Microsoft shipped a new small coding model into GitHub Copilot on 12 August 2026, and the number it wants read is 25 percent. That is how much more token-efficient Microsoft says MAI-Code-1.1-Flash is compared with the version it launched two months earlier at Microsoft Build. The number it did not publish in the announcement itself is 82.7, which is where a rival model lands on the same coding benchmark Microsoft used to sell the release.
Microsoft’s own post frames the update almost entirely against its predecessor, MAI-Code-1-Flash. The company says the new model needs 25 percent fewer tokens to finish a task, streams answers 25 percent faster, and costs about a quarter of what the June version charged. On GitHub Copilot’s command-line interface, Microsoft reports a 22 percent gain on Terminal-Bench 2.1 and a 15 percent gain on .NET-specific tasks, plus a production-side bump: developers kept 4 percent more of the model’s suggested code and returned to use it 9 percent more often. Microsoft credits the gains to training the model in what it calls “hundreds of thousands of reinforcement-learning environments” inside Copilot. No competing model is named anywhere in the post.
The Decoder, a technology outlet that tracks AI model releases, published a same-day analysis built from benchmark numbers Microsoft had posted separately in the model’s technical card rather than in the announcement. Those figures tell a more mixed story.
| Model | SWE-bench Verified | Terminal-Bench 2.1 |
|---|---|---|
| MAI-Code-1.1-Flash | 72.6% | 62.9% |
| MAI-Code-1-Flash (June) | 71.6% | 51.7% |
| Claude Haiku 4.5 | 69.8% | 49.4% |
| GPT-5.4 Mini | 69.2% | 60.7% |
| DeepSeek-V4-Flash-0731 | not published | 82.7% |
On SWE-bench Verified, Microsoft’s new model does edge past its own predecessor and past the comparable mini models from Anthropic and OpenAI, though the margins are one to three percentage points. DeepSeek, the Hangzhou-based lab whose open-weight releases have repeatedly undercut frontier pricing, did not publish a comparable SWE-bench figure. On Terminal-Bench 2.1, the same benchmark Microsoft highlighted with its 22 percent improvement claim, DeepSeek-V4-Flash-0731 scores 82.7 percent against Microsoft’s 62.9 percent, a gap of nearly 20 points.
Pricing compounds the problem. The Decoder compiled per-million-token costs showing DeepSeek undercutting Microsoft on every line item: DeepSeek charges $0.14 for input tokens, $0.0028 for cached input, and $0.28 for output, against Microsoft’s $0.20, $0.02, and $1.20. Claude Haiku 4.5 is the most expensive of the three, at $1.00, $0.10, and $5.00. Microsoft’s output-token price runs more than four times DeepSeek’s.
The Decoder’s read is structural rather than accusatory: Microsoft’s announcement leaned on percentage gains against its own prior release and avoided naming any competitor, while the head-to-head numbers sat one click away in the model card. The outlet places the release inside a pattern it has tracked before, Microsoft substituting its own MAI models for OpenAI’s and Anthropic’s inside Copilot in ways that protect Microsoft’s margin on every subscription regardless of which underlying model performs best on a given task.
Both things are true at once. MAI-Code-1.1-Flash is a real upgrade over the model it replaces and beats the small models Anthropic and OpenAI sell for the same job on SWE-bench Verified. It is also priced and benchmarked well behind DeepSeek’s comparable model on the specific test Microsoft chose to promote. Teams routing Copilot CLI workloads through the default model should benchmark DeepSeek-V4-Flash-0731 directly against MAI-Code-1.1-Flash on their own terminal tasks before treating Microsoft’s default as the fastest or cheapest option on the table.
Microsoft (microsoft.ai) and The Decoder both published their respective claims and benchmark analysis on 12 August 2026.
