Mistral Large 3 2512 vs Kimi K2.5
Mistral AIMoonshot AIboth via one key, provider price + 3%
Mistral Large 3 2512 is Mistral's flagship sparse MoE model (41B active / 675B total params), released under Apache 2.0. Kimi K2.5 is Moonshot AI's multimodal agent-focused model, continued-pretrained from Kimi K2 on roughly 15T tokens. Both ship with a 262,144-token context window, both run on OpenKey under one API key with a flat 3% fee on top of provider list price. The real difference shows up in the Design Arena rankings and in what each model can ingest as input.
Spec vs spec
| Spec | Mistral Large 3 2512 | Kimi K2.5 |
|---|---|---|
| Context window | 262K | 262K |
| Input modalities | text, image, file | text, image |
| Output modalities | text | text |
| Released | Dec 1, 2025 | Jan 27, 2026 |
| Reasoning | — | optional |
Pricing
Per 1M tokens. Provider price plus the flat 3% fee — the sum is what you pay.
mistralai/mistral-large-2512
Input · 1M tokens
$0.500 + 3%$0.515
Output · 1M tokens
$1.50 + 3%$1.54
Cache read · 1M tokens
$0.050 + 3%$0.052
FEE — FLAT, EVERY MODEL3%
moonshotai/kimi-k2.5
Input · 1M tokens
$0.375 + 3%$0.386
Output · 1M tokens
$2.02 + 3%$2.09
FEE — FLAT, EVERY MODEL3%
One workload, priced on both
10M input + 2M output tokens at each model's price, flat 3% fee included.
mistralai/mistral-large-2512
$8.24
$8.00 provider + 3%
moonshotai/kimi-k2.5Cheaper
$8.03
$7.80 provider + 3%
Benchmarks
Design Arena categories where both models have results. Higher Elo and lower rank win.
| Mistral Large 3 2512 | Kimi K2.5 | |||
|---|---|---|---|---|
| Category | Elo | Rank | Elo | Rank |
| 3D | 1176 | #51 | 1286 | #22 |
| ASCII art | 1115 | #43 | 1214 | #17 |
| Code | 1191 | #56 | 1286 | #20 |
| Data viz | 1180 | #55 | 1270 | #21 |
| Game dev | 1146 | #65 | 1272 | #23 |
| SVG | 1050 | #62 | 1210 | #25 |
| UI components | 1157 | #59 | 1290 | #19 |
| Websites | 1205 | #53 | 1291 | #16 |
Head-to-head preference voting. How we filter and rank
Pricing math
Mistral Large 3 2512 costs $0.50/M input and $1.50/M output from the provider; on OpenKey that's $0.515/M input and $1.545/M output ($0.50 × 1.03, $1.50 × 1.03). Kimi K2.5 costs $0.375/M input and $2.025/M output from the provider, or $0.38625/M input and $2.08575/M output on OpenKey (both × 1.03).
For a workload of 10M input tokens + 2M output tokens: Mistral Large 3 2512 runs $8.00, Kimi K2.5 runs $7.80. Kimi's cheaper input price (0.75x Mistral's — input ratio is 1.33 in Mistral's favor for input cost) outweighs its higher output price at this input-heavy ratio. Mistral also offers cache reads at $0.05/M, which Kimi doesn't list, so repeated-context workloads could tilt back toward Mistral.
Coding and agentic performance
This is where the two diverge hardest. Kimi K2.5 has dedicated agent-arena benchmarks — rank 2 in Godot game dev (elo 1254, 59.5% win rate), rank 14 in fullstack (elo 1182, 54.2% win rate), rank 15 in web apps (elo 1194, 50.3% win rate). Across the general model-arena categories, Kimi K2.5 ranks between 16 and 25 in every category (codecategories: rank 20, elo 1286; website: rank 16, elo 1291).
Mistral Large 3 2512 has no agent-arena entries at all — it only has model-arena scores, ranked 43 to 65 across categories (codecategories: rank 56, elo 1191; website: rank 53, elo 1205). Mistral also reports Artificial Analysis scores: coding index 20.1, agentic index 5.5, intelligence index 15.9 — Kimi K2.5 has no Artificial Analysis data in this comparison, so use the Design Arena numbers as the apples-to-apples reference.
Modality and tooling differences
Mistral Large 3 2512 accepts text, image, and file inputs; Kimi K2.5 accepts text and image only — no file upload. If your pipeline feeds PDFs or raw file attachments directly to the model, Mistral is the only option here. Both output text only.
On tooling, Kimi K2.5 supports a longer parameter list including `reasoning`, `include_reasoning`, `top_k`, `top_logprobs`, and `min_p`, and ships with reasoning enabled by default (not mandatory). Mistral Large 3 2512's supported parameters are the standard set — no reasoning toggle, no logprobs, no top_k. If you need visibility into a model's reasoning trace, Kimi K2.5 is built for it; Mistral isn't.
Context and release timing
Both models match exactly on context window — 262,144 tokens each, a 1.0 context ratio. No advantage either way for long-document work; pick based on cost and capability instead. Kimi K2.5 released 2026-01-27, about two months after Mistral Large 3 2512 (2025-12-01), and comes with continued pretraining on roughly 15T tokens on top of the original Kimi K2 base — that recency shows up in its Design Arena agent-category rankings, none of which Mistral Large 3 2512 has entries for.
Which model for which job
| Use case | Pick | Why |
|---|---|---|
| Agentic coding / autonomous dev tasks | Kimi K2.5 | Ranked #2 in Design Arena's agents/godotgamedev category (elo 1254) with no comparable agent-arena data for Mistral |
| Full-stack web app generation | Kimi K2.5 | Rank 14 in agents/fullstack (elo 1182, 54.2% win rate) vs Mistral's rank 56 in the general codecategories arena |
| File/PDF ingestion pipelines | Mistral Large 3 2512 | Only this model accepts file inputs; Kimi K2.5 supports text and image only |
| High input-token, low output-token workloads | Kimi K2.5 | Provider input price of $0.375/M is 0.75x Mistral's $0.50/M |
| Self-hosting or license flexibility | Mistral Large 3 2512 | Released under Apache 2.0; Kimi K2.5's license isn't Apache 2.0 in this record |
| Reasoning-trace visibility | Kimi K2.5 | Supports a `reasoning` parameter with reasoning enabled by default; Mistral has no reasoning parameter |
Questions
- Which model is cheaper for a typical workload?
- Kimi K2.5, marginally — a 10M input / 2M output token job costs $7.80 versus $8.00 for Mistral Large 3 2512. The gap comes from Kimi's lower input price ($0.375/M vs $0.50/M) outweighing its higher output price ($2.025/M vs $1.50/M) at this ratio.
- Does either model handle file uploads?
- Only Mistral Large 3 2512. Its input modalities are text, image, and file. Kimi K2.5 supports text and image inputs only, so PDF or raw file attachments need Mistral or a preprocessing step before hitting Kimi.
- How do their context windows compare?
- Identical — both support 262,144 tokens, giving a context ratio of exactly 1.0. Neither has an edge for long-document tasks based on window size alone.
- Which model ranks better for coding benchmarks?
- Kimi K2.5. In Design Arena's codecategories, Kimi ranks 20th (elo 1286) versus Mistral Large 3 2512 at rank 56 (elo 1191). Kimi also has dedicated agent-arena rankings — like rank 2 in Godot game dev — that Mistral has no equivalent entries for.