BenchLeader
Menu
Best LLMsChoose a modelLeaderboardsModelsBenchmarksProvidersReleases9
GoogleOpen weights: google/gemma-4-12BReasoning modelCompare with another model

Gemma 4 12B

Reasoning effort

Gemma 4 12B is a Google open-weights reasoning model, released 3 Jun 2026. Its best configuration (thinking reasoning effort) ranks #370 of 785 on the BenchLeader Index at 50.5 ±7.8, in the upper half. The ± is the point: 434 other configurations score within that range, so they and this one cannot be told apart on quality alone — price and speed are what separate them. It scores highest in instruction following (72) and lowest in knowledge (38). At $0.150 per million tokens blended it is among the cheapest fifth of ranked models. Output speed of 120 tokens per second puts it faster than most, with a first answer in 18.9 s. It has been measured at 2 reasoning-effort settings; this summary describes the best-scoring one, and the tabs above switch between them. Last measured 3 Jun 2026.

Blended price
$0.150/M
$0.100 in · $0.300 out
Output speed
120 tok/s
measured by Artificial Analysis
First answer
19 s
Context
262k
Full answer
23 s
median, reasoning included
Cost per run
–
Released
3 Jun 2026
How it scores by categoryDashed line = average model (50). One step of 15 = one standard deviation.
  1. Overall index51
  2. Reasoning41
  3. Agents & tools50
  4. Knowledge38
  5. Instruction following72
  6. Multimodal53
  7. Long context57
  8. Composite45

Versions

Google has shipped 2 models under this name. Each is ranked on its own results; a newer version often has fewer results so far, which holds its index nearer the average until more arrive.

ModelReleasedIndexRank
Gemma 4 12Bthinkingthis page3 Jun 202650.5#370
Gemma 3 12B12 Mar 202536.9#741

Reasoning-effort configurations

The same model behaves differently depending on how much it is allowed to think. Each row is one setting, scored only on the benchmarks that were run at that setting. “Not stated” collects results from publishers that did not say which setting they used; for a reasoning model that is usually its thinking mode, but we do not assume it. Pick a setting here or at the top of the page to see its price, speed and category scores.

EffortIndexRankSpeedFirst answerChat reply costCategories
no reasoning44.6#533127 tok/s2.22 s$0.0001Agents & tools 44 · Composite 40 · Instruction following 47 · Knowledge 38 · Long context 42 · Multimodal 45 · Reasoning 41
thinkingbest50.5#370120 tok/s19 s$0.0001Agents & tools 50 · Composite 45 · Instruction following 72 · Knowledge 38 · Long context 57 · Multimodal 53 · Reasoning 41

How its index has moved

9 Sept 2026 to 21 Oct 2026
374859
  • thinking
  • no reasoning

The index is recomputed from scratch every day, so a line moves when a new benchmark result lands, when a publisher revises a score, or when the models it is normalised against change. Early movement usually means the score is still settling.

Benchmark results

See all LLM benchmarks. One column per reasoning effort. Rank is among every configuration of every model on that benchmark. Hover a score for the run it came from.

Reasoning

Benchmarkno reasoningthinkingSourceTrend
GPQA Diamond (AA)not in index66.1%#33575.3%#248Artificial Analysis
Humanity's Last Exam (AA)not in index6.3%#39615.7%#251Artificial Analysis
CritPt0.0%#3150.0%#315Artificial Analysis

Coding

BenchmarkthinkingSourceTrend
Terminal-Bench 2.1 (AA)not in index27.3%#162Artificial Analysis

Agents & tools

Benchmarkno reasoningthinkingSourceTrend
Terminal-Bench Hard11.4%#21518.2%#171Artificial Analysis
τ²-Bench Telecom (AA)not in index31.9%#24936.3%#234Artificial Analysis
GDPval-AA v2.1not in index–0.0%#124Artificial Analysis

Knowledge

Benchmarkno reasoningthinkingSourceTrend
AA-Omniscience-52.8#414-52.7#413Artificial Analysis
AA-Omniscience: accuracynot in index11.9%#47715.6%#408Artificial Analysis
AA-Omniscience: non-hallucinationnot in index26.6%#23319.0%#289Artificial Analysis

Instruction following

Benchmarkno reasoningthinkingSourceTrend
IFBench45.2%#20673.5%#39Artificial Analysis

Multimodal

Benchmarkno reasoningthinkingSourceTrend
MMMU-Pro62.0%#20269.7%#152Artificial Analysis

Long context

Benchmarkno reasoningthinkingSourceTrend
AA-LCR35.0%#37563.7%#264Artificial Analysis

Composite

Benchmarkno reasoningthinkingSourceTrend
AA Intelligence Index v4.3.29.4#40114.2#299Artificial Analysis

What a task costs

Estimates from list price, output speed and time to first answer for the best configuration. “With caching” assumes three-quarters of the input is served from the prompt cache. Reasoning tokens are not modelled.

WorkloadTokens in / outCostWith cachingTime
Chat reply400 / 300$0.0001–21.4 s
Summarize a 30-page report12,000 / 600$0.0014–24.0 s
Code edit6,000 / 1,500$0.0011–31.5 s
Agentic coding session60,000 / 4,000$0.0072–52.3 s
Structured extraction2,000 / 200$0.0003–20.6 s

More from Google

New this month

See also

Best under $2/MBest under $0.50/MBest open weightsAll Google models

Questions people ask

What are Gemma 4 12B's benchmark scores?
As of 25 Oct 2026, Gemma 4 12B (thinking effort) scores 50.5 ±7.8 on the BenchLeader Index, strongest in instruction following (72) and weakest in knowledge (38). That rests on 7 independent benchmark results, and every score on this page links to its publisher.
How does Gemma 4 12B rank?
Gemma 4 12B (thinking effort) is #370 of 785 ranked configurations on 25 Oct 2026. 434 other configurations score within its margin, so the exact position matters less than the band. The leader is Claude Opus 5.5 (max effort) at 72.6.
How much does Gemma 4 12B cost per 1M tokens?
As of 25 Oct 2026, Gemma 4 12B costs $0.100 per 1M input tokens and $0.300 per 1M output tokens, $0.150 blended (three input tokens per output token).
When was Gemma 4 12B released?
Google released Gemma 4 12B on 3 Jun 2026. BenchLeader has results for it at 2 reasoning-effort settings.

Data as of 25 Oct 2026. Compare these configurations.

Cite as: BenchLeader, “Gemma 4 12B: benchmarks, pricing, speed and rank”, data as of 25 Oct 2026, https://www.benchleader.com/models/gemma-4-12b.