vs
0:3
overall capability wins (0 ties across 3 scored). Qwen3.8 Max 0902 leads overall.
- Pricing3:0
- Specs0:3
Verdict
Qwen3.8 Max 0902 offers ~7.6x larger context (1M vs 131K); Qwen3.8 Max 0902 allows ~16.0x more max output (131K vs 8K); Qwen2.5 72B ships open weights (Apache 2.0).
- Qwen3.8 Max 0902 offers ~7.6x larger context (1M vs 131K)
- Qwen3.8 Max 0902 allows ~16.0x more max output (131K vs 8K)
- Qwen2.5 72B ships open weights (Apache 2.0)
- Qwen2.5 72B is ~8.3x cheaper on a blended token basis than Qwen3.8 Max 0902
Cost for 1M in + 250K out
Illustrative chat workload at primary-provider list prices.
- Qwen2.5 72B
- $0.45
- Qwen3.8 Max 0902
- $3.50
Qwen2.5 72B is about 7.8x cheaper on this workload.
Price per 1M tokens
Full comparison
Organization
Alibaba
Alibaba
Family
Qwen
Qwen
License
Apache 2.0
Proprietary
Open weights
Yes
No
Release
Sep 19, 2024
Sep 2, 2026
Knowledge cutoff
-
-
API / provider
Together / Fireworks (ref.)
Alibaba Cloud
Modalities
text → text
text, image, video → text
Specs
Context window
131K
1M
Max output
8K
131K
Parameters
72B
2.4T (95B act.)
Pricing
Input $/1M
$0.35
$2
Output $/1M
$0.40
$6
Blended $/1M (3∶1)
$0.36
$3
Speed
tok/s
70
-
TTFT (s)
0.35
-
Reasoning
MMLU-Pro
71.1
—
GPQA Diamond
49.1
—
Humanity's Last Exam
3.6
—
AIME 2025
—
—
MATH-500
—
—
Humanity's Last Exam (with tools)
—
—
AA-Omniscience Accuracy
17.5
—
AA-LCR v1.1
—
—
CritPt
—
—
MMMU-Pro
—
—
IFBench
36.9
—
Chartography
—
—
Chartography (With Tools)
—
—
Coding
SWE-bench Verified
—
—
SWE-bench Pro
—
—
SWE-bench Multilingual
—
—
LiveCodeBench
55.5
—
Terminal-Bench 2.1
—
—
Aider Polyglot
—
—
Terminal-Bench 3
—
—
BigCodeBench
45.8
—
SciCode
—
—
CursorBench
—
—
SWE-Rebench
10.1
—
NL2Repo-Bench
—
—
DeepSWE
—
—
WebDev Arena
—
1,681
Terminal-Bench 4.0
—
—
CursorBench 4.0
—
—
FrontierCode v1.1 (Main)
—
—
Arena
LMArena Elo
1,303
—
| Metric | Qwen2.5 72B | Qwen3.8 Max 0902 | Delta |
|---|---|---|---|
| Identity | |||
| Organization | Alibaba | Alibaba | — |
| Family | Qwen | Qwen | — |
| License | Apache 2.0 | Proprietary | — |
| Open weights | Yes | No | — |
| Release | Sep 19, 2024 | Sep 2, 2026 | — |
| Knowledge cutoff | - | - | — |
| API / provider | Together / Fireworks (ref.) | Alibaba Cloud | — |
| Modalities | text → text | text, image, video → text | — |
| Specs | |||
| Context window | 131K | 1M | B +869K |
| Max output | 8K | 131K | B +123K |
| Parameters | 72B | 2.4T (95B act.) | B +2K |
| Pricing | |||
| Input $/1M | $0.35 | $2 | A +$1.65 |
| Output $/1M | $0.40 | $6 | A +$5.60 |
| Blended $/1M (3∶1) | $0.36 | $3 | A +$2.64 |
| Speed | |||
| tok/s | 70 | - | — |
| TTFT (s) | 0.35 | - | — |
| Reasoning | |||
| MMLU-Pro | 71.1 | — | — |
| GPQA Diamond | 49.1 | — | — |
| Humanity's Last Exam | 3.6 | — | — |
| AIME 2025 | — | — | — |
| MATH-500 | — | — | — |
| Humanity's Last Exam (with tools) | — | — | — |
| AA-Omniscience Accuracy | 17.5 | — | — |
| AA-LCR v1.1 | — | — | — |
| CritPt | — | — | — |
| MMMU-Pro | — | — | — |
| IFBench | 36.9 | — | — |
| Chartography | — | — | — |
| Chartography (With Tools) | — | — | — |
| Coding | |||
| SWE-bench Verified | — | — | — |
| SWE-bench Pro | — | — | — |
| SWE-bench Multilingual | — | — | — |
| LiveCodeBench | 55.5 | — | — |
| Terminal-Bench 2.1 | — | — | — |
| Aider Polyglot | — | — | — |
| Terminal-Bench 3 | — | — | — |
| BigCodeBench | 45.8 | — | — |
| SciCode | — | — | — |
| CursorBench | — | — | — |
| SWE-Rebench | 10.1 | — | — |
| NL2Repo-Bench | — | — | — |
| DeepSWE | — | — | — |
| WebDev Arena | — | 1,681 | — |
| Terminal-Bench 4.0 | — | — | — |
| CursorBench 4.0 | — | — | — |
| FrontierCode v1.1 (Main) | — | — | — |
| Arena | |||
| LMArena Elo | 1,303 | — | — |
Qwen2.5 72B
- Context
- 131K / 8K out
- Parameters
- 72B
- Price
- $0.35 / $0.40
- Speed
- 70 tok/s · 0.35s TTFT
- Modalities
- text → text
- License
- Apache 2.0
Widely deployed prior-gen open 72B - still a common self-host baseline.
Qwen3.8 Max 0902
- Context
- 1M / 131K out
- Parameters
- 2.4T (95B act.)
- Price
- $2 / $6
- Speed
- —
- Modalities
- text, image, video → text
- License
- Proprietary
September 2, 2026 hosted post-training snapshot of Qwen3.8-Max, further trained for coding and workplace/agent workflows. Same 2.4T MoE / 95B active architecture, 1M multimodal context, and $2/$6 QwenCloud pricing as the August 3 Max release. Arena Code/WebDev listed qwen3.8-max-0902 at 1688 (preliminary) on the Sep 2, 2026 snapshot.
Benchmark charts
Winner bars are emphasized. Per-benchmark deltas sit above each chart.
These two models have no overlapping published benchmarks in our dataset. Compare specs and pricing instead.