBench SRE (AA) | Artificial Analysis | – | 55.6% |
| Analyst Agent (AA) | Artificial Analysis | – | 35.0% |
| APEX-Agents (AA) | Artificial Analysis | – | 38.0% |
| Finance Agent v2 | Vals AI | – | 50.7% |
| Finance Agent v2.1 | Vals AI | – | 58.9% |
| Harvey's Legal Agent Benchmark v1.1 | Vals AI | – | 12.9% |
| Harvey's Legal Agent Benchmark v1 | Vals AI | – | 9.2% |
| Legal Research Bench | Vals AI | – | 40.4% |
| Poker Agent | Vals AI | 1131.8%† | – |
| SREBench | Vals AI | – | 0.4% |
| Tax Agent Bench | Vals AI | – | 27.3% |
| Terminal-Bench 2.0 (Vals) | Vals AI | 51.7%† | – |
| Terminal-Bench 4.0 (Vals) | Vals AI | – | 31.8% |
| Terminal-Bench Science | Vals AI | – | 2.9% |
| Arena Search | Arena (LMArena) | 1161† | – |
| DeepResearch Bench | DeepResearch Bench | 41.1%† | – |
| METR Time Horizons | METR | 75.3%† | – |
| Vending-Bench 2 | Andon Labs | 3591.3† | – |
| AutomationBench | Artificial Analysis | – | 51.0% |
| GDP.pdf (AA) | Artificial Analysis | – | 14.8% |
| EnterpriseOps-Gym | Artificial Analysis | – | 47.2% |
| AA-Briefcase v1.1 | Artificial Analysis | – | 1424 |
| Design Arena Agentic | Design Arena | 1059† | – |
| Terminal-Bench-Science 0.1 (AA) | Artificial Analysis | – | 2.4% |
| CWE-Bench (AA) | Artificial Analysis | – | 52.5% |
| DeepSec-Bench (AA) | Artificial Analysis | – | 15.8% |
| CyberGym end-to-end (AA) | Artificial Analysis | – | 74.8% |
| Maths | |||
| FrontierMath Tiers 1–3 | Epoch AI Benchmarking Hub | 67.4% | – |
| FrontierMath Tier 4 | Epoch AI Benchmarking Hub | 31.7% | – |
| OTIS Mock AIME | Epoch AI Benchmarking Hub | 96.1% | – |
| ProofBench | Vals AI | 15.0% | 42.0% |
| LiveBench Mathematics | LiveBench | 93.2%† | – |
| AIME (Vals) | Vals AI | 96.9% | – |
| MGSM | Vals AI | 94.0% | – |
| AIME 2026 | MathArena | 98.3%† | – |
| HMMT February 2026 | MathArena | 97.0%† | – |
| MathArena Apex | MathArena | 13.5%† | – |
| Arena Maths | Arena (LMArena) | 1453† | 1470 |
| Knowledge | |||
| SimpleQA Verified | Epoch AI Benchmarking Hub | 37.1% | – |
| LiveBench 2.0 Data Analysis | LiveBench | – | 63.5% |
| LiveBench Research | LiveBench | – | 75.5% |
| LiveBench 1.0 Data Analysis | LiveBench | 78.2%† | – |
| AA-Omniscience v1.1 | Artificial Analysis | – | 16.4 |
| MMLU-Pro | Vals AI | 86.2% | – |
| LegalBench | Vals AI | 82.8% | – |
| CorpFin | Vals AI | 65.9% | – |
| TaxEval | Vals AI | 75.8% | – |
| MedQA | Vals AI | 94.1% | – |
| MultiNRC | Scale AI SEAL | 42.2%† | – |
| BioMysteryBench | Vals AI | – | 67.4% |
| CaseLaw v2 | Vals AI | 66.0% | – |
| Excel Modeling Benchmark | Vals AI | – | 57.1% |
| Excel Modeling Benchmark v1.1 | Vals AI | – | 69.5% |
| MedCode | Vals AI | 49.8% | – |
| MedScribe | Vals AI | 84.4% | – |
| Public Benefits Bench v1.1 | Vals AI | – | 58.1% |
| DrugDiscoveryBench | Scale AI SEAL | 29.3% | – |
| EBR-bench | Epoch AI Benchmarking Hub | 23.0% | – |
| AA-Omniscience v1.1: accuracy | Artificial Analysis | – | 36.4% |
| AA-Omniscience v1.1: non-hallucination | Artificial Analysis | – | 68.7% |
| Global-MMLU-Lite (AA) | Artificial Analysis | 89.8%† | – |
| Instruction following | |||
| LiveBench Language | LiveBench | 79.8%† | – |
| LiveBench Instruction Following | LiveBench | 61.8%† | – |
| IFBench | Artificial Analysis | 75.4% | – |
| TutorBench | Scale AI SEAL | 53.5%† | – |
| Arena Instruction Following | Arena (LMArena) | 1457† | 1450 |
| Human preference | |||
| Arena Text | Arena (LMArena) | 1476† | 1449 |
| Arena Creative Writing | Arena (LMArena) | 1432† | 1413 |
| Arena Multi-turn | Arena (LMArena) | 1495† | 1445 |
| Arena Longer Queries | Arena (LMArena) | 1472† | 1461 |
| Multimodal | |||
| Arena Vision | Arena (LMArena) | 1268† | 1263 |
| LiveBench Design | LiveBench | – | 70.8% |
| MMMU-Pro | Artificial Analysis | 74.6%† | 76.4% |
| VISTA | Scale AI SEAL | 46.6%† | – |
| MMMU-Pro (Vals) | Vals AI | 86.7% | – |
| MortgageTax | Vals AI | 67.1% | – |
