vs
0:3
overall capability wins (1 ties across 4 scored). Gemma 4 26B leads overall.
- Arena0:1
- Pricing3:0
- Speed2:0
- Specs0:2
Verdict
Gemma 4 26B ranks higher on LMArena (+238 Elo); Gemma 4 26B offers ~31.3x larger context (256K vs 8K); Gemma 2 2B is ~6.6x cheaper on a blended token basis than Gemma 4 26B.
- Gemma 4 26B ranks higher on LMArena (+238 Elo)
- Gemma 4 26B offers ~31.3x larger context (256K vs 8K)
- Gemma 2 2B is ~6.6x cheaper on a blended token basis than Gemma 4 26B
Cost for 1M in + 250K out
Illustrative chat workload at primary-provider list prices.
- Gemma 2 2B
- $0.05
- Gemma 4 26B
- $0.30
Gemma 2 2B is about 6.0x cheaper on this workload.
Price per 1M tokens
Full comparison
Organization
Family
Gemma
Gemma
License
Gemma
Apache 2.0
Open weights
Yes
Yes
Release
Jul 31, 2024
Apr 8, 2026
Knowledge cutoff
-
-
API / provider
Groq / Together (ref.)
Google AI / self-host
Modalities
text → text
text, image → text
Specs
Context window
8K
256K
Max output
8K
8K
Parameters
2B
26B (3.8B act.)
Pricing
Input $/1M
$0.04
$0.15
Output $/1M
$0.04
$0.60
Blended $/1M (3∶1)
$0.04
$0.26
Speed
tok/s
200
95
TTFT (s)
0.12
0.28
Reasoning
MMLU-Pro
—
82.6
GPQA Diamond
—
82.3
Humanity's Last Exam
—
8.7
AIME 2025
—
—
MATH-500
—
—
Humanity's Last Exam (with tools)
—
—
AA-Omniscience Accuracy
—
—
AA-LCR v1.1
—
—
CritPt
—
—
MMMU-Pro
—
—
IFBench
—
—
Chartography
—
—
Chartography (With Tools)
—
—
Coding
SWE-bench Verified
—
—
SWE-bench Pro
—
—
SWE-bench Multilingual
—
—
LiveCodeBench
—
77.1
Terminal-Bench 2.1
—
—
Aider Polyglot
—
—
Terminal-Bench 3
—
—
BigCodeBench
—
—
SciCode
—
—
CursorBench
—
—
SWE-Rebench
—
—
NL2Repo-Bench
—
—
DeepSWE
—
—
WebDev Arena
—
1,362
Terminal-Bench 4.0
—
—
CursorBench 4.0
—
—
FrontierCode v1.1 (Main)
—
—
Arena
LMArena Elo
1,200
1,438
| Metric | Gemma 2 2B | Gemma 4 26B | Delta |
|---|---|---|---|
| Identity | |||
| Organization | — | ||
| Family | Gemma | Gemma | — |
| License | Gemma | Apache 2.0 | — |
| Open weights | Yes | Yes | — |
| Release | Jul 31, 2024 | Apr 8, 2026 | — |
| Knowledge cutoff | - | - | — |
| API / provider | Groq / Together (ref.) | Google AI / self-host | — |
| Modalities | text → text | text, image → text | — |
| Specs | |||
| Context window | 8K | 256K | B +248K |
| Max output | 8K | 8K | tie |
| Parameters | 2B | 26B (3.8B act.) | B +24 |
| Pricing | |||
| Input $/1M | $0.04 | $0.15 | A +$0.11 |
| Output $/1M | $0.04 | $0.60 | A +$0.56 |
| Blended $/1M (3∶1) | $0.04 | $0.26 | A +$0.22 |
| Speed | |||
| tok/s | 200 | 95 | A +105 |
| TTFT (s) | 0.12 | 0.28 | A +0.16 |
| Reasoning | |||
| MMLU-Pro | — | 82.6 | — |
| GPQA Diamond | — | 82.3 | — |
| Humanity's Last Exam | — | 8.7 | — |
| AIME 2025 | — | — | — |
| MATH-500 | — | — | — |
| Humanity's Last Exam (with tools) | — | — | — |
| AA-Omniscience Accuracy | — | — | — |
| AA-LCR v1.1 | — | — | — |
| CritPt | — | — | — |
| MMMU-Pro | — | — | — |
| IFBench | — | — | — |
| Chartography | — | — | — |
| Chartography (With Tools) | — | — | — |
| Coding | |||
| SWE-bench Verified | — | — | — |
| SWE-bench Pro | — | — | — |
| SWE-bench Multilingual | — | — | — |
| LiveCodeBench | — | 77.1 | — |
| Terminal-Bench 2.1 | — | — | — |
| Aider Polyglot | — | — | — |
| Terminal-Bench 3 | — | — | — |
| BigCodeBench | — | — | — |
| SciCode | — | — | — |
| CursorBench | — | — | — |
| SWE-Rebench | — | — | — |
| NL2Repo-Bench | — | — | — |
| DeepSWE | — | — | — |
| WebDev Arena | — | 1,362 | — |
| Terminal-Bench 4.0 | — | — | — |
| CursorBench 4.0 | — | — | — |
| FrontierCode v1.1 (Main) | — | — | — |
| Arena | |||
| LMArena Elo | 1,200 | 1,438 | B +238 Elo |
Gemma 2 2B
- Context
- 8K / 8K out
- Parameters
- 2B
- Price
- $0.04 / $0.04
- Speed
- 200 tok/s · 0.12s TTFT
- Modalities
- text → text
- License
- Gemma
Tiny open Gemma 2 instruct — ultra-cheap edge and high-throughput classification baseline.
Gemma 4 26B
- Context
- 256K / 8K out
- Parameters
- 26B (3.8B act.)
- Price
- $0.15 / $0.60
- Speed
- 95 tok/s · 0.28s TTFT
- Modalities
- text, image → text
- License
- Apache 2.0
Open Gemma 4 mid-size instruct model available via Google Cloud Agent Platform and self-host.
Benchmark charts
Winner bars are emphasized. Per-benchmark deltas sit above each chart.