Qwen3.8 2.4T A95B vs Qwen4 Max
Full profiles: Qwen3.8 2.4T A95B (Alibaba) and Qwen4 Max (Alibaba): every benchmark, price and rank.
Updated · Data from 8 sources
As of 19 Nov 2026, there are not enough shared benchmarks for a verdict, and Qwen4 Max is 1.3× cheaper at $2.25 vs $3.00 per 1M tokens blended. Qwen4 Max leads in every category we measure both in. Qwen4 Max streams faster.
Qwen4 Max was released on 16 Nov 2026 and has no BenchLeader Index yet (it needs five independent results across two categories). So this page gives no quality verdict: it compares what is published so far for both, price, context window, speed and the benchmarks both have results on. First results are tracked on Qwen4 Max's page.
Key facts (as of 19 Nov 2026)
- Qwen3.8 2.4T A95B costs $3.00 per 1M tokens blended (list price, 19 Nov 2026).
- Qwen4 Max costs $2.25 per 1M tokens blended (list price, 19 Nov 2026).
- Qwen4 Max streams faster than Qwen3.8 2.4T A95B (Artificial Analysis, 7-day median, 19 Nov 2026).
- Of the 11 benchmarks both models have, Qwen4 Max leads 10 (BenchLeader, 19 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 |
|---|---|---|---|
| Reasoning | Qwen4 Max | 59 vs 46 category score; HLE-Diamond (AA) 44.9% vs 20.1% | High |
| Knowledge | Qwen4 Max | 59 vs 49 category score; AA-Omniscience v1.1 27.7 vs 7.8 | High |
| Lowest cost | Qwen4 Max | $2.25 vs $3.00 per 1M blended; agentic coding session $0.108 vs $0.144 | – |
| Fastest first answer | Either | 111.5 s (Qwen3.8 2.4T A95B) vs 139.2 s (Qwen4 Max) 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
- Not enough shared benchmarks for a verdict.
- Qwen4 Max is stronger in knowledge, reasoning.
- Qwen4 Max is 1.3× cheaper ($2.25 vs $3.00 per 1M blended).
- Qwen4 Max streams 1.4× faster (59 vs 41 tokens per second).
At a glance
| Spec | Qwen3.8 2.4T A95B | Qwen4 Max |
|---|---|---|
| Developer | Alibaba | Alibaba |
| Released | 12 Aug 2026 | 16 Nov 2026 |
| Context window (tokens) | 1,000,000 | 991,808 |
| Compared at | Qwen3.8 2.4T A95B | Qwen4 Max |
| Open weights | Yes | No |
| Price in / out / cached input per 1M | $2.00 / $6.00 / $0.250 | $1.50 / $4.50 / $0.071 |
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 | Qwen3.8 2.4T A95B | Qwen4 Max | Leader |
|---|---|---|---|
| Reasoning | 45.5 | 58.5 | Qwen4 Max |
| Coding | – | 55.1 | not comparable |
| Agents & tools | 50.3 | – | not comparable |
| Knowledge | 48.8 | 59.0 | Qwen4 Max |
| Long context | 64.8 | – | 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.
Dashed: that model has no result on the axis it crosses; the line only joins its neighbours.
Not drawn, no result for these models: Maths, Instruction following, Multimodal, Human preference.
