Gemma 4 31B
- Input
- $0.165/M tok
- Output
- $0.40/M tok
- Blended
- $0.224/M tok · 3:1
- Speed
- 35tok/s
- Context
- 256Ktokens
- Cost per task
- ~$0.17/taskest.
Open weights; no paid Google API tier, price is the Artificial Analysis third-party host price.
Rankings
Coding
Mixed#73 / 137rank by quality42.1±6.6quality / 1003/7 benchmarks measuredWriting & chat
Mixed#51 / 162rank by quality61.1±3.6quality / 1005/5 benchmarks measuredVision
Verified#52 / 98rank by quality54.4±8.2quality / 1001/6 benchmarks measuredHard reasoning
Mixed#91 / 218rank by quality50.2±3.6quality / 1004/4 benchmarks measuredAgents
Verified#100 / 187rank by quality29.5±4.3quality / 1004/7 benchmarks measured
Self-hostingEstimated
How we estimate →What it would cost to run these open weights yourself on rented GPUs, compared with the API price above.
- Hardware
- 1× B200 180GB
- FP8 weights
- Throughput
- ~5,049 tok/s
- many requests batched
- Price
- $0.16–1.41 /M tok
- busy → light use
- On your own machine
- RTX 5090 32GB (int4, ~82 tok/s single-stream)
- single consumer GPU or Mac
Assumptions (5)
- FP8 weights (30.7B params) + 25% KV-cache headroom ≈ 38 GB
- 1× B200 180GB at $5.98–$14.24/GPU-hour on-demand (2026-09-24)
- ~5,049 output tok/s aggregate at batch 144 (bandwidth-bound)
- Blended 3:1 input:output; prefill ~34,629 tok/s
- Low = 75% utilization at the low GPU price; high = 20% at the high price
Benchmark scores
~ italic, dashed = no published score yet, estimated from related benchmarks.
Coding
SWE-bench VerifiedEstimated
~#13 of 21 (estimated)~73.7%
- #1Claude Opus 596.0%
- #2Claude Sonnet 585.2%
- #12Flint74.2%
- ~#13Gemma 4 31BEstimated73.7%
- #13Claude Haiku 4.573.3%
No published score — estimated from related benchmarks
SWE-bench Pro (Public)Estimated
~#20 of 24 (estimated)~53.0%
- #1Claude Fable 5.293.0%
- #2Claude Opus 5.589.9%
- #18Gemini 3.5 Flash-Lite54.2%
- ~#20Gemma 4 31BEstimated53.0%
- #20Nemotron 3.5 Super VL51.8%
No published score — estimated from related benchmarks
DeepSWE v1.1Estimated
~#38 of 46 (estimated)~46.2%
- #1Gemini 4 Argon78.8%
- #2DeepSeek V4.1 Flash 101777.0%
- #37Qwen3.8 Max (0902)51.0%
- ~#38Gemma 4 31BEstimated46.2%
- #38Ripple-65B Preview44.3%
No published score — estimated from related benchmarks
Terminal-Bench 4.0Verified
#63 of 1010.0%
- #1Kimi K3.1 Pro67.2%
- #2Claude Sonnet 5.563.6%
- #63Devstral Small 20.0%
- #63Gemma 4 31B0.0%
- #63gpt-oss-120b0.0%
artificialanalysis.ai · 2026-09-24
LiveCodeBenchLab-reported
#11 of 2780.0%
- #1Qwen3.8-Omni-Flash92.6%
- #2Qwen3.8 Flash-Next91.9%
- #10Ternary Bonsai 27B82.8%
- #11Gemma 4 31B80.0%
- #12gpt-oss-20b77.7%
huggingface.co · 2026-04-02
SciCodeVerified
#60 of 11845.5%
- #1Claude Fable 5.267.7%
- #2Claude Opus 5.566.9%
- #60Apodex 1.145.5%
- #60Gemma 4 31B45.5%
- #62Ring-2.6-1T45.0%
artificialanalysis.ai · 2026-09-24
AA Coding Agent IndexEstimated
~#24 of 26 (estimated)~32.3
- #1Claude Opus 5.561.7
- #2Claude Opus 559.7
- #23Gemini 3.8 Flash33.3
