Neither wins outright: Google: Gemma 4 31B (batch) is the pick when you need reasoning / thinking mode, Mistral: Codestral 2508 (batch) when cost matters at scale.
| Spec | Google: Gemma 4 31B (batch) | Mistral: Codestral 2508 (batch) |
|---|---|---|
| Provider | Mistral AI | |
| Input price (per 1M tokens) | $0.39 | $0.15 |
| Output price (per 1M tokens) | $0.97 | $0.45 |
| Blended price (3:1 input/output) | $0.53 | $0.22 |
| Cheaper than … of all models we track | 52% | 68% |
| Context window | 262K tokens (~393 pages) | 256K tokens (~384 pages) |
| Larger context than … of all models | 48% | 45% |
| Max output per response | 236K tokens (~176,900 words) | 205K tokens (~153,600 words) |
| Input types | image, text, video | text, file |
| Output types | text | text |
| Free variant | No | No |
| Listed since | 2026-04-02 | 2025-08-01 |
Monthly cost at list price for five common workloads. Token counts per unit are stated so you can scale them to your own traffic. A workload that does not fit a model's context window is marked rather than priced.
| Workload (per month) | Google: Gemma 4 31B (batch) | Mistral: Codestral 2508 (batch) | Difference |
|---|---|---|---|
| Customer-support chatbot 50,000 conversations — each a 6-turn conversation (1,500 in / 400 out tokens) | $48.65 | $20.25 | Mistral: Codestral 2508 (batch) saves $28.40 |
| RAG search over documents 20,000 questions — each a question answered from ~8 retrieved passages (6,000 in / 500 out tokens) | $56.50 | $22.50 | Mistral: Codestral 2508 (batch) saves $34.00 |
| Coding / tool-using agent 2,000 tasks — each a multi-step task with tool calls and file context (60,000 in / 6,000 out tokens) | $58.44 | $23.40 | Mistral: Codestral 2508 (batch) saves $35.04 |
| Batch summarisation 10,000 documents — each a 6,000-word report summarised to a page (8,000 in / 600 out tokens) | $37.02 | $14.70 | Mistral: Codestral 2508 (batch) saves $22.32 |
| Long-document analysis 500 documents — each a 150-page contract or codebase read in one pass (100,000 in / 2,000 out tokens) | $20.47 | $7.95 | Mistral: Codestral 2508 (batch) saves $12.52 |
| Capability | Google: Gemma 4 31B (batch) | Mistral: Codestral 2508 (batch) | What it lets you do |
|---|---|---|---|
| Tool / function calling | ✓ Yes | ✓ Yes | call your APIs and run agent loops |
| Structured output (JSON schema) | ✓ Yes | ✓ Yes | return JSON that validates against your schema |
| Reasoning / thinking mode | ✓ Yes | — | spend extra tokens thinking before answering hard problems |
| Image input | ✓ Yes | — | read screenshots, charts and scanned pages |
| Audio input | — | — | take speech or audio directly |
| File / PDF input | — | ✓ Yes | accept documents without your own parsing step |
| Image output | — | — | generate images, not just text |
| Built-in web search | — | — | answer from live web results |
| Seed (reproducible sampling) | — | ✓ Yes | repeat a generation for tests and evals |
| Log-probabilities | — | — | read token confidence for classification and scoring |
From each model's published parameters and input/output types. “—” means not advertised, not proven absent.
| Date | Model | Input / 1M | Output / 1M |
|---|---|---|---|
| 2026-09-10 | Mistral: Codestral 2508 (batch) | $0.15 | $0.45 |
Neither wins outright: Google: Gemma 4 31B (batch) is the pick when you need reasoning / thinking mode, Mistral: Codestral 2508 (batch) when cost matters at scale. In short: Google: Gemma 4 31B (batch) when you need reasoning / thinking mode, you need image input, you want the newer model; Mistral: Codestral 2508 (batch) when cost matters at scale, you need file / PDF input, you need seed (reproducible sampling).
Mistral: Codestral 2508 (batch) is cheaper: $0.22 vs $0.53 per 1M tokens at a typical 3:1 input/output mix, 58% less. For 50,000 support conversations a month that is $20.25 against $48.65.
Google: Gemma 4 31B (batch) costs $0.39 per 1M input tokens and $0.97 per 1M output tokens; Mistral: Codestral 2508 (batch) costs $0.15 input and $0.45 output. Output tokens are billed separately, so long answers shift the gap.
Google: Gemma 4 31B (batch): 262K tokens, about 393 pages, against 256K (about 384 pages) for Mistral: Codestral 2508 (batch).
Tool calling — Google: Gemma 4 31B (batch): yes, Mistral: Codestral 2508 (batch): yes. JSON-schema structured output — Google: Gemma 4 31B (batch): yes, Mistral: Codestral 2508 (batch): yes.
Only Google: Gemma 4 31B (batch) accepts image input; Mistral: Codestral 2508 (batch) is text-only for input.
No free variant is listed for either; both are pay-per-token.
Google: Gemma 4 31B (batch) can write up to 236K tokens (about 176,900 words) per response; Mistral: Codestral 2508 (batch) up to 205K (about 153,600 words).
Google: Gemma 4 31B (batch) was listed on 2026-04-02; Mistral: Codestral 2508 (batch) on 2025-08-01.
Google: Gemma 4 31B (batch): price unchanged since we started tracking it on 2026-08-29. Mistral: Codestral 2508 (batch): input price cut 50% ($0.30 → $0.15/1M), output price cut 50% ($0.90 → $0.45/1M) on 2026-09-10.
Gemma 4 31B Instruct is Google DeepMind's 30.7B dense multimodal model supporting text and image input with text output. Features a 256K token context window, configurable thinking/reasoning mode, native function...
Full Google: Gemma 4 31B (batch) specs →Mistral's cutting-edge language model for coding released end of July 2025. Codestral specializes in low-latency, high-frequency tasks such as fill-in-the-middle (FIM), code correction and test generation. [Blog Post](https://mistral.ai/news/codestral-25-08)
Full Mistral: Codestral 2508 (batch) specs →