GPT-5.5 vs Qwen3.8 27B
Full profiles: GPT-5.5 (OpenAI) and Qwen3.8 27B (Alibaba): every benchmark, price and rank.
Updated · Data from 13 sources
As of 16 Nov 2026, GPT-5.5 is ahead of Qwen3.8 27B on quality (64.1 vs 55.1 on the BenchLeader Index), but Qwen3.8 27B is 10× cheaper at $1.09 vs $11.25 per 1M tokens blended. GPT-5.5 leads in every category we measure both in.
Key facts (as of 16 Nov 2026)
- GPT-5.5 scores 64.1 on the BenchLeader Index and Qwen3.8 27B scores 55.1 (BenchLeader, 16 Nov 2026).
- GPT-5.5 costs $11.25 per 1M tokens blended (list price, 16 Nov 2026).
- Qwen3.8 27B costs $1.09 per 1M tokens blended (list price, 16 Nov 2026).
- Of the 12 benchmarks both models have, GPT-5.5 leads 9 (BenchLeader, 16 Nov 2026).
- On Arena WebDev, GPT-5.5 is rated 1512 and Qwen3.8 27B 1590 (LMArena, as published by 16 Nov 2026).
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 | Either | 61 vs 62 category score; Arena WebDev 1512 vs 1590 | Either |
| Agents & tool use | GPT-5.5 | 62 vs 44 category score; LiveBench Agentic Coding 54.0% vs 61.4% | High |
| Maths | GPT-5.5 | 69 vs 55 category score; ProofBench 50.0% vs 16.0% | High |
| Reasoning | GPT-5.5 | 68 vs 58 category score; LiveBench 1.0 Reasoning 89.7% vs 80.0% | High |
| Images and other media | GPT-5.5 | 62 vs 61 category score | Medium |
| Chat quality (human preference) | – | Only Qwen3.8 27B has an Arena rating (61 category score) | Not comparable |
| Lowest cost | Qwen3.8 27B | $1.06 vs $11.25 per 1M blended; agentic coding session $0.037 vs $0.420 | – |
| Fastest first answer | Qwen3.8 27B | 0.7 s (Qwen3.8 27B) vs 38.3 s (GPT-5.5 (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-5.5 (xhigh) leads on quality: 64.1 vs 55.1.
- GPT-5.5 (xhigh) is stronger in agents & tools, maths, multimodal, reasoning.
- Qwen3.8 27B is stronger in coding.
- Qwen3.8 27B is 10× cheaper ($1.09 vs $11.25 per 1M blended).
At a glance
| Spec | GPT-5.5 | Qwen3.8 27B |
|---|---|---|
| Developer | OpenAI | Alibaba |
| Released | 23 Apr 2026 | 14 Aug 2026 |
| Context window (tokens) | 1,050,000 | 262,144 |
| Compared at | GPT-5.5 (xhigh) | Qwen3.8 27B |
| Open weights | No | Yes |
| Price in / out / cached input per 1M | $5.00 / $30.00 / $0.500 | $0.420 / $3.00 / $0.085 |
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-5.5 (xhigh) | Qwen3.8 27B | Leader |
|---|---|---|---|
| Reasoning | 68.1 | 57.6 | GPT-5.5 |
| Coding | 60.7 | 61.6 | level |
| Agents & tools | 61.8 | 44.2 | GPT-5.5 |
| Maths | 69.3 | 54.6 | GPT-5.5 |
| Knowledge | 63.6 | – | not comparable |
| Instruction following | 74.0 | – | not comparable |
| Human preference | – | 61.3 | not comparable |
| Multimodal | 62.3 | 60.7 | GPT-5.5 |
| Long context | 66.9 | – | 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: Knowledge, Instruction following, Long context, Human preference.
