GoogleGemma 4 26B A4B
Gemma 4 26B A4B is a Google open-weights model, released 2 Apr 2026. Its best configuration ranks #212 of 785 on the BenchLeader Index at 56.2 ±7.5, in the upper half. The ± is the point: 363 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 maths (65) and lowest in coding (44). At $0.198 per million tokens blended it is cheaper than most ranked models. Output speed of 46 tokens per second puts it in the slowest quarter, with a first token in 0.5 s. It has been measured at 4 reasoning-effort settings; this summary describes the best-scoring one, and the tabs above switch between them. Last measured 23 Oct 2026.
- Blended price
- $0.198/M
- $0.130 in · $0.400 out
- Output speed
- 46 tok/s
- OpenRouter traffic, 7-day median; not yet measured by Artificial Analysis
- First answer
- 0.50 s
- Context
- 262k
- Full answer
- –
- Cost per run
- –
- Released
- 2 Apr 2026
- Overall index56
- Reasoning62
- Coding44
- Maths65
- Human preference62
- Multimodal59
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.
| Effort | Index | Rank | Speed | First answer | Chat reply cost | Categories |
|---|---|---|---|---|---|---|
| no reasoning | 46.6 | #482 | 96 tok/s | 0.96 s | $0.0002 | Agents & tools 56 · Composite 44 · Instruction following 48 · Knowledge 33 · Long context 46 · Multimodal 50 · Reasoning 41 |
| minimal | – | – | 46 tok/s | 0.50 s | $0.0002 | Maths 59 · Reasoning 54 |
| thinking | 50.0 | #383 | 81 tok/s | 26 s | $0.0001 | Agents & tools 42 · Composite 48 · Instruction following 71 · Knowledge 39 · Long context 58 · Multimodal 52 · Reasoning 41 |
| not statedbest | 56.2 | #212 | 46 tok/s | 0.50 s | $0.0002 | Coding 44 · Human preference 62 · Maths 65 · Multimodal 59 · Reasoning 62 |
How its index has moved
8 Sept 2026 to 23 Oct 2026- not stated
- 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
| Benchmark | no reasoning | minimal | thinking | not stated | Source | Trend |
|---|---|---|---|---|---|---|
| GPQA Diamond | – | 73.2%#158 | – | – | Epoch AI Benchmarking Hub | |
| Arena Hard Promptseffort not stated | – | – | – | 1461#92 | Arena (LMArena) | |
| GPQA Diamond (AA)not in index | 71.4%#286 | – | 79.2%#200 | – | Artificial Analysis | |
| Humanity's Last Exam (AA)not in index | 11.5%#292 | – | 19.3%#221 | – | Artificial Analysis | |
| CritPt | 0.0%#315 | – | 0.0%#315 | – | Artificial Analysis | |
| LMCAnot in indexeffort not stated | – | – | – | 29.7%#136 | Epoch AI Benchmarking Hub | |
| DTBenchnot in indexeffort not stated | – | – | – | 74.9%#150 | Epoch AI Benchmarking Hub |
Coding
| Benchmark | thinking | not stated | Source | Trend |
|---|---|---|---|---|
| SciCodeeffort not stated | – | 40.0%#162 | SciCode | |
| WeirdMLeffort not stated | – | 35.2%#127 | WeirdML | |
| Arena Codingeffort not stated | – | 1483#103 | Arena (LMArena) | |
| Arena WebDeveffort not stated | – | 1360#106 | Arena (LMArena) | |
| SciCode (AA)not in index | 40.0%#174 | – | Artificial Analysis | |
| ALE-Benchnot in indexeffort not stated | – | 927.2#62 | Epoch AI Benchmarking Hub | |
| Terminal-Bench 2.1 (AA)not in index | 39.0%#141 | – | Artificial Analysis |
Agents & tools
| Benchmark | no reasoning | thinking | not stated | Source | Trend |
|---|---|---|---|---|---|
| Terminal-Bench Hard | 25.0%#138 | 13.6%#201 | – | Artificial Analysis | |
| τ²-Bench Telecom (AA)not in index | 40.4%#223 | 43.6%#214 | – | Artificial Analysis | |
| GDPval-AA v2.1not in index | – | 3.3%#118 | – | Artificial Analysis | |
| τ³-Banking (AA) | – | 12.0%#134 | – | Artificial Analysis | |
| ITBench SRE (AA)not in index | – | 23.6%#38 | – | Artificial Analysis | |
| AutomationBenchnot in index | – | 1.9%#183 | – | Artificial Analysis | |
| FACTS Searchnot in indexeffort not stated | – | – | 60.3%#21 |
Maths
| Benchmark | minimal | not stated | Source | Trend |
|---|---|---|---|---|
| OTIS Mock AIME | 82.2%#108 | – | Epoch AI Benchmarking Hub | |
| Arena Mathseffort not stated | – | 1469#53 | Arena (LMArena) |
Knowledge
| Benchmark | no reasoning | thinking | Source | Trend |
|---|---|---|---|---|
| AA-Omniscience | -62.3#479 | -50.8#395 | Artificial Analysis | |
| AA-Omniscience: accuracynot in index | 15.5%#412 | 19.1%#318 | Artificial Analysis | |
| AA-Omniscience: non-hallucinationnot in index | 8.0%#471 | 13.6%#360 | Artificial Analysis |
Instruction following
| Benchmark | no reasoning | thinking | not stated | Source | Trend |
|---|---|---|---|---|---|
| IFBench | 45.4%#204 | 72.5%#48 | – | Artificial Analysis | |
| Arena Instruction Followingnot in indexeffort not stated | – | – | 1438#83 | Arena (LMArena) |
Human preference
| Benchmark | not stated | Source | Trend |
|---|---|---|---|
| Arena Text | 1438#97 | Arena (LMArena) | |
| Arena Creative Writingnot in index | 1399#108 | Arena (LMArena) | |
| Arena Multi-turnnot in index | 1448#90 | Arena (LMArena) | |
| Arena Longer Queriesnot in index | 1446#96 | Arena (LMArena) |
Multimodal
| Benchmark | no reasoning | thinking | not stated | Source | Trend |
|---|---|---|---|---|---|
| Arena Visioneffort not stated | – | – | 1260#56 | Arena (LMArena) | |
| MMMU-Pro | 66.7%#169 | 69.3%#153 | – | Artificial Analysis | |
| MMMU-Pro (official)not in indexeffort not stated | – | – | 73.8%#9 | MMMU |
Long context
| Benchmark | no reasoning | thinking | Source | Trend |
|---|---|---|---|---|
| AA-LCR | 42.3%#350 | 65.7%#247 | Artificial Analysis |
Composite
| Benchmark | no reasoning | thinking | not stated | Source | Trend |
|---|---|---|---|---|---|
| Epoch Capabilities Indexnot in indexeffort not stated | – | – | 141.9#104 | Epoch AI Benchmarking Hub | |
| AA Intelligence Index v4.3.2 | 13.1#314 | 16.7#269 | – | Artificial Analysis |
Where to run it
Every provider serving this model through OpenRouter, with throughput and first-token latency measured on live traffic over the last 30 minutes and each provider’s own price. Purple marks the best in each column.
| Provider | Speed | First token | Input $/M | Output $/M | Context | Quantisation |
|---|---|---|---|---|---|---|
| NovitaAI | 167 tok/s | 0.80 s | $0.130 | $0.400 | 262k | bf16 |
| Makora | 110 tok/s | 0.38 s | $0.070 | $0.240 | 256k | – |
| Parasail | 86 tok/s | 0.46 s | $0.130 | $0.400 | 262k | bf16 |
| DeepInfra | 71 tok/s | 0.44 s | $0.070 | $0.340 | 262k | fp8 |
| Venice | 67 tok/s | 0.73 s | $0.130 | $0.400 | 256k | bf16 |
| NextBit | 63 tok/s | 0.35 s | $0.068 | $0.225 | 262k | bf16 |
| io.net | 52 tok/s | 0.60 s | $0.090 | $0.310 | 262k | bf16 |
| Cloudflare | 51 tok/s | 0.46 s | $0.100 | $0.300 | 256k | – |
| Google Vertex | 39 tok/s | 1.24 s | $0.150 | $0.600 | 262k | – |
| Darkbloom | 35 tok/s | 0.94 s | $0.042 | $0.220 | 131k | – |
| SiliconFlow | 29 tok/s | 1.21 s | $0.140 | $0.400 | 262k | fp8 |
| CoreWeave | 26 tok/s | 0.51 s | $0.100 | $0.300 | 262k | bf16 |
| DekaLLM | 15 tok/s | 1.29 s | $0.060 | $0.330 | 262k | bf16 |
Price history
The price above, per 1M tokens, every day it changed.
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 at $0.050 per 1M. Reasoning tokens are not modelled.
| Workload | Tokens in / out | Cost | With caching | Time |
|---|---|---|---|---|
| Chat reply | 400 / 300 | $0.0002 | $0.0001 | 7.0 s |
| Summarize a 30-page report | 12,000 / 600 | $0.0018 | $0.0011 | 13.5 s |
| Code edit | 6,000 / 1,500 | $0.0014 | $0.0010 | 33.1 s |
| Agentic coding session | 60,000 / 4,000 | $0.0094 | $0.0058 | 1.5 min |
| Structured extraction | 2,000 / 200 | $0.0003 | $0.0002 | 4.8 s |
More from Google
See also
Questions people ask
- What are Gemma 4 26B A4B's benchmark scores?
- As of 25 Oct 2026, Gemma 4 26B A4B scores 56.2 ±7.5 on the BenchLeader Index, strongest in maths (65) and weakest in coding (44). That rests on 8 independent benchmark results, and every score on this page links to its publisher.
- How does Gemma 4 26B A4B rank?
- Gemma 4 26B A4B is #212 of 785 ranked configurations on 25 Oct 2026. 363 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 26B A4B cost per 1M tokens?
- As of 25 Oct 2026, Gemma 4 26B A4B costs $0.130 per 1M input tokens and $0.400 per 1M output tokens, $0.198 blended (three input tokens per output token). The list price last changed on 19 Sept 2026.
- When was Gemma 4 26B A4B released?
- Google released Gemma 4 26B A4B on 2 Apr 2026. BenchLeader has results for it at 4 reasoning-effort settings.
Data as of 25 Oct 2026. Compare these configurations.
Cite as: BenchLeader, “Gemma 4 26B A4B: benchmarks, pricing, speed and rank”, data as of 25 Oct 2026, https://www.benchleader.com/models/gemma-4-26b-a4b.