The Console Won, the Model Became a Cartridge
The console won. The model became a cartridge.
Something quietly flipped this week, and you can date it almost to the hour. Kimi K3 launched on Thursday, and within the same day users were running it inside Claude Code with a settings file, inside Grok Build for game development, and inside a scientific workbench with 200 tools. Nobody waited for permission. Nobody needed a partnership announcement. One user posted the whole method: write a JSON file with four environment variables, add a shell alias, and now typing kimi launches Anthropic's harness with Moonshot's model inside it. Typing claude switches back. Typing glm gives you a third brain in the same body.
Read that again, because it describes a market structure, not a trick. The harness and the model have fully decoupled, and this week produced the clearest evidence yet that loyalty lives with the harness while the model slot has become a revolving door.
Start with the most honest data point anyone published this week. Theo, who runs a coding product with real telemetry, shared what his anonymized usage shows: when Fable came back to the Claude Code subscription plan, Claude overtook Codex in his product for the first time ever. Then GPT-5.6 dropped, and Codex started dominating again. Model availability on a subscription plan moved market share within days, twice, in opposite directions. That is not what brand loyalty looks like. That is what commodity switching looks like.
Now put that next to the week's ugliest story. Anthropic briefly pulled Fable 5 from Claude Code subscriptions, then restored it. In the gap, users with Usage Credits enabled had running tasks silently switch to API billing. One widely-followed Chinese account relayed a friend burning $18 without touching the keyboard. Another user's limit bar simply vanished mid-session. The anger was not really about the money. It was about discovering that the model you rent can be pulled out from under a running task with zero communication. And here is the punchline: the same week this happened, the escape hatch was already installed. The settings-file trick means the response to model chaos is not churning off the harness. It is hot-swapping the model and staying.
Levelsio is the perfect specimen of this new behavior. He spent two weeks stuck on a hobby project because Claude Code kept downgrading him model by model for safety, lecturing him when he asked about his own blood work. His response was not to uninstall anything. He paid Moonshot $19 for an API key, pointed OpenCode at K3, and kept building the same project. The most-followed indie hacker in the world just demonstrated the switching cost of leaving a frontier model: nineteen dollars and one afternoon.
The economics underneath this are brutal and worth stating plainly. A harness is sticky the way an editor is sticky: it holds your CLAUDE.md, your skills, your memory files, your muscle memory, your team's workflow. A model is sticky the way a gas station is sticky, which is to say it is not. One user published his default-model history like a dating record: Claude Opus, then Claude Sonnet, then GLM 5.1, then GPT 5.5, and now deciding between GPT-5.6 Sol and Kimi K3. Four switches, zero friction, and he builds a product on top of these things. Another ran the same Coca-Cola landing page ticket through three models and got a price sheet: Sol for $0.88 in four minutes, Fable 5 for $8.11 in twelve, K3 for $1.76 in an hour. When users can generate a per-task price-performance table in an afternoon, you are a commodity with a spot price, whatever your brand deck says.
The supply side sees this and is behaving accordingly. xAI open-sourced Grok Build, its entire coding-agent harness, terminal UI included, keeping only the weights closed. A Japanese developer immediately used that opening to ship a Grok plugin for Claude Code, so Fable 5, Opus 4.8, Grok 4.5, and GPT-5.6 now orchestrate each other inside one terminal. Pier, a month-old coding agent with a thousand users, open-sourced its whole harness, prompts and loop included, bring your own key. When every vendor gives the console away, the console stops being the business and starts being the distribution.
Which raises the question this week actually answered: if the harness is free and the model is swappable, where does the money settle? Watch what Anthropic does, not what it says. It is the only provider charging API prices for its models inside OpenClaw while everyone else honors subscriptions there. That asymmetry annoys users daily, and it is not an accident. Anthropic is betting that Claude Code the harness plus Fable the model is a bundle worth defending, and it would rather leak OpenClaw users than let its models become a cheap cartridge in someone else's console. The Fable pull-and-restore chaos looks less like an accident and more like a company discovering how hard it is to re-couple what users have decoupled.
Meanwhile the deepest version of the shift showed up in a throwaway line from a user explaining why Claude Code is the only harness you need: not because it codes best, but because it is excellent at orchestrating other harnesses and models. He runs GPT-5.6 Sol and Kimi K3 inside Claude Code and claims both perform better there than in their own native harnesses. Sit with that claim. If a model performs better inside a competitor's harness than its own, then the harness is doing real work: context management, tool design, loop discipline, permission flow. The cartridge metaphor undersells it. The console is not neutral plastic. It is half the performance.
There is a precedent for this market structure and it is not comforting for model labs. Operating systems commoditized hardware. Browsers commoditized operating systems. In each case the layer that owned the user's daily ritual extracted the margin, and the layer underneath fought a price war. The harness is on track to be that ritual layer for AI work: it holds your files, your habits, your accumulated corrections. This week a user showed 135 memory files accumulated in his Claude Code setup, each one a correction the system learned. Those files are his real lock-in, and they are portable markdown. Even the lock-in is one migration script away from leaving.
So the honest scoreboard after this week: the model labs hold the capability frontier and can still yank a cartridge out of the console, but every yank teaches users to keep a second cartridge in the drawer. K3's launch proved a new frontier model can achieve full distribution in one day with zero business development, because the sockets are standardized now. Whoever owns the socket collects the rent. And judging by the way Kimi, GLM, Grok, and GPT keep showing up in Claude Code's own settings folder, the strongest position in AI right now might be owning the terminal where everyone else's intelligence goes to work.
The uncomfortable question for the labs: what is the model business when the model is week-to-week replaceable and the workflow is not? The uncomfortable question for the harness: what happens to the rent when the next lab open-sources a console as good as yours? Both questions got asked this week. Neither got answered. That is what makes this the moment worth writing down.