Where sources disagree
Across the 12 benchmarks both have, GPT-5.5 leads 9, Qwen3.8 27B leads 3. Every overall index that rates both puts GPT-5.5 ahead.
| Overall index | GPT-5.5 (xhigh) | Qwen3.8 27B | Leader |
|---|---|---|---|
| BenchLeader Index | 64.1 | 55.1 | GPT-5.5 |
| Arena Text | 1481† | 1434 | GPT-5.5 |
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-5.5 (xhigh) | Qwen3.8 27B |
|---|---|---|
| Blended price per 1M tokens | $11.25 | $1.06 |
| Input / output per 1M | $5.00 / $30.00 | $0.420 / $3.00 |
| Cached input per 1M | $0.500 | $0.085 |
| Price per task: chat reply | $0.011 | $0.0011 |
| Price per task: summarize a 30-page report | $0.078 | $0.0068 |
| Price per task: code edit | $0.075 | $0.0070 |
| Price per task: agentic coding session | $0.420 | $0.037 |
| Price per task: structured extraction | $0.016 | $0.0014 |
| Agentic coding session, input mostly cached | $0.217 | $0.022 |
| Output speed | 81 tok/s | 73 tok/s |
| Time to first answer | 38.3 s | 0.7 s |
| Time to a complete answer, last measured 13 Nov 2026 | 104.3 s | – |
Each model is compared at its best-ranked reasoning-effort setting. GPT-5.5 (xhigh) scores 64.1 and takes 38 s to its first answer token; GPT-5.5 (high) scores 63.5 in 16 s, GPT-5.5 scores 61.0 in 3 s. Qwen3.8 27B scores 55.1 and takes 1 s to its first answer token; Qwen3.8 27B (xhigh) scores 52.2 in 109 s, Qwen3.8 27B (medium) scores 50.2 in 24 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
161 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-5.5 (xhigh) | Qwen3.8 27B |
|---|---|---|---|
| BenchLeader Index | BenchLeader | 64.1 | 55.1 |
| Agents & tools score | BenchLeader | 61.8 | 44.2 |
| Coding score | BenchLeader | 60.7 | 61.6 |
| Instruction following score | BenchLeader | 74.0 | – |
| Knowledge score | BenchLeader | 63.6 | – |
| Long context score | BenchLeader | 66.9 | – |
| Maths score | BenchLeader | 69.3 | 54.6 |
| Multimodal score | BenchLeader | 62.3 | 60.7 |
| Reasoning score | BenchLeader | 68.1 | 57.6 |
| Human preference score | BenchLeader | – | 61.3 |
| Blended price $/M | BenchLeader | $11.25 | $1.06 |
| Output speed | BenchLeader | 81 tok/s | 73 tok/s |
| Time to first answer | BenchLeader | 38.3 s | 0.7 s |
| Context window | BenchLeader | 1,050,000 tokens | 262,144 tokens |
| Reasoning | |||
| GPQA Diamond | Epoch AI Benchmarking Hub | 90.7%† | – |
| SimpleBench | SimpleBench | 69.0%† | 60.2% |
| Arena Hard Prompts | Arena (LMArena) | 1499† | 1459 |
| LiveBench 2.0 Reasoning | LiveBench | – | 80.0%† |
| LiveBench 1.0 Reasoning | LiveBench | 89.7% | 80.0% |
| GPQA Diamond (AA) | Artificial Analysis | 93.5% | 90.5%† |
| Humanity's Last Exam (AA) | Artificial Analysis | 45.8% | 33.9%† |
| HLE-Diamond (AA) | Artificial Analysis | – | 15.3%† |
| GPQA Diamond (Vals) | Vals AI | 93.2% | 88.9%† |
