Microsoft Just Slashed Copilot Pricing 73% — And Snuck Its Own Model Into the Driver's Seat

If you build software for a living, the only line you need to read from Microsoft's August 11 announcement is this one: the new MAI-Code-1.1-Flash model costs one quarter of what the previous version did. That is a 73% price cut for the underlying AI powering GitHub Copilot, delivered just ten weeks after Microsoft first shipped the original. The upgrade cycle says more about the competitive panic gripping the AI coding market than it does about any single metric. Microsoft is iterating on first-party models at a pace that suggests a company racing to catch a moving target. MAI-Code-1-Flash, introduced at Build in June, is already being put out to pasture. The successor pulls in 22% better scores on Terminal-Bench 2.1 within the Copilot CLI, pushes .NET coding tasks up 15%, and streams tokens 25% faster. The company also claims a 4% bump in code acceptance in production and a 9% increase in how often developers come back to the tool. Oh, and the new model can see. Native vision support means Copilot can now accept screenshots, diagrams, and UI mockups alongside code. A practical addition, particularly for developers working through design handoffs without leaving the IDE.

The Fine Print on That 73% Price Cut

Microsoft's headline number is accurate as far as it goes, but how the discount translates to your monthly bill depends on subscription type. Under the usage-based pricing that GitHub rolled out in June, MAI-Code-1.1-Flash is billed at $0.20 per million input tokens and $1.20 per million output tokens. Cached input costs $0.02 per million tokens. That is down from the $0.75 input and $4.50 output rates of its predecessor. For annual subscribers, the model carries a 0.25x premium request multiplier — meaning each request burns one-quarter of a standard AI credit instead of a full one. Microsoft's official line was that improvements in training and serving efficiency allowed them to offer a stronger, faster model at one quarter of the price of 1.0 — and pass those savings reliably to customers.

The Architecture Under the Hood

According to the model card, MAI-Code-1.1-Flash is a 138-billion-parameter sparse Mixture-of-Experts model. The catch is that only about 5 billion parameters are active per request. This is the same design philosophy that has come to dominate the AI coding space: keep the compute footprint small during inference while retaining a large enough total parameter pool to hold knowledge. The training run spans March through August 2026, using a compressed checkpoint of MAI-Thinking-1 with approximately 2 million synthetic agent tasks during intermediate training, followed by reinforcement learning across more than 150,000 environments. Context window is 256,000 tokens. Microsoft describes the tuning focus as agentic coding through the GitHub Copilot harness, repository questions, refactoring, and tool use. The corporate language is predictable — ship, learn, improve, repeat — but the efficiency gains are real and measurable. Twenty-five percent fewer tokens to complete comparable tasks forms the basis for the pricing change. If the math holds in production, Microsoft is no longer selling a marginally cheaper commodity. It is selling a genuinely cheaper way to run coding assistance at scale.

The DeepSeek Problem

This is where the narrative gets messy, and where Microsoft's carefully worded announcement starts to feel like damage control rather than a victory lap. Independent benchmarks tell a different story than Microsoft's internal metrics. On Terminal-Bench 2.1, MAI-Code-1.1-Flash scores 62.9%. DeepSeek-V4-Flash-0731 — the open-weight model from China that has been eating everyone's lunch since late July — scores 82.7% on the same benchmark. On SWE-bench Verified, Microsoft's model comes in at 72.6%, slightly edging out Claude Haiku 4.5 at 69.8% and GPT-5.4 mini at 69.2%, according to data reported by THE DECODER. Price tells a similar story. DeepSeek-V4-Flash undercuts Microsoft on every line item. Input comes in at $0.14 per million tokens versus Microsoft's $0.20, and output at $0.28 versus Microsoft's $1.20. The gap on cached input is even wider, with DeepSeek charging just $0.0028 per million tokens. A Chinese tech blogger, cocoloop.cn, put it bluntly during the rollout discussion: Microsoft's new MAI Code 1.1 Flash boasts 25% higher token efficiency and costs a quarter of the previous version. But looking at benchmarks, both price and performance are crushed by Deepseek V4 Flash. This is the same old playbook.

A Strategy Whitepaper or a Margin Play?

