Agnes 2.5 Pro Beta vs Qwen3.8 Max
Full profiles: Agnes 2.5 Pro Beta (Sapiens AI) and Qwen3.8 Max (Alibaba): every benchmark, price and rank.
Updated · Data from 12 sources
As of 16 Nov 2026, there are not enough shared benchmarks for a verdict, and Agnes 2.5 Pro Beta is 20× cheaper at $0.150 vs $3.00 per 1M tokens blended. Choose Agnes 2.5 Pro Beta for agents & tools (64 vs 59) and long context (66 vs 65); choose Qwen3.8 Max for multimodal (66 vs 58).
Key facts (as of 16 Nov 2026)
- Agnes 2.5 Pro Beta costs $0.150 per 1M tokens blended (list price, 16 Nov 2026).
- Qwen3.8 Max costs $3.00 per 1M tokens blended (list price, 16 Nov 2026).
- Of the 11 benchmarks both models have, Qwen3.8 Max leads 10 (BenchLeader, 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 |
|---|---|---|---|
| Agents & tool use | Agnes 2.5 Pro Beta | 64 vs 59 category score; τ³-Banking (AA) 35.7% vs 51.3% | Medium |
| Long documents | Agnes 2.5 Pro Beta | 66 vs 65 category score; AA-LCR v1.1 83.0% vs 80.3% | Medium |
| Images and other media | Qwen3.8 Max | 66 vs 58 category score; MMMU-Pro 82.8% vs 75.4% | High |
| Chat quality (human preference) | – | Only Qwen3.8 Max has an Arena rating (68 category score) | Not comparable |
| High volume on a budget | Neither flagship | A smaller model is usually the better fit: Cheapest good LLMs, Fastest LLMs | – |
Key differences
- Not enough shared benchmarks for a verdict.
- Agnes 2.5 Pro Beta is stronger in agents & tools, long context.
- Qwen3.8 Max (max) is stronger in multimodal.
- Agnes 2.5 Pro Beta is 20× cheaper ($0.150 vs $3.00 per 1M blended).
At a glance
| Spec | Agnes 2.5 Pro Beta | Qwen3.8 Max |
|---|---|---|
| Developer | Sapiens AI | Alibaba |
| Released | 26 Aug 2026 | 3 Aug 2026 |
| Context window (tokens) | 1,000,000 | 1,000,000 |
| Compared at | Agnes 2.5 Pro Beta | Qwen3.8 Max (max) |
| Open weights | No | No |
| Price in / out / cached input per 1M | – | $2.00 / $6.00 / $0.250 |
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 | Agnes 2.5 Pro Beta | Qwen3.8 Max (max) | Leader |
|---|---|---|---|
| Reasoning | – | 51.5 | not comparable |
| Coding | – | 61.6 | not comparable |
| Agents & tools | 64.4 | 58.5 | Agnes 2.5 Pro Beta |
| Maths | – | 65.2 | not comparable |
| Knowledge | – | 58.0 | not comparable |
| Human preference | – | 67.5 | not comparable |
| Multimodal | 57.7 | 65.7 | Qwen3.8 Max |
| Long context | 66.2 | 64.8 | Agnes 2.5 Pro Beta |
| Composite | – | 45.0 | 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: Reasoning, Coding, Maths, Knowledge, Instruction following, Human preference.
Where sources disagree
Across the 11 benchmarks both have, Agnes 2.5 Pro Beta leads 1, Qwen3.8 Max leads 10.