| Kagi LLM Benchmark | Kagi LLM Benchmark | 88.8%† | – |
| ARC-AGI-1 | ARC Prize | 95.0% | 87.5%† |
| ARC-AGI-2 | ARC Prize | 85.0% | 42.4%† |
| ARC-AGI-3 | ARC Prize | 0.4%† | – |
| CritPt | Artificial Analysis | 27.1% | 5.4%† |
| CritPt v2 hard subset (AA) | Artificial Analysis | – | 4.0%† |
| Arena Expert | Arena (LMArena) | 1510† | 1473 |
| Chess Puzzles | Epoch AI Benchmarking Hub | 63.0% | – |
| Mystery Game Puzzles | Epoch AI Benchmarking Hub | 56.0% | – |
| CL-bench Life | Epoch AI Benchmarking Hub | 22.2%† | – |
| LMCA | Epoch AI Benchmarking Hub | 54.3% | 41.4% |
| DTBench | Epoch AI Benchmarking Hub | 96.0% | 88.0% |
| ForecastBench | ForecastBench | 64.2%† | – |
| BTF-3 | Epoch AI Benchmarking Hub | 14.3%† | – |
| Coding | |||
| SciCode | SciCode | 55.8% | 46.6%† |
| WeirdML | WeirdML | 84.9% | – |
Show all 161 benchmarks
| Metric | Publisher | GPT-5.5 (xhigh) | Qwen3.8 27B |
|---|---|---|---|
| Coding | |||
| FrontierCode | Cognition | 43.0%† | – |
| GSO-Bench | GSO-Bench | 40.2% | – |
| Arena Coding | Arena (LMArena) | 1510† | 1496 |
| Arena WebDev | Arena (LMArena) | 1512 | 1590 |
| LiveBench 2.0 Coding | LiveBench | – | 69.3%† |
| LiveBench 1.0 Coding | LiveBench | 82.2% | 75.7% |
| LiveCodeBench | Vals AI | 85.3% | 84.0%† |
| IOI | Vals AI | – | 39.1%† |
| SWE-bench (Vals) | Vals AI | 82.6% | 86.0%† |
| Code Migration | Vals AI | 45.2% | 14.2%† |
| ProgramBench | Vals AI | 0.5% | 0.0%† |
| Vibe Code Bench v1.1 | Vals AI | 69.8% | 64.8%† |
| SWE Atlas v1.0: Codebase QnA | Scale AI SEAL | 45.4% | – |
| SWE Atlas v1.0: Refactoring | Scale AI SEAL | 44.8% | – |
| SWE Atlas v1.0: Test Writing | Scale AI SEAL | 42.6% | – |
| MirrorCode | Epoch AI Benchmarking Hub | 23.3% | – |
| Surface Evolver Bench | Epoch AI Benchmarking Hub | 88.1%† | 45.0%† |
| DeepSWE v1.1 | Epoch AI Benchmarking Hub | 67.0% | – |
| ALE-Bench | Epoch AI Benchmarking Hub | 1943.0 | – |
| Terminal-Bench 4.0 (AA) | Artificial Analysis | 14.7% | 5.6%† |
| Terminal-Bench 2.1 (AA) | Artificial Analysis | 84.3% | 79.8%† |
| GameBench 2 Overall | DuelLab GameBench 2 | 49.2† | – |
| GameBench 2 by reasoning setting | DuelLab GameBench 2 | 59.9 | 43.8† |
| Design Arena Website | Design Arena | 1248† | – |
| FrontierCode 1.1 (AA) | Artificial Analysis | – | 34.2%† |
| FrontierCode 2.0 (AA) | Artificial Analysis | – | 14.9%† |
| Agents & tools | |||
| Terminal-Bench | Terminal-Bench | 84.7%† | – |
| OSWorld-Verified 2.0 | OSWorld | 13.0% | – |
| Remote Labor Index | Scale AI / CAIS | 6.3%† | – |
| APEX-Agents | Mercor | 55.1%† | 47.5% |
| Arena Agent | Arena (LMArena) | – | -4.8 |
| LiveBench Long Agentic | LiveBench | – | 15.9%† |
| LiveBench Browser | LiveBench | – | 24.2%† |
| LiveBench Agentic Coding | LiveBench | 54.0% | 61.4% |
| Terminal-Bench Hard | Artificial Analysis | 60.6% | – |
