DeepSeek V4 Flash Crushes Microsoft's MAI Code 1.1 Flash

Microsoft has quietly released MAI Code 1.1 Flash, the newest coding model feeding GitHub Copilot, and early benchmark numbers paint an awkward picture: the model trails DeepSeek's V4 Flash on both price and performance, the two numbers developers actually care about most.

The announcement itself reads like a solid upgrade. Microsoft says the new model writes better code than its June predecessor, runs 25 percent more token-efficiently, and costs a quarter as much to use. Developers accepted 4 percent more of its suggested code, and the company says the model was trained on hundreds of thousands of reinforcement-learning environments built inside Copilot itself.

Those claims are all relative to Microsoft's own previous model, though. Against the open-weight competition, the story changes quickly.

Benchmarks show a wide gap

Microsoft's own figures show MAI Code 1.1 Flash edging out its predecessor and the mini-models from Anthropic and OpenAI. On SWE-bench Verified, it scores 72.6 percent, ahead of MAI Code 1 Flash at 71.6 percent, Claude Haiku 4.5 at 69.8 percent, and GPT-5.4 mini at 69.2 percent.

Terminal Bench 2.1, which tests real-world agentic coding workflows, is where the gap becomes uncomfortable. MAI Code 1.1 Flash improved from its predecessor's 51.7 percent to 62.9 percent, comfortably ahead of Haiku 4.5's 49.4 percent. DeepSeek V4 Flash, meanwhile, sits at 82.7 percent, a roughly 20-point lead over Microsoft's new model.

• SWE-bench Verified: MAI Code 1.1 Flash 72.6%, predecessor 71.6%, Haiku 4.5 69.8%
• Terminal Bench 2.1: MAI Code 1.1 Flash 62.9%, predecessor 51.7%, DeepSeek V4 Flash 82.7%
• DeepSeek did not publish a SWE-bench Verified score, keeping its strongest cards in hand

Microsoft buries these results deep in the model card and leads the official announcement with vaguer metrics instead, such as code survival rising 4 percent and return visits up 9 percent, with no direct comparison against DeepSeek anywhere in the post.

The price gap is just as clear

MAI Code 1.1 Flash is positioned as a budget model, and it is cheaper than what Microsoft shipped in June. But cheaper than before is not the same as cheapest, and on raw pricing DeepSeek still wins every row of the comparison.

Output tokens are where it hurts most: Microsoft charges $1.20 per million, more than four times DeepSeek's $0.28. Cost per token does not tell the whole story without factoring in usage efficiency, but with DeepSeek also ahead on the benchmark that matters, the gap in its favor is likely significant either way.

The strategy question writes itself. Microsoft has spent months praising open-weight models, even describing itself as a champion of open AI, yet it keeps investing in proprietary in-house models that trail the freely available alternatives on the metrics developers check first.

The likely answer is margin math. Microsoft recently reshuffled its Copilot lineup, swapping out OpenAI and Anthropic models for its own cheaper MAI alternatives to cut costs. The trade-off was worse performance for better margins, and MAI Code 1.1 Flash fits the same pattern.

Customers in the GitHub ecosystem can still pick different models per app and use case, but Microsoft will almost certainly make its own models the default over time. That alone locks up a massive share of the market, since most users never actively choose a specific AI model anyway.

For developers budgeting API calls at scale, the takeaway is simple. Microsoft's 25 percent efficiency claim is a starting point for your own testing, not a reason to switch. DeepSeek V4 Flash remains cheaper per token and stronger on real-world coding benchmarks, and until Microsoft's in-house models close that gap, the open-weight alternative keeps winning on the numbers that matter.