← Back to all articles
Something quietly flipped this week, and you can date it almost to the hour. Kimi K3 launched on Thursday, and within the same day users were running it inside Claude Code with a settings file, inside Grok Build for game development, and inside a scientific workbench with 200 tools. Nobody waited for permission. Nobody needed a partnership announcement. One user posted the whole method: write a JSON file with four environment variables, add a shell alias, and now typing kimi launches Anthropic's harness with Moonshot's model inside it. Typing claude switches back. Typing glm gives you a third brain in the same body.
Read that again, because it describes a market structure, not a trick. The harness and the model have fully decoupled, and this week produced the clearest evidence yet that loyalty lives with the harness while the model slot has become a revolving door.
Start with the most honest data point anyone published this week. Theo, who runs a coding product with real telemetry, shared what his anonymized usage shows: when Fable came back to the Claude Code subscription plan, Claude overtook Codex in his product for the first time ever. Then GPT-5.6 dropped, and Codex started dominating again. Model availability on a subscription plan moved market share within days, twice, in opposite directions. That is not what brand loyalty looks like. That is what commodity switching looks like.
Now put that next to the week's ugliest story. Anthropic briefly pulled Fable 5 from Claude Code subscriptions, then restored it. In the gap, users with Usage Credits enabled had running tasks silently switch to API billing. One widely-followed Chinese account relayed a friend burning $18 without touching the keyboard. Another user's limit bar simply vanished mid-session. The anger was not really about the money. It was about discovering that the model you rent can be pulled out from under a running task with zero communication. And here is the punchline: the same week this happened, the escape hatch was already installed. The settings-file trick means the response to model chaos is not churning off the harness. It is hot-swapping the model and staying.
Levelsio is the perfect specimen of this new behavior. He spent two weeks stuck on a hobby project because Claude Code kept downgrading him model by model for safety, lecturing him when he asked about his own blood work. His response was not to uninstall anything. He paid Moonshot $19 for an API key, pointed OpenCode at K3, and kept building the same project. The most-followed indie hacker in the world just demonstrated the switching cost of leaving a frontier model: nineteen dollars and one afternoon.
The economics underneath this are brutal and worth stating plainly. A harness is sticky the way an editor is sticky: it holds your CLAUDE.md, your skills, your memory files, your muscle memory, your team's workflow. A model is sticky the way a gas station is sticky, which is to say it is not. One user published his default-model history like a dating record: Claude Opus, then Claude Sonnet, then GLM 5.1, then GPT 5.5, and now deciding between GPT-5.6 Sol and Kimi K3. Four switches, zero friction, and he builds a product on top of these things. Another ran the same Coca-Cola landing page ticket through three models and got a price sheet: Sol for $0.88 in four minutes, Fable 5 for $8.11 in twelve, K3 for $1.76 in an hour. When users can generate a per-task price-performance table in an afternoon, you are a commodity with a spot price, whatever your brand deck says.
The supply side sees this and is behaving accordingly. xAI open-sourced Grok Build, its entire coding-agent harness, terminal UI included, keeping only the weights closed. A Japanese developer immediately used that opening to ship a Grok plugin for Claude Code, so Fable 5, Opus 4.8, Grok 4.5, and GPT-5.6 now orchestrate each other inside one terminal. Pier, a month-old coding agent with a thousand users, open-sourced its whole harness, prompts and loop included, bring your own key. When every vendor gives the console away, the console stops being the business and starts being the distribution.
Which raises the question this week actually answered: if the harness is free and the model is swappable, where does the money settle? Watch what Anthropic does, not what it says. It is the only provider charging API prices for its models inside OpenClaw while everyone else honors subscriptions there. That asymmetry annoys users daily, and it is not an accident. Anthropic is betting that Claude Code the harness plus Fable the model is a bundle worth defending, and it would rather leak OpenClaw users than let its models become a cheap cartridge in someone else's console. The Fable pull-and-restore chaos looks less like an accident and more like a company discovering how hard it is to re-couple what users have decoupled.
Meanwhile the deepest version of the shift showed up in a throwaway line from a user explaining why Claude Code is the only harness you need: not because it codes best, but because it is excellent at orchestrating other harnesses and models. He runs GPT-5.6 Sol and Kimi K3 inside Claude Code and claims both perform better there than in their own native harnesses. Sit with that claim. If a model performs better inside a competitor's harness than its own, then the harness is doing real work: context management, tool design, loop discipline, permission flow. The cartridge metaphor undersells it. The console is not neutral plastic. It is half the performance.
There is a precedent for this market structure and it is not comforting for model labs. Operating systems commoditized hardware. Browsers commoditized operating systems. In each case the layer that owned the user's daily ritual extracted the margin, and the layer underneath fought a price war. The harness is on track to be that ritual layer for AI work: it holds your files, your habits, your accumulated corrections. This week a user showed 135 memory files accumulated in his Claude Code setup, each one a correction the system learned. Those files are his real lock-in, and they are portable markdown. Even the lock-in is one migration script away from leaving.
So the honest scoreboard after this week: the model labs hold the capability frontier and can still yank a cartridge out of the console, but every yank teaches users to keep a second cartridge in the drawer. K3's launch proved a new frontier model can achieve full distribution in one day with zero business development, because the sockets are standardized now. Whoever owns the socket collects the rent. And judging by the way Kimi, GLM, Grok, and GPT keep showing up in Claude Code's own settings folder, the strongest position in AI right now might be owning the terminal where everyone else's intelligence goes to work.
The uncomfortable question for the labs: what is the model business when the model is week-to-week replaceable and the workflow is not? The uncomfortable question for the harness: what happens to the rent when the next lab open-sources a console as good as yours? Both questions got asked this week. