GoogleGemma 4 26B A4B
Best configuration ranks #209 of 610 on the BenchLeader Index at 54.0 ±5.4. Last measured 8 Sept 2026. Released 2 Apr 2026.
- Blended price
- $0.190/M
- $0.120 in · $0.400 out
- Output speed
- 30 tok/s
- First answer
- 0.63 s
- Context
- 262k
- Overall index54
- Reasoning63
- Coding46
- Agents & tools45
- Knowledge40
- Instruction following70
- Human preference63
- Multimodal57
- Long context59
- Composite50
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. “Default” means the publisher did not say which setting was used.
| Effort | Index | Rank | Speed | First answer | Chat reply cost | Categories |
|---|---|---|---|---|---|---|
| no reasoning | 47.8 | #345 | 79 tok/s | 1.34 s | $0.0002 | Agents & tools 55 · Composite 46 · Instruction following 47 · Knowledge 35 · Long context 47 · Multimodal 51 |
| minimal | – | – | 30 tok/s | 0.63 s | $0.0002 | Maths 60 · Reasoning 54 |
| defaultbest | 54.0 | #209 | 30 tok/s | 0.63 s | $0.0002 | Agents & tools 45 · Coding 46 · Composite 50 · Human preference 63 · Instruction following 70 · Knowledge 40 · Long context 59 · Multimodal 57 · Reasoning 63 |
Benchmark results
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 | default | Source | Trend |
|---|---|---|---|---|---|
| GPQA Diamond | – | 73.2%#145 | – | Epoch AI Benchmarking Hub | |
| LMArena Hard Prompts | – | – | 1462#76 | LMArena | |
| GPQA Diamond (AA)not in index | 71.4%#276 | – | 79.2%#194 | Artificial Analysis | |
| Humanity's Last Exam (AA)not in index | 11.5%#228 | – | 19.3%#165 | Artificial Analysis |
Coding
| Benchmark | no reasoning | minimal | default | Source | Trend |
|---|---|---|---|---|---|
| SciCode | – | – | 40.0%#106 | SciCode | |
| WeirdML | – | – | 35.2%#122 | WeirdML | |
| LMArena Coding | – | – | 1481#89 | LMArena | |
| LMArena WebDev | – | – | 1362#84 | LMArena |
Agents & tools
| Benchmark | no reasoning | minimal | default | Source | Trend |
|---|---|---|---|---|---|
| Terminal-Bench Hard | 25.0%#133 | – | 13.6%#194 | Artificial Analysis | |
| τ²-Bench Telecom (AA)not in index | 40.4%#217 | – | 43.6%#208 | Artificial Analysis |
Maths
| Benchmark | no reasoning | minimal | default | Source | Trend |
|---|---|---|---|---|---|
| OTIS Mock AIME | – | 82.2%#93 | – | Epoch AI Benchmarking Hub |
Knowledge
| Benchmark | no reasoning | minimal | default | Source | Trend |
|---|---|---|---|---|---|
| AA-Omniscience | -62.3#403 | – | -50.8#322 | Artificial Analysis |
Instruction following
| Benchmark | no reasoning | minimal | default | Source | Trend |
|---|---|---|---|---|---|
| IFBench | 45.4%#201 | – | 72.5%#48 | Artificial Analysis |
Human preference
| Benchmark | no reasoning | minimal | default | Source | Trend |
|---|---|---|---|---|---|
| LMArena Text | – | – | 1438#82 | LMArena |
Multimodal
| Benchmark | no reasoning | minimal | default | Source | Trend |
|---|---|---|---|---|---|
| LMArena Vision | – | – | 1259#43 | LMArena | |
| MMMU-Pro | 66.7%#137 | – | 69.3%#121 | Artificial Analysis | |
| MMMU-Pro (official)not in index | – | – | 73.8%#9 | MMMU |
Long context
| Benchmark | no reasoning | minimal | default | Source | Trend |
|---|---|---|---|---|---|
| AA-LCR | 42.3%#280 | – | 65.7%#188 | Artificial Analysis |
Composite
| Benchmark | no reasoning | minimal | default | Source | Trend |
|---|---|---|---|---|---|
| Epoch Capabilities Indexnot in index | – | – | 141.9#89 | Epoch AI Benchmarking Hub | |
| AA Intelligence Index | 13.1#250 | – | 16.7#206 | 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 |
|---|---|---|---|---|---|---|
| Cloudflare | 50 tok/s | 0.48 s | $0.100 | $0.300 | 256k | – |
| NextBit | 47 tok/s | 0.42 s | $0.100 | $0.400 | 262k | bf16 |
| Parasail | 46 tok/s | 0.63 s | $0.130 | $0.400 | 262k | bf16 |
| NovitaAI | 41 tok/s | 0.94 s | $0.130 | $0.400 | 262k | bf16 |
| SiliconFlow | 30 tok/s | 1.32 s | $0.120 | $0.400 | 262k | fp8 |
| Venice | 26 tok/s | 1.11 s | $0.130 | $0.400 | 256k | bf16 |
| Google Vertex | 21 tok/s | 2.73 s | $0.150 | $0.600 | 262k | – |
| DeepInfra | 9 tok/s | 0.27 s | $0.070 | $0.340 | 262k | fp8 |
| Darkbloom | 7 tok/s | 0.51 s | $0.042 | $0.220 | 131k | – |
Price history
Listed price per 1M tokens over time, as recorded by OpenRouter for the provider with the longest history.
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.
| Workload | Tokens in / out | Cost | With caching | Time |
|---|---|---|---|---|
| Chat reply | 400 / 300 | $0.0002 | – | 10.6 s |
| Summarise a 30-page report | 12,000 / 600 | $0.0017 | – | 20.6 s |
| Code edit | 6,000 / 1,500 | $0.0013 | – | 50.6 s |
| Agentic coding session | 60,000 / 4,000 | $0.0088 | – | 2.2 min |
| Structured extraction | 2,000 / 200 | $0.0003 | – | 7.3 s |
See also
Data as of 9 Sept 2026. Compare these configurations.