The developer community has not been shy about drawing conclusions. On Hacker News, the predecessor model sparked a 540-point thread with 255 comments, and the 1.1 release is generating similar debate. The most pointed critique came from THE DECODER, which observed that Microsoft buries the benchmark results in the model card and ships press releases touting vague improvement metrics like code survival rose 4% and return visits increased 9% while skipping direct comparisons with competitors. The editorial conclusion there was less polite: it is hard to see the point of Microsoft's MAI models when they trail OpenWeight alternatives on both price and performance, the same models Microsoft keeps saying it is a massive fan of. There is a broader strategic context worth noting. Microsoft has reportedly been quietly replacing OpenAI and Anthropic models in Copilot with cheaper MAI alternatives for months. Bloomberg confirmed the company is routing tens of thousands of prompts away from OpenAI and Anthropic models to MAI models. The PowerPoint team claims MAI models cut compute costs by up to 85% compared to using OpenAI's image models. Analysts have flagged that Copilot's old fixed-price model was losing money at scale — the more users, the bigger the losses. The shift to usage-based pricing in June, combined with the aggressive cost reduction on first-party models, points to one conclusion: Microsoft is playing defense on margins.

Developers Feel the Squeeze

The pricing overhaul has not gone down smoothly. Long-time Greg, a developer on Reddit, posted his monthly bill going from roughly $29 to nearly $750. Another user reported a jump from $50 to around $3,000. Annual contract holders discovered hidden rate multipliers — choosing advanced reasoning models consumed credits four times faster than simple requests. Claude Opus 4.7 interactions went from 7.5 premium requests to a 27x multiplier penalty. One developer reported burning through an entire month's quota in 72 hours of light usage. GitHub defended the change as necessary for the economics to work, but the community response has been hostile. GitHub Community forums carry formal complaints alleging the pricing change violates annual subscription agreements. The deeper irritation is model lock-in. Free and Student plan users lost manual model selection entirely on June 24. They now get whatever automatic routing Microsoft decides, and developers have reported in GitHub Discussion 204673 that Auto mode has been steering requests to MAI-Code-1-Flash with noticeably worse results for complex tasks. One senior developer wrote that we are moving backwards, requesting metrics on whether Auto mode actually satisfies user needs.

Why Not Just Use DeepSeek?

The obvious question — one that follows Microsoft executives into every interview — is why the company is sinking resources into a more expensive, weaker proprietary model when an open-weight alternative delivers better performance at lower cost. Part of the answer may be political, not technical. Microsoft has thus far resisted offering DeepSeek in its core products, though reports from Axios suggest the company has been evaluating adding DeepSeek V4 to Copilot Cowork as a low-cost option. If those talks materialize, it would mark the first time a major U.S. tech company embedded a Chinese large language model into a core enterprise product. Microsoft has reportedly studied routing simpler tasks to DeepSeek while keeping the most complex or sensitive workloads on OpenAI models. Anthropic's recent security dispute with Alibaba — in which Claude Code was accused of containing backdoors and subsequently banned from several Chinese tech firms — has added another layer of friction to cross-border AI deployments. But from a pure engineering and economics standpoint, the case for DeepSeek is starting to look irrefutable.

The Forward Path Is Uncomfortable

Microsoft is rolling out MAI-Code-1.1-Flash across every significant Copilot surface — Visual Studio Code, Visual Studio, JetBrains IDEs, the CLI, cloud agents, GitHub Mobile, Eclipse, and Xcode. Free and Student plan users get it through automatic model selection. Pro, Pro+, Max, Business, and Enterprise users can manually select it if they prefer. Business and Enterprise plan administrators must explicitly enable the policy — it is off by default. MAI-Code-1-Flash gets deprecated on September 10, 2026. The rapid iteration cadence — two major flash model releases in ten weeks — gives Microsoft a credible feedback loop, processing real production data from millions of Copilot sessions. Their broader model roadmap includes MAI-Cyber-1-Flash, expected to halve MDASH costs, and the recently released MAI-Image-2.6, which reached second place on the Arena leaderboard with a score of 1336. But the uncomfortable fact remains: on raw performance, Microsoft's own flagship coding model trails a Chinese open-weight competitor by roughly 20 points on a benchmark that people actually care about. DeepSeek-V4-Flash was trained to be the cheapest per-token model on the market, hitting 82.7% on Terminal Bench 2.1 while accepting a 256K context window and high-throughput agentic workloads. It lacks vision support, but for the coding tasks where it matters, developers have been perfectly happy to route around that limitation. Whether Microsoft can close the gap in future iterations, or whether the margin play is the whole game, remains to be seen. The company has opened a GitHub Community Discussion thread for developer feedback on the new model. It is an invitation worth accepting — the developers reading the benchmarks likely understand the stakes better than the marketing team writing the press release.