vs
0:3
overall capability wins (0 ties across 3 scored). Devstral 2 leads overall.
- Pricing3:0
- Speed2:0
- Specs0:3
Verdict
Devstral 2 offers ~7.8x larger context (256K vs 33K); Devstral 2 allows ~7.8x more max output (64K vs 8K); Codestral 22B is ~2.7x cheaper on a blended token basis than Devstral 2.
- Devstral 2 offers ~7.8x larger context (256K vs 33K)
- Devstral 2 allows ~7.8x more max output (64K vs 8K)
- Codestral 22B is ~2.7x cheaper on a blended token basis than Devstral 2
Cost for 1M in + 250K out
Illustrative chat workload at primary-provider list prices.
- Codestral 22B
- $0.35
- Devstral 2
- $0.90
Codestral 22B is about 2.6x cheaper on this workload.
Price per 1M tokens
Full comparison
Organization
Mistral
Mistral
Family
Mistral
Mistral
License
MNPL
Apache 2.0
Open weights
Yes
Yes
Release
May 29, 2024
Dec 1, 2025
Knowledge cutoff
-
-
API / provider
Mistral / Together (ref.)
Mistral
Modalities
text → text
text → text
Specs
Context window
33K
256K
Max output
8K
64K
Parameters
22B
123B
Pricing
Input $/1M
$0.20
$0.40
Output $/1M
$0.60
$2
Blended $/1M (3∶1)
$0.30
$0.80
Speed
tok/s
90
85
TTFT (s)
0.3
0.35
Reasoning
MMLU-Pro
—
76.2
GPQA Diamond
—
59.4
Humanity's Last Exam
—
3.6
AIME 2025
—
36.7
MATH-500
—
—
Humanity's Last Exam (with tools)
—
—
AA-Omniscience Accuracy
—
20.8
AA-LCR v1.1
—
32.3
CritPt
—
—
MMMU-Pro
—
—
IFBench
—
38.1
Chartography
—
—
Chartography (With Tools)
—
—
Coding
SWE-bench Verified
—
—
SWE-bench Pro
—
—
SWE-bench Multilingual
—
—
LiveCodeBench
—
44.8
Terminal-Bench 2.1
—
30.3
Aider Polyglot
—
—
Terminal-Bench 3
—
—
BigCodeBench
41.8
—
SciCode
—
32.8
CursorBench
—
—
SWE-Rebench
—
42.0
NL2Repo-Bench
—
—
DeepSWE
—
—
WebDev Arena
—
1,194
Terminal-Bench 4.0
—
—
CursorBench 4.0
—
—
FrontierCode v1.1 (Main)
—
—
Arena
LMArena Elo
—
—
| Metric | Codestral 22B | Devstral 2 | Delta |
|---|---|---|---|
| Identity | |||
| Organization | Mistral | Mistral | — |
| Family | Mistral | Mistral | — |
| License | MNPL | Apache 2.0 | — |
| Open weights | Yes | Yes | — |
| Release | May 29, 2024 | Dec 1, 2025 | — |
| Knowledge cutoff | - | - | — |
| API / provider | Mistral / Together (ref.) | Mistral | — |
| Modalities | text → text | text → text | — |
| Specs | |||
| Context window | 33K | 256K | B +223K |
| Max output | 8K | 64K | B +56K |
| Parameters | 22B | 123B | B +101 |
| Pricing | |||
| Input $/1M | $0.20 | $0.40 | A +$0.20 |
| Output $/1M | $0.60 | $2 | A +$1.40 |
| Blended $/1M (3∶1) | $0.30 | $0.80 | A +$0.50 |
| Speed | |||
| tok/s | 90 | 85 | A +5.00 |
| TTFT (s) | 0.3 | 0.35 | A +0.05 |
| Reasoning | |||
| MMLU-Pro | — | 76.2 | — |
| GPQA Diamond | — | 59.4 | — |
| Humanity's Last Exam | — | 3.6 | — |
| AIME 2025 | — | 36.7 | — |
| MATH-500 | — | — | — |
| Humanity's Last Exam (with tools) | — | — | — |
| AA-Omniscience Accuracy | — | 20.8 | — |
| AA-LCR v1.1 | — | 32.3 | — |
| CritPt | — | — | — |
| MMMU-Pro | — | — | — |
| IFBench | — | 38.1 | — |
| Chartography | — | — | — |
| Chartography (With Tools) | — | — | — |
| Coding | |||
| SWE-bench Verified | — | — | — |
| SWE-bench Pro | — | — | — |
| SWE-bench Multilingual | — | — | — |
| LiveCodeBench | — | 44.8 | — |
| Terminal-Bench 2.1 | — | 30.3 | — |
| Aider Polyglot | — | — | — |
| Terminal-Bench 3 | — | — | — |
| BigCodeBench | 41.8 | — | — |
| SciCode | — | 32.8 | — |
| CursorBench | — | — | — |
| SWE-Rebench | — | 42.0 | — |
| NL2Repo-Bench | — | — | — |
| DeepSWE | — | — | — |
| WebDev Arena | — | 1,194 | — |
| Terminal-Bench 4.0 | — | — | — |
| CursorBench 4.0 | — | — | — |
| FrontierCode v1.1 (Main) | — | — | — |
| Arena | |||
| LMArena Elo | — | — | — |
Codestral 22B
- Context
- 33K / 8K out
- Parameters
- 22B
- Price
- $0.20 / $0.60
- Speed
- 90 tok/s · 0.3s TTFT
- Modalities
- text → text
- License
- MNPL
Original open-weight Codestral 22B — fill-in-the-middle coding model under Mistral's MNPL.
Devstral 2
- Context
- 256K / 64K out
- Parameters
- 123B
- Price
- $0.40 / $2
- Speed
- 85 tok/s · 0.35s TTFT
- Modalities
- text → text
- License
- Apache 2.0
Open-weight agentic coding model for autonomous software engineering (successor path: Medium 3.5).
Benchmark charts
Winner bars are emphasized. Per-benchmark deltas sit above each chart.
These two models have no overlapping published benchmarks in our dataset. Compare specs and pricing instead.