Kimi K2.5 vs Qwen3 Max
Moonshot AIQwenboth via one key, provider price + 3%
Both models shipped with the same 262,144-token context window, but that's where the similarity ends. Kimi K2.5 (Moonshot AI, released 2026-01-27) is a multimodal model with a self-directed agent design; Qwen3 Max (Qwen, released 2025-09-23) is text-only with a June 2025 knowledge cutoff and adds prompt caching. Across every Design Arena coding category both models were tested in, Kimi K2.5 ranks higher. On price, Kimi K2.5's input tokens cost 48% of what Qwen3 Max charges.
Spec vs spec
| Spec | Kimi K2.5 | Qwen3 Max |
|---|---|---|
| Context window | 262K | 262K |
| Max output | — | 33K |
| Input modalities | text, image | text |
| Output modalities | text | text |
| Knowledge cutoff | — | Jun 30, 2025 |
| Released | Jan 27, 2026 | Sep 23, 2025 |
| Reasoning | optional | optional |
Pricing
Per 1M tokens. Provider price plus the flat 3% fee — the sum is what you pay.
moonshotai/kimi-k2.5
Input · 1M tokens
$0.375 + 3%$0.386
Output · 1M tokens
$2.02 + 3%$2.09
FEE — FLAT, EVERY MODEL3%
qwen/qwen3-max
Input · 1M tokens
$0.780 + 3%$0.803
Output · 1M tokens
$3.90 + 3%$4.02
Cache read · 1M tokens
$0.156 + 3%$0.161
Cache write · 1M tokens
$0.975 + 3%$1.00
FEE — FLAT, EVERY MODEL3%
One workload, priced on both
10M input + 2M output tokens at each model's price, flat 3% fee included.
moonshotai/kimi-k2.5Cheaper
$8.03
$7.80 provider + 3%
qwen/qwen3-max
$16.07
$15.60 provider + 3%
Benchmarks
Design Arena categories where both models have results. Higher Elo and lower rank win.
| Kimi K2.5 | Qwen3 Max | |||
|---|---|---|---|---|
| Category | Elo | Rank | Elo | Rank |
| 3D | 1286 | #22 | 1150 | #62 |
| ASCII art | 1214 | #17 | 1175 | #32 |
| Code | 1286 | #20 | 1159 | #66 |
| Data viz | 1270 | #21 | 1149 | #64 |
| Game dev | 1272 | #23 | 1160 | #62 |
| SVG | 1210 | #25 | 1069 | #60 |
| UI components | 1290 | #19 | 1132 | #67 |
| Websites | 1291 | #16 | 1161 | #66 |
Head-to-head preference voting. How we filter and rank
Pricing math on a real workload
On OpenKey, Kimi K2.5 costs $0.38625/M input and $2.08575/M output (provider price $0.375/$2.025 x 1.03 fee). Qwen3 Max costs $0.8034/M input and $4.017/M output (provider $0.78/$3.9 x 1.03). Run a 10M-input / 2M-output workload and Kimi K2.5 costs $7.80 versus Qwen3 Max's $15.60 — exactly double. The input price ratio alone is 0.48, meaning Kimi K2.5's input tokens cost less than half of Qwen3 Max's. If your workload leans input-heavy (RAG, long documents, agent context), that gap compounds fast. Qwen3 Max does offer cache-read pricing at $0.156/M and cache-write at $0.975/M — Kimi K2.5 has no cache pricing listed, so if you repeatedly hit the same long context, Qwen3 Max's caching can partially close the cost gap. Both run on OpenKey with one key and the same flat 3% fee on top of provider list price.
Coding and agent benchmarks
Design Arena data is one-sided here. In the agents arena, Kimi K2.5 ranks #2 in Godot game dev (elo 1254, 59.5% win rate) and #14–20 across fullstack, mobile apps, web apps, and Android native. Qwen3 Max has no agents-arena scores in this dataset — only models-arena results, all in the rank 60–67 range: website (rank 66, elo 1161), code categories (rank 66, elo 1159), UI components (rank 67, elo 1132), SVG (rank 60, elo 1069). Kimi K2.5's matching models-arena scores are consistently stronger: code categories rank 20 (elo 1286), website rank 16 (elo 1291), UI components rank 19 (elo 1290). For anything resembling agentic coding or UI generation, Kimi K2.5 is the clear performer on this data.
Modality and tooling differences
Kimi K2.5 accepts text and image input; Qwen3 Max is text-only. If your app needs to read screenshots, diagrams, or UI mockups as part of a coding task, Qwen3 Max can't do it — only Kimi K2.5 supports that input path. On the tooling side, Kimi K2.5 exposes 19 supported parameters including reasoning, top_k, min_p, and logit_bias, versus Qwen3 Max's 11 parameters. Qwen3 Max caps max output at 32,768 tokens; Kimi K2.5 has no listed max-completion cap. Neither model's reasoning is mandatory, but Kimi K2.5 has reasoning enabled by default — useful if you want chain-of-thought traces without extra config.
When to pick each
Use Kimi K2.5 when cost per token and coding-agent quality matter — it's cheaper on every axis and ranks higher on every shared Design Arena category. Use it too when the task involves images alongside text. Use Qwen3 Max when you need a fixed knowledge cutoff (2025-06-30) for reproducibility, when prompt caching will meaningfully cut costs on repeated long contexts, or when you're already tuned to Qwen's instruction-following behavior for multilingual tasks. Context length is a non-factor — both sit at 262,144 tokens, a 1.0 ratio.
Which model for which job
| Use case | Pick | Why |
|---|---|---|
| Coding agents / fullstack dev | Kimi K2.5 | Ranks #14 fullstack, #2 Godot game dev vs Qwen3 Max's rank 60+ across all models-arena coding categories |
| High-volume input workloads (RAG, long docs) | Kimi K2.5 | Input price ratio of 0.48 — less than half Qwen3 Max's per-token input cost |
| Repeated long-context calls | Qwen3 Max | Only Qwen3 Max offers cache-read pricing at $0.156/M to cut repeat-context costs |
| Image-in-context tasks (screenshots, mockups) | Kimi K2.5 | Only model here with image input modality |
| Reproducible outputs pinned to a knowledge date | Qwen3 Max | Has a stated knowledge cutoff of 2025-06-30; Kimi K2.5 lists none |
| Budget-constrained 10M-in/2M-out batch job | Kimi K2.5 | Costs $7.80 for that workload vs $15.60 for Qwen3 Max |
Questions
- Which model is cheaper on OpenKey?
- Kimi K2.5, by a wide margin. It runs $0.38625/M input and $2.08575/M output versus Qwen3 Max's $0.8034/M input and $4.017/M output. A 10M-input/2M-output job costs $7.80 on Kimi K2.5 versus $15.60 on Qwen3 Max — exactly 2x.
- Does either model support images?
- Only Kimi K2.5. Its input modalities are text and image; output is text-only. Qwen3 Max is text-to-text only, with no image input support in this catalog data.
- How do they compare on coding benchmarks?
- Kimi K2.5 outranks Qwen3 Max in every shared Design Arena models-arena category: code categories (rank 20 vs 66), website (rank 16 vs 66), UI components (rank 19 vs 67), and SVG (rank 25 vs 60).
- Is context length a differentiator here?
- No. Both models have a 262,144-token context window, giving a context ratio of 1.0. The decision comes down to price, modality, and benchmark performance, not context size.