Xiaomi's MiMo Code claims a 65%+ win rate over Claude Code on tasks longer than 200 steps — but the benchmarks are self-reported

The claim that matters

Past 200 execution steps, Xiaomi says MiMo Code beats Claude Code more than 65% of the time. Below that threshold, the two tools split roughly even. That asymmetry is the whole argument: the gains aren't about raw model quality, they're about what happens to an AI coding agent when a task gets long enough that context management becomes the binding constraint.

The caveat that matters equally: every number in this article comes from Xiaomi. MiMo Code does not yet appear on the official SWE-bench or Terminal Bench 2 leaderboards, and cross-comparing self-run figures against leaderboard submissions carries real configuration risk.

What MiMo Code actually does differently

Most AI coding agents degrade over long sessions because their context windows — the working memory available to the model at any moment — fill up. Earlier decisions, file structures, and task state get compressed or dropped. Developers end up re-explaining their own projects.

MiMo Code attacks this with a four-layer persistent memory system: a project-level `MEMORY.md` file, session checkpoints, scratch notes, and per-task progress logs, all indexed via SQLite FTS5 full-text search for fast retrieval. Crucially, a separate "checkpoint-writer" subagent handles note-taking in parallel, so the primary coding agent doesn't have to pause its work to update state. When the context window approaches its limit, the system rebuilds from structured checkpoints rather than relying on lossy compression.

A `/dream` command, run roughly every seven days, reviews historical sessions, deduplicates them, and compresses them into long-term memory. A "distill" function mines past sessions for repeated workflows that can be automated — an approach similar to what OpenAI and Anthropic have pursued in their own recent model updates.

What the benchmarks actually show

Xiaomi reports the following head-to-head figures for MiMo Code + MiMo-V2.5-Pro versus Claude Code + Claude Sonnet 4.6:

- **SWE-bench Verified:** 82% vs. 79% - **SWE-bench Pro:** 62% vs. 55% - **Terminal Bench 2:** 73% vs. 69%

Running the same MiMo-V2.5-Pro model through both harnesses, MiMo Code scores five points higher on SWE-bench Pro (62% vs. 57%) and five points higher on Terminal Bench 2 (73% vs. 68%), suggesting the scaffolding itself — not just the model — is doing measurable work.

The selective comparison is worth noting. Xiaomi benchmarked exclusively against Claude Code. On the official Terminal Bench 2 leaderboard at tbench.ai, OpenAI's Codex CLI running GPT-5.5 scores 82.2% — roughly nine points above MiMo Code's self-reported 73%. On SWE-bench Pro, the picture reverses: OpenAI reports GPT-5.5 at 58.6%, below MiMo Code's claimed 62%. Neither comparison is clean, and MiMo Code's absence from official leaderboards makes direct verification impossible for now.

Xiaomi's own human evaluation — 576 developers, 474 private repositories, 1,213 judged head-to-head pairs — is the most interesting data point, precisely because standard benchmarks, as Xiaomi concedes, "still measure one-shot problem-solving ability" and don't capture multi-session design goals.

The pricing angle

For many developers, the more immediate draw may be the bundled model access. MiMo Code ships with free, limited-time access to MiMo-V2.5 — a 310-billion-parameter sparse mixture-of-experts model with a one-million-token context window, priced at $0.40 per million input tokens and $2.00 per million output tokens when billed. The larger MiMo-V2.5-Pro runs $1.00/$3.00 per million tokens up to 256K context. Both sit well below comparable offerings from Anthropic, OpenAI, and Google.

What enterprises should weigh

MiMo Code's MIT license, inspectable OpenCode lineage, and bring-your-own-model support make it a low-friction evaluation candidate. The persistent memory architecture addresses a genuine pain point. The countervailing considerations are equally real: "free for a limited time" model access routes code context through Xiaomi's servers — a non-starter for organizations with data-residency or IP policies — the benchmark edge is unverified, and V0.1.0 signals exactly the maturity level it implies. Organizations subject to U.S. government procurement restrictions on Chinese technology vendors should factor that context into any adoption decision.