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 | Agnes 2.5 Pro Beta | Qwen3.8 Max (max) |
|---|---|---|
| Blended price per 1M tokens | – | $3.00 |
| Input / output per 1M | – / – | $2.00 / $6.00 |
| Cached input per 1M | – | $0.250 |
| Price per task: chat reply | – | $0.0026 |
| Price per task: summarize a 30-page report | – | $0.028 |
| Price per task: code edit | – | $0.021 |
| Price per task: agentic coding session | – | $0.144 |
| Price per task: structured extraction | – | $0.0052 |
| Agentic coding session, input mostly cached | – | $0.065 |
| Cost per benchmark task (AA Index) | – | $10.02 |
| Output tokens per task (AA Index) | – | 171k |
| Of which reasoning | – | 120k |
| Time per task (AA Index) | – | 48 min |
| Output speed | – | 59 tok/s |
| Time to first answer | – | 119.6 s |
| Time to a complete answer, last measured 13 Nov 2026 | – | 294.1 s |
Each model is compared at its best-ranked reasoning-effort setting. Agnes 2.5 Pro Beta. Qwen3.8 Max (max) scores 58.7 and takes 120 s to its first answer token; Qwen3.8 Max (xhigh) scores 56.6 in 3 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
137 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 | Agnes 2.5 Pro Beta | Qwen3.8 Max (max) |
|---|---|---|---|
| BenchLeader Index | BenchLeader | – | 58.7 |
| Agents & tools score | BenchLeader | 64.4 | 58.5 |
| Long context score | BenchLeader | 66.2 | 64.8 |
| Multimodal score | BenchLeader | 57.7 | 65.7 |
| Coding score | BenchLeader | – | 61.6 |
| Composite score | BenchLeader | – | 45.0 |
| Human preference score | BenchLeader | – | 67.5 |
| Knowledge score | BenchLeader | – | 58.0 |
| Maths score | BenchLeader | – | 65.2 |
| Reasoning score | BenchLeader | – | 51.5 |
| Blended price $/M | BenchLeader | – | $3.00 |
| Output speed | BenchLeader | – | 59 tok/s |
| Time to first answer | BenchLeader | – | 119.6 s |
| Context window | BenchLeader | 1,000,000 tokens | 1,000,000 tokens |
| Reasoning | |||
| GPQA Diamond | Epoch AI Benchmarking Hub | – | 92.7%† |
| Arena Hard Prompts | Arena (LMArena) | – | 1507 |
| LiveBench 2.0 Reasoning | LiveBench | – | 88.2% |
| LiveBench 1.0 Reasoning | LiveBench | – | 88.2% |
| GPQA Diamond (AA) | Artificial Analysis | 90.5% | 92.8% |
| Humanity's Last Exam (AA) | Artificial Analysis | 37.5% | 43.1% |
| HLE-Diamond (AA) | Artificial Analysis | – | 24.0% |
| GPQA Diamond (Vals) | Vals AI | – | 93.7% |
| CritPt | Artificial Analysis | 15.7% | 20.0% |
| CritPt v2 hard subset (AA) | Artificial Analysis | – | 29.5% |
| MysteryMechanism | Vals AI | – | 23.9% |
| Arena Expert | Arena (LMArena) | – | 1514 |
| Chess Puzzles | Epoch AI Benchmarking Hub | – | 46.0%† |
| Mystery Game Puzzles | Epoch AI Benchmarking Hub | – | 38.0%† |
| LMCA | Epoch AI Benchmarking Hub | – | 46.2%† |
| DTBench | Epoch AI Benchmarking Hub | – | 92.0%† |
| Coding | |||
| SciCode | SciCode | – | 53.2% |
| Arena Coding | Arena (LMArena) | – | 1523 |
| Arena WebDev | Arena (LMArena) | – | 1676 |
| LiveBench 2.0 Coding | LiveBench | – | 60.9% |
| LiveBench 1.0 Coding | LiveBench | – | 72.9% |
| SciCode (AA) | Artificial Analysis | 48.8% | 53.2% |
| LiveCodeBench | Vals AI | – | 87.8% |
| IOI | Vals AI | – | 68.9% |
| SWE-bench (Vals) | Vals AI | – | 85.6% |
Show all 137 benchmarks
| Metric | Publisher | Agnes 2.5 Pro Beta | Qwen3.8 Max (max) |
|---|---|---|---|
| Coding | |||
| Code Migration | Vals AI | – | 24.0% |
| ProgramBench | Vals AI | – | 0.0% |
| SWE-bench Multimodal | Vals AI | – | 42.5% |
| Vibe Code Bench 1-100 | Vals AI | – | 12.8% |
| Vibe Code Bench v1.1 | Vals AI | – | 64.7% |
| SWE Atlas v1.1 | Scale AI SEAL | – | 64.4% |
| SWE Atlas v1.1: Codebase QnA | Scale AI SEAL | – | 81.5% |
| SWE Atlas v1.1: Test Writing | Scale AI SEAL | – | 58.9% |