| τ²-Bench Telecom (AA) | Artificial Analysis | 93.9% | – |
| Terminal-Bench 2.1 (Vals) | Vals AI | 57.3%† | 58.4%† |
| MCP Atlas | Scale AI SEAL | 75.3% | – |
| HiL-Bench | Scale AI SEAL | 39.7%† | – |
| τ³-bench Banking | τ²-bench | 44.6% | – |
| GDPval-AA v2.1, normalised score | Artificial Analysis | – | 46.0%† |
| GDPval-AA v2.1 | Artificial Analysis | 1357 | 1426† |
| τ³-Banking (AA) | Artificial Analysis | 39.0% | 48.0%† |
| ITBench SRE (AA) | Artificial Analysis | 45.8% | – |
| Analyst Agent (AA) | Artificial Analysis | 50.0% | – |
| APEX-Agents (AA) | Artificial Analysis | 37.7% | – |
| CyberBench | Vals AI | – | 67.8%† |
| Finance Agent v2 | Vals AI | 51.8% | 48.5%† |
| Harvey's Legal Agent Benchmark v1 | Vals AI | 3.8% | 11.3%† |
| Legal Research Bench | Vals AI | 40.4% | 36.1%† |
| SkillsBench | Vals AI | 62.2% | 38.1%† |
| SREBench | Vals AI | 3.8% | – |
| Tax Agent Bench | Vals AI | 21.8% | 30.9%† |
| Terminal-Bench 2.0 (Vals) | Vals AI | 73.2%† | – |
| Terminal-Bench Science | Vals AI | – | 0.0%† |
| Time Horizon Index: KSP | Vals AI | 8.3% | – |
| Arena Search | Arena (LMArena) | 1226† | – |
| Arena Agent Task Completion | Arena (LMArena) | – | -2.7 |
| DeepResearch Bench | DeepResearch Bench | 54.0%† | – |
| PostTrainBench | Epoch AI Benchmarking Hub | 23.9% | – |
| ExploitBench | Epoch AI Benchmarking Hub | 47.4%† | – |
| GBAEval | Epoch AI Benchmarking Hub | 53.2%† | – |
| Vending-Bench 2 | Andon Labs | 7523.8† | – |
| AutomationBench | Artificial Analysis | 47.3% | 48.2%† |
| GDP.pdf (AA) | Artificial Analysis | 21.2% | 16.4%† |
| EnterpriseOps-Gym | Artificial Analysis | 46.6% | 44.2%† |
| AA-Briefcase v1.1 | Artificial Analysis | 1135 | 1391† |
| Design Arena Agentic | Design Arena | 1113† | – |
| RoboDojo (LLM as robot policy) | RoboDojo | 1.1† | – |
| Terminal-Bench-Science 0.1 (AA) | Artificial Analysis | – | 0.5%† |
| CWE-Bench (AA) | Artificial Analysis | – | 26.7%† |
| DeepSec-Bench (AA) | Artificial Analysis | – | 11.5%† |
| CyberGym end-to-end (AA) | Artificial Analysis | – | 0.0%† |
| Maths | |||
| FrontierMath Tiers 1–3 | Epoch AI Benchmarking Hub | 85.3% | – |
| FrontierMath Tier 4 | Epoch AI Benchmarking Hub | 72.5% | – |
| OTIS Mock AIME | Epoch AI Benchmarking Hub | 84.4%† | – |
| ProofBench | Vals AI | 50.0% | 16.0% |
| LiveBench Mathematics | LiveBench | 95.9% | 86.2% |
| AIME 2026 | MathArena | 100.0% | – |
| HMMT February 2026 | MathArena | 98.5% | – |
| USAMO 2026 | MathArena | 98.2% | – |
| MathArena Apex | MathArena | 80.2% | – |
| Arena Maths | Arena (LMArena) | 1500† | 1451 |
| FrontierMath Erdős | Epoch AI Benchmarking Hub | 0.0% | – |
| Knowledge | |||
| SimpleQA Verified | Epoch AI Benchmarking Hub | 63.0% | – |
| LiveBench 2.0 Data Analysis | LiveBench | – | 76.6%† |
| LiveBench Research | LiveBench | – | 61.3%† |
