MiniMax M2.5 vs Qwen3 235B A22B Instruct 2507

Pricing math

MiniMax M2.5 costs $0.12/M input and $0.48/M output at the provider level; on OpenKey that's $0.12 × 1.03 = $0.1236/M input and $0.48 × 1.03 = $0.4944/M output. Qwen3 235B A22B Instruct 2507 is cheaper on both ends: $0.09/M input ($0.0927 on OpenKey) and $0.10/M output ($0.103 on OpenKey).

Run the numbers on a 10M-input/2M-output workload and the gap is stark: MiniMax M2.5 costs $2.16, Qwen3 235B costs $1.10 — MiniMax is 1.33x more expensive on input price alone, and roughly 2x more expensive on this specific workload because its output price is nearly 5x higher per token. If you're processing large batches and quality differences don't matter for the task, Qwen3 is the clear cost winner.

Coding and design benchmarks

Design Arena scores separate these two models clearly. MiniMax M2.5 ranks in the 26-40 range across categories (3d: rank 31, elo 1246; codecategories: rank 29, elo 1256; website: rank 26, elo 1265; svg: rank 26, elo 1208). Qwen3 235B ranks 77-94 across the same categories (codecategories: rank 81, elo 1088; gamedev: rank 94, elo 1018; website: rank 83, elo 1101).

Win rates tell the same story: MiniMax M2.5 hits 51-58% win rate across categories, Qwen3 235B sits at 35-48%. If the task is UI generation, game dev prototyping, or website scaffolding, MiniMax M2.5's benchmark lead is not marginal — it's a different tier.

Context and output limits

Qwen3 235B A22B Instruct 2507 has the longer context window at 262,144 tokens versus MiniMax M2.5's 204,800 tokens — a context ratio of 0.78, meaning M2.5 holds about 78% as much context as Qwen3. But MiniMax M2.5 flips that on max output: 196,608 tokens versus Qwen3's 16,384. If your workload needs long generated output (large code files, long-form docs), M2.5's 12x larger output ceiling matters more than the input context gap. If you're stuffing huge documents into the prompt and only need short answers back, Qwen3's larger input window wins.

Reasoning and tool support

MiniMax M2.5 runs with mandatory reasoning and exposes `reasoning_effort` and `include_reasoning` as supported parameters — useful for agent workflows where you want visibility into the model's intermediate steps. Qwen3 235B A22B Instruct 2507 has no reasoning mode; it supports the standard parameter set (temperature, top_p, tool_choice, tools, structured_outputs) but skips reasoning-specific controls entirely. If your pipeline depends on chain-of-thought tracing or reasoning-effort tuning, MiniMax M2.5 is the only one of the two that supports it.

Questions

Which model is cheaper to run at scale?

Qwen3 235B A22B Instruct 2507 is cheaper: $0.09/M input and $0.10/M output at the provider level versus MiniMax M2.5's $0.12/M input and $0.48/M output. On a 10M-input/2M-output workload, Qwen3 costs $1.10 total versus MiniMax M2.5's $2.16.

Which model has a bigger context window?

Qwen3 235B A22B Instruct 2507 supports 262,144 tokens of context versus MiniMax M2.5's 204,800 tokens — a context ratio of 0.78. But MiniMax M2.5 supports 196,608 max output tokens versus Qwen3's 16,384, so the advantage flips depending on whether you need long input or long output.

Which model performs better on coding tasks?

MiniMax M2.5 leads on Design Arena's codecategories benchmark with rank 29 and 1256 elo, versus Qwen3 235B's rank 81 and 1088 elo. MiniMax also wins across every other category tested, including website (rank 26 vs rank 83) and gamedev (rank 33 vs rank 94).

Does either model support reasoning mode?

MiniMax M2.5 has mandatory reasoning and exposes reasoning_effort and include_reasoning as supported parameters. Qwen3 235B A22B Instruct 2507 has no reasoning field and doesn't support these parameters — it's a standard instruction-tuned model.