← Back to feed

Xiaomi’s MiMo-V2.5-Pro entered public beta on April 22 with no pricing change from its predecessor: $1/M input, $3/M output for prompts under 256K tokens, $2/$6 for 256K–1M token inputs. The model scores 54 on the Artificial Analysis Intelligence Index — a five-point gain over V2-Pro’s 49 — tying Kimi K2.6 at the #4-5 global slot, three points behind the three-way tie at 57 held by Claude Opus 4.7, Gemini 3.1 Pro Preview, and GPT-5.4.

The performance story centres on token efficiency. On ClawEval, V2.5-Pro scores 64% Pass^3 using roughly 70K tokens per trajectory — 40-60% fewer than Claude Opus 4.6, Gemini 3.1 Pro, and GPT-5.4 at comparable capability levels. Lower token burn at the same intelligence tier translates directly to cost savings in long-horizon agentic workloads, where token counts accumulate across hundreds of tool calls.

Key Numbers

• Intelligence Index: 54 (AA, v4.0) — global #4-5, tied with Kimi K2.6
• Pricing: $1/$3 per million tokens (≤256K); $2/$6 per million tokens (256K–1M)
• Context window: 1M tokens
• ClawEval: 64% Pass^3 at ~70K tokens/trajectory
• Token cost to run AA Intelligence Index: ~$348 (vs $2,486 for Claude Opus 4.6)

What It Actually Does

Xiaomi published two long-horizon benchmarks in the launch documentation:

SysY Compiler in Rust: Starting from the Peking University Compiler Principles course spec, V2.5-Pro built a complete SysY compiler — lexer, parser, AST, Koopa IR codegen, RISC-V assembly backend, and performance optimisation — in 4.3 hours across 672 tool calls. It scored 233/233 on the course’s hidden test suite. The first compile passed 137/233 tests cold, a 59% cold-start accuracy suggesting the architecture was designed correctly before any test ran. After a refactoring at turn 512 regressed two tests, the model self-diagnosed and recovered.

Full-featured desktop video editor: 8,192 lines of code, multi-track timeline, clip trimming, cross-fades, audio mixing, and export pipeline — produced across 1,868 tool calls over 11.5 hours of autonomous operation.

Xiaomi claims V2.5-Pro can sustain tasks spanning more than 1,000 tool calls while maintaining instruction fidelity and contextual coherence — a threshold where most current models degrade noticeably.

Competitive Position

At AA Index 54 with $1/$3 per million token pricing, MiMo-V2.5-Pro undercuts every model above it on the global leaderboard by a significant margin. Claude Opus 4.7 runs at $15/$75 per million. Gemini 3.1 Pro Preview and GPT-5.4 are comparable in cost-per-call but cost roughly 5-6x more per million output tokens.

The V2.5-Pro is positioned for production agent pipelines — multi-step automation, software engineering tasks, long-context orchestration — where per-call latency and per-token cost determine total job cost more than raw benchmark position. For a team running agents at scale, the token efficiency advantage compounds over millions of calls.

Weights have not been released. The model is API-only, available via Xiaomi’s AI Studio and API Platform. Context-aware pricing kicks in for prompts over 256K tokens, where input cost doubles.