Xiaomi MiMo Code Enters AI Coding Race Claiming 200 Step Endurance Over Claude Code

Xiaomi has entered the competitive AI coding model arena with MiMo Code, an agent that claims to sustain effective coding sessions beyond 200 steps. The claim directly challenges existing models such as Claude Code, which reportedly stalls around 30 steps during production refactoring tasks.

According to benchmarks reported by The New Stack, this endurance gap is a critical limitation in current coding agents. Most agents can build a simple application during a lunch break but lose coherence on long refactoring tasks. MiMo Code asserts it maintains accuracy and context across more than 200 steps, a threshold no major competitor has publicly claimed.

The Cost of a Coding Test

The cost of running these models varies sharply. In one documented test, a coding session using Anthropic's Claude Fable model cost $9. The same task completed with OpenAI's GPT-5.5 cost only $1.50. This price difference is driving developers to adopt model triage, selecting cheaper models for routine tasks and reserving expensive models for complex critical work.

• MiMo Code claims 200+ step endurance, surpassing Claude Code's typical 30 step limit.
• Claude Fable cost $9 per test; GPT-5.5 cost $1.50 per test.
• Anthropic's Fable 5 and Mythos 5 models remain suspended due to government review.

Regulatory Hurdles and Market Uncertainty

Meanwhile, Anthropic faces regulatory pressure. Its Fable 5 and Mythos 5 flagship models were disabled after U.S. government concerns. The company said the ball is in Anthropic's court, indicating ongoing negotiations. This suspension leaves a gap in the high-end coding assistant market, which competitors like Xiaomi and OpenAI are eager to fill.

What comes next? The endurance benchmark is becoming a key differentiator. Developers now prioritize models that can maintain context over long sessions rather than just speed or accuracy. Xiaomi's entry with MiMo Code signals that the next phase of AI coding competition will focus on sustained reasoning, not just single-turn responses.