| VPCT | Epoch AI Benchmarking Hub | 84.0% | – |
| Code2Video-AA | Artificial Analysis | – | 821 |
| AA-3D | Artificial Analysis | – | 859 |
| Long context | |||
| AA-LCR v1.1 | Artificial Analysis | 82.7% | 88.3% |
| Arena Document | Arena (LMArena) | 1425† | – |
| MLCR | Artificial Analysis | – | 18.3% |
| Composite | |||
| Epoch Capabilities Index | Epoch AI Benchmarking Hub | 154.5† | – |
| LiveBench 2.0 | LiveBench | – | 54.9% |
| LiveBench 1.0 | LiveBench | 74.6%† | – |
| AA Intelligence Index v5.0 | Artificial Analysis | – | 28.5 |
| AA Intelligence Index v4.3.2 | Artificial Analysis | 30.4 | 43.7 |
| AA Progress Index v1 | Artificial Analysis | 295.5 | 318.2 |
| Vals Index | Vals AI | – | 54.8 |
| Vals Finance Index | Vals AI | – | 55.4% |
| Vals RSI Index | Vals AI | 15.8% | – |
| Vals Science Index | Vals AI | – | 29.6% |
| AA Cyber Index | Artificial Analysis | – | 47.7 |
| Safety & honesty | |||
| FORTRESS | Scale AI SEAL | 17.5%† | – |
| MASK | Scale AI SEAL | 86.7%† | – |
| PropensityBench | Scale AI SEAL | 34.4%† | – |
| Metadata | |||
| AA Openness Index | Artificial Analysis | – | 33.3 |
Data as of 2026-11-14. Best configuration of each model; every score links to its source on the model pages.
What changed
- Step 5 Preview: 1 score revision.
- GPT-5.2 (xhigh): 1 new result (MirrorCode).
- Step 5 Preview: 7 score revisions.
- Step 5 Preview: 2 new results (Vals Index, Harvey's Legal Agent Benchmark v1.1).
- Step 5 Preview: 1 score revision.
- Step 5 Preview: BenchLeader Index 58.0 → 54.7.
- GPT-5.2 (xhigh): 1 new result (AA Progress Index v1).
- Step 5 Preview: 5 new results (AA Intelligence Index v5.0, AA Progress Index v1, HLE-Diamond (AA)…); 2 score revisions; Step 5 Preview cost per task moved from $1.03/M to $2.35/M (128% up).
- Step 5 Preview: 3 score revisions.
- Step 5 Preview: 1 score revision.
Related comparisons
- GPT-5.2 vs Claude Opus 5.5
- GPT-5.2 vs GPT-6 Astra
- Step 5 Preview vs Claude Opus 5.5
- Step 5 Preview vs GPT-6 Astra
- Step 5 Preview vs GPT-6.1 Sol
- Step 5 Preview vs Gemini 4 Argon
- GPT-5.2 vs Claude Fable 5.1
- GPT-5.2 vs GPT-6.1 Sol
Rankings: Best LLM overall · Best LLM for agents · Best cheap LLMs
GPT-5.2 vs Step 5 Preview: questions
- Is GPT-5.2 better than Step 5 Preview?
- GPT-5.2 (xhigh) and Step 5 Preview are neck and neck on quality (59.9 vs 54.3, within the ±6.0 margin). The BenchLeader Index combines every independent quality benchmark; the two are within the margin of error as of 14 Nov 2026, so check the category scores for your use.
- GPT-5.2 vs Step 5 Preview: which is cheaper, and will that change?
- Step 5 Preview is cheaper today: $1.43 against $4.81 per million tokens, blended at three input tokens per output token. No price change has been announced for either.
- Is GPT-5.2 or Step 5 Preview better for coding?
- GPT-5.2 scores higher in coding (63 vs 50 on the category index, where 50 is average).
- Which is faster, GPT-5.2 or Step 5 Preview?
- Step 5 Preview streams ~2× faster (133 vs 71 output tokens per second, Artificial Analysis 7-day median).
- Is GPT-5.2 or Step 5 Preview better for agents and tool use?
- GPT-5.2 scores higher in agentic tasks (54 vs 50 on the category index, where 50 is average).
More questions (4)
- Which is better for long documents?
- Step 5 Preview scores higher on long-context benchmarks (69 vs 66). Step 5 Preview accepts more context: 1,000,000 against 400,000 tokens for GPT-5.2.
- Which has lower latency?
- Step 5 Preview answers first: 24.3 s against 80.1 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.
- Do independent benchmarks agree on GPT-5.2 vs Step 5 Preview?
- Across the 6 benchmarks both have, GPT-5.2 leads 0, Step 5 Preview leads 6. Every overall index that rates both puts Step 5 Preview ahead.
- When was Step 5 Preview released?
- Step 5 Preview was released on 18 Sep 2026.