- ~#24Gemma 4 31BEstimated32.3
- #24GPT-6 Luna31.1
No published score — estimated from related benchmarks
Writing & chat
IFBenchVerified
#23 of 13975.6%
- #1MiniMax-M382.9%
- #2Qwen3.8 2.4T A95B82.8%
- #22Qwen3.5 122B A10B75.7%
- #23Gemma 4 31B75.6%
- #24GPT-5.3 Codex75.4%
artificialanalysis.ai · 2026-09-24
LMArena Text, Non-English (Elo)Few votesVerified
#26 of 561437 Elo
- #1Gemini 4 ArgonFew votes1512
- #2Claude Opus 5.51497
- #25GPT-5.4 mini1439
- #26Gemma 4 31BFew votes1437
- #27GPT-6 Luna1436
lmarena.ai · 2026-09-24
#23 of 251368 Elo
eqbench.com · 2026-09-24
Vision
CharXiv ReasoningEstimated
~#10 of 12 (estimated)~77.4%
- #1Gemini 3.8 Flash86.2%
- #2Kimi K384.8%
- #9Inkling78.1%
- ~#10Gemma 4 31BEstimated77.4%
- #10Claude Sonnet 577.0%
No published score — estimated from related benchmarks
OmniDocBench v1.5Estimated
~#5 of 8 (estimated)~83.0%
- #1MiniMax-M391.6%
- #2Kimi K391.1%
- #4Gemini 3.1 Pro85.3%
- ~#5Gemma 4 31BEstimated83.0%
- #5Claude Haiku 4.579.6%
No published score — estimated from related benchmarks
GDP.pdf (AA)Estimated
~#33 of 48 (estimated)~11.7%
- #1Muse Wonder Full32.6%
- #2GPT-6 Astra31.0%
- #31Qwen3.7 Plus12.2%
- ~#33Gemma 4 31BEstimated11.7%
- #33Muse Glimmer10.0%
No published score — estimated from related benchmarks
ChartographyEstimated
~#19 of 25 (estimated)~16.6%
- #1Claude Fable 5.273.4%
- #2GPT-6 Astra71.0%
- #18Claude Sonnet 516.6%
- ~#19Gemma 4 31BEstimated16.6%
- #19GLM-5.3 Flash16.3%
No published score — estimated from related benchmarks
Hard reasoning
Humanity's Last ExamVerified
#69 of 19823.6%
- #1Claude Fable 5.274.3%
- #2Claude Opus 5.561.4%
- #68Ling 3.0 Flash23.7%
- #69Gemma 4 31B23.6%
- #70K2 Horizon MoVA 36B A4B23.4%
artificialanalysis.ai · 2026-09-24
AIME (latest)Lab-reported
#13 of 3389.2%
- #1GLM-5.3 NVFP499.5%
- #2Beam 1.0 501B A23B97.7%
- #12gpt-oss-20b89.3%
- #13Gemma 4 31B89.2%
- #14Nova 2.0 Pro Preview89.0%
huggingface.co · 2026-04-02
Agents
τ²-bench (Telecom)Verified
#48 of 11959.9%
- #1Step 3.7 Flash98.5%
- #2DeepSeek-V4-Pro-0813-nvfp4-DSpark98.3%
- #47gpt-oss-20b60.2%
- #48Gemma 4 31B59.9%
- #49Claude Haiku 4.554.7%
artificialanalysis.ai · 2026-09-24
OSWorld-VerifiedEstimated
~#12 of 15 (estimated)~63.4%
- #1Kimi K384.8%
- #2Qwen3.8 27B84.3%
- #11Muse Glimmer65.9%
- ~#12Gemma 4 31BEstimated63.4%
- #12Nemotron 3.5 Super VL NVFP452.3%
No published score — estimated from related benchmarks
OSWorld 2.0Estimated
~last of 5 (estimated)below measured range (<47.5%)
No published score — estimated from related benchmarks
BrowseCompEstimated
~#19 of 22 (estimated)~61.6%
- #1Atria Dawn Preview92.5%
- #2GPT-6 Astra91.5%
- #18Nemotron 3.5 Bolt 74B-A9B70.0%
- ~#19Gemma 4 31BEstimated61.6%
- #19Mistral Medium 3.548.6%
No published score — estimated from related benchmarks
GDPval-AA (Elo)Verified
#62 of 91606 Elo
- #1Claude Opus 5.51866
- #2Muse Wonder Full1857
- #61Nemotron 3.5 Lightning624
- #62Gemma 4 31B606
- #63gpt-oss-120b596
artificialanalysis.ai · 2026-09-24