A Chinese open model showed up inside GitHub Copilot this week, one click from OpenAI
GitHub Copilot now serves Kimi K2.7 Code, an open model from China's Moonshot, generally available to millions of developers, while z.ai ships its own GLM-5.2 agent harness and even OpenAI ships a Claude Code plugin. Why the coding tool is going model-neutral, and what a commodity shelf means for the closed labs.
A Chinese open model showed up inside GitHub Copilot this week, one click away from OpenAI.
It's Thursday, July second. Here's the rundown: how the open models finally got their distribution, why the coding tool is splitting from the model, and what that does to the closed labs.
A developer opening Copilot this week found a new name in the model dropdown. Kimi K2.7 Code, an open model from the Chinese lab Moonshot, generally available. Microsoft is serving it to the largest population of professional developers on earth.
Nine days after an open model caught Claude on a benchmark, the question of whether these models could reach Western developers has its answer. They reached them through Microsoft.
And it's not a one-off. The same week, z.ai shipped ZCode, its own agent harness built on GLM 5.2, a straight answer to Claude Code. And OpenAI, of all vendors, shipped a plugin to run its Codex agent from inside Claude Code.
Three moves, one direction. The coding tool is decoupling from the model it runs. The harness is turning into neutral ground, and neutral ground is the best distribution a cheap open model could ask for.
The pushback is that most enterprises leave Copilot on the default, and listing Kimi is just a procurement checkbox nobody uses.
Fair. But look at what actually moves a default. A model that costs a fraction as much, clears the bar on everyday coding, and sits one dropdown away is what a cost-conscious platform team switches to once the finance review lands. That's how a default changes, one budget meeting at a time.
This is the distribution story we've tracked all month, running backwards. When Anthropic put Claude in Slack, the lesson was that owning the channel beats owning the model.
Copilot turns that logic against the closed labs. The tool most developers use is model-neutral, and a neutral tool is a commodity shelf. OpenAI's models now share that shelf with a Chinese open model that costs a tenth as much.
For anyone building: treat the model as a swappable part and invest in the harness around it. Keep two models wired in, route by cost and task.
Run an open model against a real workload this week and measure the gap to your default. On routine coding, it may be smaller than the price difference is.
To the tape. We're watching Microsoft, Moonshot, Zhipu, and OpenAI. Microsoft's the interesting one: making Copilot a model marketplace hedges its OpenAI dependence. Whichever model wins, the distribution layer is Microsoft's.
We covered our GLM short to watch nine days ago, and ZCode is why. An owned harness is the one place an open-weights lab can capture the value the weights give away. If it pulls developers the way Claude Code did, Zhipu has a business, not just a leaderboard spot.
As always, the tape is the desk's scorecard, not advice.
Quick break — two from the desk.
One we know well: vote dot direct. If you're on an H O A or a board, it runs your elections digitally — secure, verifiable, no paper, no clipboard in the lobby. Point your council to vote dot direct.
And if this is your ten minutes of A I for the day, get the written edition too. The full wire, free, every morning — leave your email at nextbig dot dev.
Our call: within six months, a Chinese open model becomes the model most sessions actually run in at least one big Western developer tool, either set as the default or winning usage outright, not just sitting as one more option in the menu.
What proves us wrong: if by January second no major Western developer tool has made a Chinese open model its default, or reported it as the most-used model on its platform.
The developers who come out ahead over the next year are the ones who stopped marrying a model. GitHub just made the divorce a dropdown. That's the rundown.
A developer opening GitHub Copilot this week found a new name in the model dropdown: Kimi K2.7 Code, an open model from the Chinese lab Moonshot, now generally available. Microsoft is serving it to the largest population of professional developers on earth, one click away from the OpenAI models Copilot was built on. Nine days after an open Chinese model caught Claude on a working benchmark, the question of whether those models could reach Western developers has its answer. They reached them through Microsoft.
It is not a one-off. The same week, z.ai shipped ZCode, its own agent harness built around GLM 5.2, a direct answer to Claude Code with the open model underneath. And OpenAI, of all vendors, published a plugin that drives its Codex agent from inside Claude Code. Three moves, one direction: the coding tool is decoupling from the model it runs. The harness is turning into neutral ground, and neutral ground is the best distribution a cheap open model could ask for.
Be fair to the skeptical read. Most enterprises will leave Copilot on its default model, and offering Kimi is partly a procurement checkbox, the option a big customer asks for and never uses. Grant all of it. Then look at what actually moves a default. A model that costs a fraction as much, clears the bar on everyday coding, and sits one dropdown away is what a cost-conscious platform team reaches for once the finance review lands. Getting listed is how the open model gets through the door. That is how a default changes, one budget meeting at a time.
This is the distribution story we have tracked all month, running backwards. When Anthropic put Claude inside Slack, the lesson was that owning the channel beats owning the model, because the tool you control is the moat. Copilot and ZCode turn that same logic against the closed labs. The tool most developers actually use is model-neutral, and a neutral tool is a commodity shelf. OpenAI's models now share that shelf with a Chinese open model that costs a tenth as much, and the lab that does not own the shelf does not decide what sits at eye level.
For anyone building, the move is to treat the model as a swappable part and put your investment in the harness around it. A workflow welded to one vendor's model is paying a premium you can now design away: wire your tools to a neutral layer, keep two models configured, route by cost and task. Run Kimi K2.7 Code or a GLM 5.2 harness against a real workload this week and measure the gap to your default; on routine coding it may be smaller than the price difference is. The developers who come out ahead over the next year are the ones who stopped marrying a model. GitHub just made the divorce a dropdown.