| LiveBench 1.0 Data Analysis | LiveBench | 81.6% | 76.6% |
| AA-Omniscience v1.1 | Artificial Analysis | – | -6.4† |
| MMLU-Pro | Vals AI | 88.1% | 84.3%† |
| LegalBench | Vals AI | 86.5% | 82.4%† |
| CorpFin | Vals AI | 68.4% | – |
| TaxEval | Vals AI | 75.0% | 70.8%† |
| CaseLaw v2 | Vals AI | 66.2% | – |
| Excel Modeling Benchmark | Vals AI | 64.5% | 59.7%† |
| MedCode | Vals AI | 49.1% | 28.7%† |
| MedScribe | Vals AI | 86.9% | 83.8%† |
| Public Benefits Bench v1.1 | Vals AI | 60.9% | – |
| DrugDiscoveryBench | Scale AI SEAL | 51.6% | – |
| EBR-bench | Epoch AI Benchmarking Hub | 34.3% | – |
| GDP.pdf | Epoch AI Benchmarking Hub | 26.0% | – |
| AA-Omniscience v1.1: accuracy | Artificial Analysis | – | 17.3%† |
| AA-Omniscience v1.1: non-hallucination | Artificial Analysis | – | 71.4%† |
| Instruction following | |||
| LiveBench Language | LiveBench | 87.4% | 74.3% |
| LiveBench Instruction Following | LiveBench | 70.7% | 72.7% |
| IFBench | Artificial Analysis | 75.8% | – |
| Arena Instruction Following | Arena (LMArena) | 1475† | 1419 |
| Human preference | |||
| Arena Text | Arena (LMArena) | 1476† | 1434 |
| EQ-Bench 4 | EQ-Bench | 1295† | – |
| EQ-Bench Creative Writing v3 | EQ-Bench | 1845† | 1671 |
| Arena Creative Writing | Arena (LMArena) | 1448† | 1359 |
| Arena Multi-turn | Arena (LMArena) | 1484† | 1436 |
| Arena Longer Queries | Arena (LMArena) | 1486† | 1443 |
| Multimodal | |||
| Arena Vision | Arena (LMArena) | 1296† | 1270 |
| LiveBench Design | LiveBench | – | 72.4%† |
| MMMU-Pro | Artificial Analysis | 79.9% | 76.3%† |
| MMMU-Pro (Vals) | Vals AI | 88.3% | 83.9%† |
| MortgageTax | Vals AI | 68.8% | 64.9%† |
| GeoBench | GeoBench | – | 38.0% |
| Blueprint-Bench 2 | Epoch AI Benchmarking Hub | 36.2%† | – |
| Furniture Assembly | Epoch AI Benchmarking Hub | 44.2% | – |
| Code2Video-AA | Artificial Analysis | 1000 | 949† |
| Long context | |||
| AA-LCR v1.1 | Artificial Analysis | 84.3% | 82.0%† |
| Arena Document | Arena (LMArena) | 1479† | – |
| MLCR | Artificial Analysis | 28.9% | 22.8%† |
| Composite | |||
| Epoch Capabilities Index | Epoch AI Benchmarking Hub | 160.1† | 152.6 |
| LiveBench 2.0 | LiveBench | – | 54.0%† |
| LiveBench 1.0 | LiveBench | 80.2% | 75.3% |
| AA Intelligence Index v5.0.1 | Artificial Analysis | – | 20.1† |
| AA Intelligence Index v4.3.2 | Artificial Analysis | 38.4 | 33.7† |
| AA Progress Index v1 | Artificial Analysis | 320 | 303.1† |
| Vals RSI Index | Vals AI | 16.5% | – |
| Vals Multimodal Index | Vals AI | 68.1% | – |
| AA Cyber Index | Artificial Analysis | – | 12.7† |
| Safety & honesty | |||
| FORTRESS | Scale AI SEAL | 16.3% | – |
| Metadata | |||
| AA Openness Index | Artificial Analysis | – | 38.9† |
Data as of 2026-11-16. Best configuration of each model; every score links to its source on the model pages.