Where sources disagree
Across the 11 benchmarks both have, Qwen3.8 2.4T A95B leads 1, Qwen4 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 | Qwen3.8 2.4T A95B | Qwen4 Max |
|---|---|---|
| Blended price per 1M tokens | $3.00 | $2.25 |
| Input / output per 1M | $2.00 / $6.00 | $1.50 / $4.50 |
| Cached input per 1M | $0.250 | $0.071 |
| Price per task: chat reply | $0.0026 | $0.0020 |
| Price per task: summarize a 30-page report | $0.028 | $0.021 |
| Price per task: code edit | $0.021 | $0.016 |
| Price per task: agentic coding session | $0.144 | $0.108 |
| Price per task: structured extraction | $0.0052 | $0.0039 |
| Agentic coding session, input mostly cached | $0.065 | $0.044 |
| Cost per benchmark task (AA Index) | – | $3.27 |
| Output tokens per task (AA Index) | – | 105k |
| Of which reasoning | – | 69k |
| Time per task (AA Index) | – | 30 min |
| Output speed | 41 tok/s | 59 tok/s |
| Time to first answer | 111.5 s | 139.2 s |
| Time to a complete answer, last measured 13 Nov 2026 | 248.5 s | – |
Each model is compared at its best-ranked reasoning-effort setting. Qwen3.8 2.4T A95B scores 51.6 and takes 111 s to its first answer token. Qwen4 Max and takes 139 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
61 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 | Qwen3.8 2.4T A95B | Qwen4 Max |
|---|---|---|---|
| BenchLeader Index | BenchLeader | 51.6 | – |
| Agents & tools score | BenchLeader | 50.3 | – |
| Knowledge score | BenchLeader | 48.8 | 59.0 |
| Long context score | BenchLeader | 64.8 | – |
| Reasoning score | BenchLeader | 45.5 | 58.5 |
| Coding score | BenchLeader | – | 55.1 |
| Blended price $/M | BenchLeader | $3.00 | $2.25 |
| Output speed | BenchLeader | 41 tok/s | 59 tok/s |
| Time to first answer | BenchLeader | 111.5 s | 139.2 s |
| Context window | BenchLeader | 1,000,000 tokens | 991,808 tokens |
| Reasoning | |||
| LiveBench 2.0 Reasoning | LiveBench | – | 92.2%† |
| GPQA Diamond (AA) | Artificial Analysis | 93.5% | – |
| Humanity's Last Exam (AA) | Artificial Analysis | 42.5% | – |
| HLE-Diamond (AA) | Artificial Analysis | 20.1% | 44.9% |
| CritPt | Artificial Analysis | 20.0% | – |
| CritPt v2 hard subset (AA) | Artificial Analysis | – | 43.0% |
| Coding | |||
| Arena WebDev | Arena (LMArena) | – | 1750† |
| LiveBench 2.0 Coding | LiveBench | – | 69.8%† |
| SciCode (AA) | Artificial Analysis | 54.0% | – |
| Vibe Code Bench v1.1 | Vals AI | – | 86.7%† |
| Terminal-Bench 4.0 (AA) | Artificial Analysis | 11.1% | 48.5% |
| Terminal-Bench 2.1 (AA) | Artificial Analysis | 82.0% | – |
| GameBench 2 Overall | DuelLab GameBench 2 | 52.8 | – |
| GameBench 2 by reasoning setting | DuelLab GameBench 2 | 49.4† | – |
| FrontierCode 2.0 (AA) | Artificial Analysis | – | 26.4% |
| Agents & tools | |||
| Arena Agent | Arena (LMArena) | – | 7.3† |
| LiveBench Long Agentic | LiveBench | – | 46.6%† |
| LiveBench Browser | LiveBench | – | 80.7%† |
| MCP Atlas | Scale AI SEAL | 84.5%† | – |
| GDPval-AA v2.1, normalised score | Artificial Analysis | 55.4% | – |
| GDPval-AA v2.1 | Artificial Analysis | 1614 | 1809 |
| τ³-Banking (AA) | Artificial Analysis | 49.1% | – |
| Analyst Agent (AA) | Artificial Analysis | – | 46.3% |
| APEX-Agents (AA) | Artificial Analysis | – | 59.1% |
| Finance Agent v2.1 | Vals AI | – | 66.7%† |
Show all 61 benchmarks
| Metric | Publisher | Qwen3.8 2.4T A95B | Qwen4 Max |
|---|---|---|---|
| Agents & tools | |||
| Harvey's Legal Agent Benchmark v1.1 | Vals AI | – | 21.3%† |
| Legal Research Bench | Vals AI | – | 41.4%† |
| Tax Agent Bench | Vals AI | – | 40.4%† |
| Terminal-Bench 4.0 (Vals) | Vals AI | – | 48.5%† |
| Arena Agent Task Completion | Arena (LMArena) | – | 8.8† |
| AutomationBench | Artificial Analysis | 57.3% | 72.6% |
| GDP.pdf (AA) | Artificial Analysis | 15.6% | 26.2% |