Kimi K2.7 Code is generally available in GitHub Copilot
Microsoft added Moonshot's open Kimi K2.7 Code to Copilot as a first-class, generally available model, sitting in the same dropdown as the OpenAI models the product was built on. The significance is distribution, not benchmarks: an open Chinese model just got handed the largest professional developer audience in the world as a selectable default-in-waiting.
z.ai ships ZCode, a Claude Code-style harness built on GLM 5.2
Zhipu's z.ai released its own agent harness with GLM 5.2 underneath, a direct answer to Claude Code that pairs a cheap open model with an owned tool. It is the other half of the distribution play: if the incumbents will list your model, ride their shelf; if they will not, ship your own harness and pull developers to you.
Even OpenAI shipped a plugin to run Codex from inside Claude Code
OpenAI published codex-plugin-cc, letting developers drive its Codex agent from within Anthropic's Claude Code to review code or delegate tasks. When the vendor with the most to lose from tool neutrality ships a plugin into a rival's tool, the direction is settled: the harness and the model are separating, and everyone is planning for a world where developers mix and match.
CursorBench 3.1 raises the bar for agentic coding evaluation
Cursor refreshed its public evaluation suite, the kind of infrastructure that decides which model a tool defaults to. As tools go model-neutral, these benchmarks become the referee: the model that wins CursorBench or its peers is the one a platform team can justify switching the default to. The evals are where the commodity shelf gets ranked.
Senior SWE-Bench grades agents as senior engineers, not interns
Senior SWE-Bench, from Snorkel, tests whether agents can operate at a senior engineer's level rather than closing toy tickets, a harder and more honest bar than the SWE-Bench scores labs like to quote. The gap between models narrows fastest on easy tasks and holds on the hard ones. If an open model clears a senior-level eval, the last argument for paying flagship prices on routine work gets thinner.
Yesterday: Sonnet 5 cut the price of a capable agent
Anthropic put flagship-grade agentic coding at mid-tier pricing with Sonnet 5. Set it beside Kimi landing in Copilot and the squeeze is two-sided: the closed labs are cutting their own prices while open models walk in through the incumbents' tools. Both pressures push the same way, toward cheaper tokens and swappable models.
Monday: the benchmark that started the run
Semgrep's cyber eval put Zhipu's open GLM 5.2 level with Claude at a fraction of the cost. That was the capability proof. This week is the distribution proof: the same class of model is now inside Copilot and shipping its own harness. Capability without distribution is a demo; this week it stopped being a demo.
The open-China models cleared their last real hurdle this week, and it was not a benchmark. Kimi K2.7 Code went generally available inside GitHub Copilot, handed the largest developer audience on earth in a dropdown next to OpenAI. z.ai shipped ZCode, its own Claude Code answer running GLM 5.2. OpenAI itself shipped a plugin into Claude Code. The coding tool is separating from the model it runs, and a model-neutral tool is a commodity shelf where a model that costs a tenth as much competes on even footing with the flagships. Concede that most teams will stay on the default for now; defaults move once a cheaper option that clears the bar is one click away and the bill is visible. For builders the play is to stop marrying a model: invest in the harness, keep two models wired in, and route by cost and task. Run an open model against your real workload this week. The gap may be smaller than the price difference.
Within six months, a Chinese open model becomes the model most sessions actually run in at least one Western AI developer tool with a large user base, whether by being set as the default or by winning usage outright, rather than sitting as one more option in the menu.
GitHub Copilot already ships Kimi K2.7 Code generally available, and z.ai ships its own GLM harness. The coding tool is decoupling from the model, and a model-neutral platform has every incentive to route to a capable option that costs a fraction as much. Once an open model clears the everyday-coding bar and sits one dropdown away, the economics pull the default toward it, and a tool's default and usage mix are things platforms disclose.
If, by January 2, 2027, no Western developer tool with a large user base has made a Chinese open model its default or reported it as the most-used model on its platform, the call is wrong.
Making Copilot a model marketplace hedges Microsoft's OpenAI dependence: whichever model wins, the distribution layer is Microsoft's. The cost is that it commoditizes the model layer it partly owns through OpenAI, but owning the neutral shelf is the stronger position.
The durable value in developer AI is shifting from the model to the tool that distributes it. Copilot listing an open Chinese model shows Microsoft would rather own the channel than defend any single model's margin.
Kimi K2.7 Code inside Copilot is a distribution win most labs would trade a benchmark for. The watch is the same one that dogs every open-weights leader: it captures developer mindshare far better than it captures revenue.
Riding Microsoft's shelf puts Kimi in front of millions of developers at zero acquisition cost. But the value shows up as savings on their bills, not as sales on Moonshot's, unless it converts the reach into a paid tier or an owned tool.
We covered the GLM short to watch nine days ago; ZCode is why. Shipping an owned harness is the move that could finally convert the open-model lead into a product developers pay for, rather than a benchmark others monetize.
An owned tool is the one place an open-weights lab can capture value the weights give away. If ZCode pulls developers the way Claude Code did, Zhipu has a business, not just a leaderboard position.
Shipping a Codex plugin into Claude Code is OpenAI conceding the multi-tool world it would rather not have. Its model is now one option on a neutral shelf inside a competitor's product, which is exactly the position a commoditizing model layer creates.
When even the leader plans for developers mixing tools and models, tool neutrality is the base case. OpenAI's defense is capability and its own surfaces; the plugin is an admission that neither fully holds developers in.