| SWE Atlas v1.1: Refactoring | Scale AI SEAL | – | 52.9% |
| MirrorCode | Epoch AI Benchmarking Hub | – | 52.2%† |
| DeepSWE v1.1 | Epoch AI Benchmarking Hub | – | 57.5%† |
| ALE-Bench | Epoch AI Benchmarking Hub | – | 1679.1† |
| FrontierSWE | Epoch AI Benchmarking Hub | – | 17.8%† |
| Terminal-Bench 4.0 (AA) | Artificial Analysis | – | 38.9% |
| Terminal-Bench 2.1 (AA) | Artificial Analysis | 69.7% | 88.8% |
| Startrise Frontend | Startrise | – | 71.3% |
| GameBench 2 Overall | DuelLab GameBench 2 | – | 54.2† |
| GameBench 2 by reasoning setting | DuelLab GameBench 2 | – | 69.6† |
| Design Arena Website | Design Arena | – | 1269 |
| FrontierCode 1.1 (AA) | Artificial Analysis | – | 43.2% |
| FrontierCode 2.0 (AA) | Artificial Analysis | – | 23.4% |
| Agents & tools | |||
| Terminal-Bench | Terminal-Bench | – | 27.0% |
| APEX-Agents | Mercor | – | 63.3% |
| Arena Agent | Arena (LMArena) | – | 0.1 |
| LiveBench Long Agentic | LiveBench | – | 26.6% |
| LiveBench Browser | LiveBench | – | 49.4% |
| LiveBench Agentic Coding | LiveBench | – | 64.7% |
| Terminal-Bench 2.1 (Vals) | Vals AI | – | 67.4% |
| τ³-bench Banking | τ²-bench | – | 55.1%† |
| GDPval-AA v2.1, normalised score | Artificial Analysis | 41.5% | 58.3% |
| GDPval-AA v2.1 | Artificial Analysis | 1335 | 1674 |
| τ³-Banking (AA) | Artificial Analysis | 35.7% | 51.3% |
| ITBench SRE (AA) | Artificial Analysis | – | 40.3% |
| Analyst Agent (AA) | Artificial Analysis | – | 45.0% |
| APEX-Agents (AA) | Artificial Analysis | – | 42.4% |
| CyberBench | Vals AI | – | 28.6% |
| Finance Agent v2 | Vals AI | – | 50.6% |
| Harvey's Legal Agent Benchmark v1 | Vals AI | – | 10.4% |
| Legal Research Bench | Vals AI | – | 47.6% |
| SkillsBench | Vals AI | – | 42.0% |
| Tax Agent Bench | Vals AI | – | 32.1% |
| Terminal-Bench 4.0 (Vals) | Vals AI | – | 34.3% |
| Terminal-Bench Science | Vals AI | – | 12.9% |
| Arena Search | Arena (LMArena) | – | 1227 |
| Arena Agent Task Completion | Arena (LMArena) | – | 4.5 |
| AutomationBench | Artificial Analysis | – | 56.2% |
| GDP.pdf (AA) | Artificial Analysis | – | 22.8% |
| EnterpriseOps-Gym | Artificial Analysis | – | 47.6% |
| AA-Briefcase v1.1 | Artificial Analysis | – | 1604 |
| GDPx | GDPx by SemiAnalysis | – | 3.5% |
| Design Arena Agentic | Design Arena | – | 1264 |
| Terminal-Bench-Science 0.1 (AA) | Artificial Analysis | – | 11.9% |
| MKA (AA) | Artificial Analysis | – | 37.4% |
| FACTS Search | – | 85.0% | |
| Maths | |||
| FrontierMath Tiers 1–3 | Epoch AI Benchmarking Hub | – | 74.7%† |
| FrontierMath Tier 4 | Epoch AI Benchmarking Hub | – | 46.3%† |
| OTIS Mock AIME | Epoch AI Benchmarking Hub | – | 100.0%† |
| ProofBench | Vals AI | – | 58.0% |
| LiveBench Mathematics | LiveBench | – | 91.3% |
| Arena Maths | Arena (LMArena) | – | 1505 |
| Knowledge | |||
| SimpleQA Verified | Epoch AI Benchmarking Hub | – | 47.3%† |
| LiveBench 2.0 Data Analysis | LiveBench | – | 78.4% |
| LiveBench Research | LiveBench | – | 72.8% |
| LiveBench 1.0 Data Analysis | LiveBench | – | 78.4% |
| AA-Omniscience v1.1 | Artificial Analysis | – | 10.5 |
| MMLU-Pro | Vals AI | – | 88.6% |
| LegalBench | Vals AI | – | 83.6% |
| CorpFin | Vals AI | – | 65.8% |
| TaxEval | Vals AI | – | 75.5% |
| Excel Modeling Benchmark | Vals AI | – | 60.1% |
| Excel Modeling Benchmark v1.1 | Vals AI | – | 66.0% |
| MedCode | Vals AI | – | 40.7% |
| MedScribe | Vals AI | – | 85.0% |
| Public Benefits Bench v1.1 | Vals AI | – | 67.1% |
| Access to Justice Benchmark | Vals AI | – | 76.4% |
| Financial Coaching Bench | Vals AI | – | 75.6% |
| Immigration Bench | Vals AI | – | 70.1% |
| GDP.pdf | Epoch AI Benchmarking Hub | – | 23.2%† |
| AA-Omniscience v1.1: accuracy | Artificial Analysis | – | 29.5% |
| AA-Omniscience v1.1: non-hallucination | Artificial Analysis | – | 72.9% |
