Vertical AI Models
MAI-Code 1.1 Flash falls short on price and performance against DeepSeek
Microsoft's new coding model for GitHub Copilot, MAI-Code 1.1 Flash, launched with a price cut to a quarter of its predecessor. But head-to-head with China's DeepSeek, it reportedly trails on both price and benchmark performance. The "cheap specialized models over frontier chasing" pitch from June's Build is already being tested.
The "cheap specialist" pitch
stumbled on its first comparison
MAI-Code 1.1 Flash began rolling out across GitHub Copilot on August 11 with a sharp price cut — but its performance reputation didn't keep pace.
According to Microsoft AI's own announcement, MAI-Code 1.1 Flash, the coding model for GitHub Copilot, cut pricing to "a quarter" of the previous MAI-Code 1 Flash, claimed a 22% improvement on Terminal-Bench 2.1 via the Copilot CLI and a 15% improvement on .NET tasks, and began its rollout across Copilot surfaces on August 11.
Around the same time, tech outlet The Decoder told a different story. MAI-Code 1.1 Flash reportedly trails China's "DeepSeek-V4-Flash-0731" on both price and coding performance. Line up the numbers and the gap is clear: input pricing runs $0.20 per 1M tokens for MAI versus $0.14 for DeepSeek. The bigger gap is on output — $1.20 for MAI versus $0.28 for DeepSeek, roughly a 4.3x difference.
Coding benchmarks tell a split story too. On SWE-bench Verified, both score an identical 71.6%. But on Terminal-Bench, which evaluates tasks involving terminal operations and multi-step edits, MAI lands at 54.8% (v2.0) versus DeepSeek's 82.7% (v2.1). Even accounting for the version difference, a gap of more than 27 points is hard to wave away.
| MAI-Code 1.1 Flash | DeepSeek V4 Flash-0731 |
|---|---|
| Input price (1M tokens) $0.20 | $0.14 |
| Output price (1M tokens) $1.20 | $0.28 |
| SWE-bench Verified 71.6% | 71.6% (tie) |
| Terminal-Bench 54.8% (v2.0) | 82.7% (v2.1) |
"Cheap specialized models over frontier chasing."
At Build in June, that line was Microsoft's coding-strategy banner.
The cheap math hasn't
caught up to DeepSeek yet
Output token pricing and agentic-task benchmarks. Lined up as numbers, the gap is bigger than rounding error.
This comparison matters because Microsoft's whole pitch was choosing not to chase frontier models. At Build on June 2, 2026, Microsoft unveiled seven in-house models including MAI-Code-1-Flash, framing it as a shift toward "long-term self-sufficiency" away from licensing OpenAI's models. The design logic — cheap, token-efficient, good enough for production — is defensible on its own terms. But this week's reporting exposes how that logic fares when placed directly against an open Chinese model on the same turf.
The output-price gap especially deserves attention. For token-heavy workloads like automated test generation or documentation generation, the gap between $1.20 and $0.28 per 1M tokens compounds fast into monthly costs. That said, Microsoft claims 1.1 cuts token consumption 25% versus the prior generation and streams 25% faster — so judging by per-token price alone, without factoring in actual tokens consumed, doesn't tell the full cost story.
So what should Copilot
model-pickers actually do
The people most affected are engineers who choose coding models day to day, and the PMs who own Copilot's licensing budget.
If you pick models inside Copilot
Don't leave MAI-Code 1.1 Flash as your default coding model. The Terminal-Bench gap tends to show up most in tasks involving terminal operations and multi-step edits, so A/B it against DeepSeek-class alternatives and your existing picks on real tasks, and split usage by task type.
If you manage Copilot cost and licensing
A 4.3x output-price gap compounds fast at team scale. Don't budget off the SWE-bench Verified tie alone — weigh it against agentic-task volume and Terminal-Bench-style numbers before locking in a default model.
The next question is whether
the gap closes in the next release
Microsoft cut output pricing to a quarter and improved Terminal-Bench 2.1 scores by 22% in under six months, going from 1.0 to 1.1 — a fast iteration pace. DeepSeek, meanwhile, keeps shipping too, with "V4-Flash-0731" dated July 31. Expect the back-and-forth to continue for a while. In the meantime, three practical steps make sense.
- Don't default to MAI-Code 1.1 Flash inside Copilot — test alternatives by task type
- Run explicit cost comparisons for output-heavy workloads (test generation, documentation, etc.)
- Track whether the Terminal-Bench gap narrows in the next model update
A benchmark gap doesn't always
equal a felt difference
In fairness, a benchmark score gap doesn't necessarily scale one-to-one with real-world feel. The Terminal-Bench comparison pits MAI's v2.0 result against DeepSeek's v2.1 result, so a same-version re-test would be more conclusive. Microsoft also tunes MAI-Code directly on Copilot's production harness — its file-editing tools, terminal integration, and multi-step task loops — which could produce day-to-day usability differences that generic benchmarks don't capture either way. Given how much ground Microsoft covered between 1.0 and 1.1 in under six months, there's real room for this gap to narrow in the next release.