DuelLab GameBench 2, release gb2-20261111-7b598bb13de4 (https://benchmarks.duellab.org/releases/gb2-20261111-7b598bb13de4/), CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Scores mapped to BenchLeader model names; not endorsed by DuelLab.
What changed
- Qwen3.8 27B: BenchLeader Index 53.2 → 55.1.
- Qwen3.8 27B: Qwen3.8 27B cached input price moved from $0.10/M to $0.085/M (15% down).
- GPT-5.5 (xhigh): 1 new result (GDPval-AA v2.1).
- Qwen3.8 27B: Qwen3.8 27B input price moved from $0.45/M to $0.42/M (7% down).
- GPT-5.5 (xhigh): BenchLeader Index 64.6 → 64.1.
- GPT-5.5 (xhigh): 1 score revision.
- Qwen3.8 27B: BenchLeader Index 54.0 → 53.2.
- GPT-5.5 (xhigh): 1 new result (MirrorCode).
- Qwen3.8 27B: 2 new results (SimpleBench, GeoBench).
- Qwen3.8 27B: 5 score revisions.
Related comparisons
- GPT-5.5 vs Claude Fable 5.2
- GPT-5.5 vs Gemini 4 Carbon
- GPT-5.5 vs Claude Opus 5.5
- GPT-5.5 vs GPT-6 Astra
- Qwen3.8 27B vs Claude Opus 5
- GPT-5.5 vs Claude Fable 5.1
- GPT-5.5 vs GPT-6.1 Sol
- GPT-5.5 vs Claude Fable 5
Rankings: Best LLM overall · Best LLM for agents · Best cheap LLMs
GPT-5.5 vs Qwen3.8 27B: questions
- Is GPT-5.5 better than Qwen3.8 27B?
- GPT-5.5 (xhigh) leads on quality: 64.1 vs 55.1. The BenchLeader Index combines every independent quality benchmark; GPT-5.5 (xhigh) is ahead overall as of 16 Nov 2026, but check the category scores for your use.
- GPT-5.5 vs Qwen3.8 27B: which is cheaper, and will that change?
- Qwen3.8 27B is cheaper today: $1.09 against $11.25 per million tokens, blended at three input tokens per output token. No price change has been announced for either.
- Is GPT-5.5 or Qwen3.8 27B better for coding?
- They are level in coding (61 vs 62 on the category index). On Arena WebDev, the top shared coding benchmark, it is 1512 for GPT-5.5 against 1590 for Qwen3.8 27B.
- Which is faster, GPT-5.5 or Qwen3.8 27B?
- GPT-5.5 streams faster (81 vs 73 output tokens per second, Artificial Analysis 7-day median).
- Is GPT-5.5 or Qwen3.8 27B better for agents and tool use?
- GPT-5.5 scores higher in agentic tasks (62 vs 44 on the category index, where 50 is average).
More questions (4)
- Which is better for long documents?
- GPT-5.5 accepts more context: 1,050,000 against 262,144 tokens for Qwen3.8 27B.
- Which has lower latency?
- Qwen3.8 27B answers first: 0.7 s against 38.3 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-5.5 vs Qwen3.8 27B?
- Across the 12 benchmarks both have, GPT-5.5 leads 9, Qwen3.8 27B leads 3. Every overall index that rates both puts GPT-5.5 ahead.
- When was Qwen3.8 27B released?
- Qwen3.8 27B was released on 14 Aug 2026.