| FACTS Parametric | – | 30.8% | |
| Instruction following | |||
| LiveBench Language | LiveBench | – | 79.7% |
| LiveBench Instruction Following | LiveBench | – | 74.1% |
| Arena Instruction Following | Arena (LMArena) | – | 1478 |
| Human preference | |||
| Arena Text | Arena (LMArena) | – | 1484 |
| Arena Creative Writing | Arena (LMArena) | – | 1472 |
| Arena Multi-turn | Arena (LMArena) | – | 1495 |
| Arena Longer Queries | Arena (LMArena) | – | 1496 |
| Multimodal | |||
| Arena Vision | Arena (LMArena) | – | 1312 |
| LiveBench Design | LiveBench | – | 76.6% |
| MMMU-Pro | Artificial Analysis | 75.4% | 82.8% |
| MMMU-Pro (Vals) | Vals AI | – | 88.0% |
| MortgageTax | Vals AI | – | 64.0% |
| SAGE | Vals AI | – | 81.8% |
| Humanity's Sixth Sense | Scale AI SEAL | – | 33.0% |
| Furniture Assembly | Epoch AI Benchmarking Hub | – | 20.0%† |
| VIALS-AA | Handshake AI | – | 31.7% |
| Code2Video-AA | Artificial Analysis | – | 1021 |
| AA-3D | Artificial Analysis | – | 954 |
| FACTS Multimodal | – | 51.5% | |
| Long context | |||
| AA-LCR v1.1 | Artificial Analysis | 83.0% | 80.3% |
| MLCR | Artificial Analysis | – | 26.1% |
| FACTS Grounding | – | 73.8% | |
| Composite | |||
| Epoch Capabilities Index | Epoch AI Benchmarking Hub | – | 157.6 |
| LiveBench 2.0 | LiveBench | – | 60.9% |
| LiveBench 1.0 | LiveBench | – | 78.5% |
| AA Intelligence Index v5.0.1 | Artificial Analysis | – | 33.6 |
| AA Intelligence Index v4.3.2 | Artificial Analysis | 35.2 | 45.4 |
| AA Progress Index v1 | Artificial Analysis | 299.1 | 323.7 |
| Vals RSI Index | Vals AI | – | 21.8% |
| Vals Multimodal Index | Vals AI | – | 65.4% |
| Safety & honesty | |||
| DistressBench | Scale AI SEAL | – | 77.6% |
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 Max (max): 1 score revision.
- Qwen3.8 Max (max): 1 score revision.
- Qwen3.8 Max (max): 6 score revisions.
- Qwen3.8 Max (max): 1 score revision.
- Qwen3.8 Max (max): 1 new result (FACTS Search); 1 score revision.
- Qwen3.8 Max (max): 1 score revision; cached input price moved from $0.22/M to $0.25/M (13% up).
- Qwen3.8 Max (max): cached input price moved from $0.25/M to $0.22/M (11% down).
- Qwen3.8 Max (max): 1 score revision; cached input price moved from $0.22/M to $0.25/M (13% up).
- Qwen3.8 Max (max): cached input price moved from $0.25/M to $0.22/M (11% down).
- Qwen3.8 Max (max): 1 score revision.
Related comparisons
- Qwen3.8 Max vs GPT-6 Astra
- Agnes 2.5 Pro Beta vs Claude Opus 5
- Agnes 2.5 Pro Beta vs GPT-6 Sol
- Agnes 2.5 Pro Beta vs Muse Wonder Full
- Agnes 2.5 Pro Beta vs Muse Spark 1.3
- Agnes 2.5 Pro Beta vs Kimi K3
- Qwen3.8 Max vs GPT-6.1 Sol
- Qwen3.8 Max vs Gemini 4 Argon
Rankings: Best LLM overall · Best LLM for agents · Best cheap LLMs
Agnes 2.5 Pro Beta vs Qwen3.8 Max: questions
- Agnes 2.5 Pro Beta vs Qwen3.8 Max: which is cheaper, and will that change?
- Agnes 2.5 Pro Beta is cheaper today: $0.150 against $3.00 per million tokens, blended at three input tokens per output token. No price change has been announced for either.
- Is Agnes 2.5 Pro Beta or Qwen3.8 Max better for agents and tool use?
- Agnes 2.5 Pro Beta scores higher in agentic tasks (64 vs 59 on the category index, where 50 is average). On τ³-Banking (AA), the top shared agents & tools benchmark, it is 35.7% for Agnes 2.5 Pro Beta against 51.3% for Qwen3.8 Max.
- Which is better for long documents?
- Agnes 2.5 Pro Beta scores higher on long-context benchmarks (66 vs 65). Both accept 1,000,000 tokens of context.
- Do independent benchmarks agree on Agnes 2.5 Pro Beta vs Qwen3.8 Max?
- Across the 11 benchmarks both have, Agnes 2.5 Pro Beta leads 1, Qwen3.8 Max leads 10.
- When was Agnes 2.5 Pro Beta released?
- Agnes 2.5 Pro Beta was released on 26 Aug 2026.