| Harvey LAB-AA v1.1 | Harvey | – | 9.4% |
| EnterpriseOps-Gym | Artificial Analysis | 47.4% | 48.5% |
| AA-Briefcase v1.1 | Artificial Analysis | 1430 | 1686 |
| Design Arena Agentic | Design Arena | – | 1356 |
| Terminal-Bench-Science 0.1 (AA) | Artificial Analysis | – | 23.3% |
| CWE-Bench (AA) | Artificial Analysis | 28.3% | – |
| DeepSec-Bench (AA) | Artificial Analysis | 9.5% | – |
| CyberGym end-to-end (AA) | Artificial Analysis | 0.0% | – |
| Knowledge | |||
| LiveBench 2.0 Data Analysis | LiveBench | – | 83.4%† |
| LiveBench Research | LiveBench | – | 84.7%† |
| AA-Omniscience v1.1 | Artificial Analysis | 7.8 | 27.7 |
| Excel Modeling Benchmark v1.1 | Vals AI | – | 70.8%† |
| AA-Omniscience v1.1: accuracy | Artificial Analysis | 31.4% | 43.7% |
| AA-Omniscience v1.1: non-hallucination | Artificial Analysis | 65.5% | 71.6% |
| Human preference | |||
| EQ-Bench Creative Writing v3 | EQ-Bench | 1844 | – |
| Multimodal | |||
| LiveBench Design | LiveBench | – | 79.1%† |
| Long context | |||
| AA-LCR v1.1 | Artificial Analysis | 80.3% | – |
| MLCR | Artificial Analysis | 30.6% | 19.4% |
| Composite | |||
| LiveBench 2.0 | LiveBench | – | 73.1%† |
| AA Intelligence Index v5.0.1 | Artificial Analysis | – | 42.9 |
| AA Intelligence Index v4.3.2 | Artificial Analysis | 40.0 | – |
| AA Progress Index v1 | Artificial Analysis | 317.8 | – |
| Vals Index | Vals AI | – | 63.2† |
| Vals Finance Index | Vals AI | – | 62.0%† |
| Vals Legal Index | Vals AI | – | 70.7%† |
| AA Cyber Index | Artificial Analysis | 12.6 | – |
| Safety & honesty | |||
| FORTRESS | Scale AI SEAL | 21.6%† | – |
| DistressBench | Scale AI SEAL | 77.5% | – |
| Metadata | |||
| AA Openness Index | Artificial Analysis | 27.8 | – |
Data as of 2026-11-19. Best configuration of each model; every score links to its source on the model pages.
DuelLab GameBench 2, release gb2-20261114-ca113f475f5e (https://benchmarks.duellab.org/releases/gb2-20261114-ca113f475f5e/), 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 2.4T A95B: BenchLeader Index 51.0 → 51.6.
- Qwen3.8 2.4T A95B: BenchLeader Index 51.5 → 51.0.
- Qwen3.8 2.4T A95B: BenchLeader Index 52.0 → 51.5.
- Qwen3.8 2.4T A95B: 1 score revision.
- Qwen3.8 2.4T A95B: BenchLeader Index 53.0 → 52.0.
- Qwen3.8 2.4T A95B: 1 new result (HLE-Diamond (AA)).
- Qwen3.8 2.4T A95B: BenchLeader Index 56.1 → 53.0.
- Qwen3.8 2.4T A95B: 1 new result (AA Progress Index v1); 2 score revisions.
- Qwen3.8 2.4T A95B: 5 new results (AA Cyber Index, CWE-Bench (AA), DeepSec-Bench (AA)…).
- Qwen3.8 2.4T A95B: 1 score revision.
Related comparisons
- Qwen3.8 2.4T A95B vs GPT-5.6 Sol
- Qwen4 Max vs Qwen3.8 Max
- Qwen4 Max vs Inkling 2 Small
- Qwen3.8 2.4T A95B vs Gemini 4 Argon
- Qwen3.8 2.4T A95B vs Claude Sonnet 5.5
- Qwen3.8 2.4T A95B vs Muse Spark 1.3
- Qwen3.8 2.4T A95B vs Gemini 3 Pro
- Qwen3.8 2.4T A95B vs Qwen3.8 Max
Rankings: Best LLM overall · Best LLM for agents · Best cheap LLMs
Qwen3.8 2.4T A95B vs Qwen4 Max: questions
- Qwen3.8 2.4T A95B vs Qwen4 Max: which is cheaper, and will that change?
- Qwen4 Max is cheaper today: $2.25 against $3.00 per million tokens, blended at three input tokens per output token. No price change has been announced for either.
- Which is faster, Qwen3.8 2.4T A95B or Qwen4 Max?
- Qwen4 Max streams faster (59 vs 41 output tokens per second, Artificial Analysis 7-day median).
- Which is better for long documents?
- Both take about 1M tokens: Qwen3.8 2.4T A95B 1,000,000, Qwen4 Max 991,808.
- Which has lower latency?
- Qwen3.8 2.4T A95B answers first: 111.5 s against 139.2 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 Qwen3.8 2.4T A95B vs Qwen4 Max?
- Across the 11 benchmarks both have, Qwen3.8 2.4T A95B leads 1, Qwen4 Max leads 10.
More questions (1)
- When was Qwen4 Max released?
- Qwen4 Max was released on 16 Nov 2026.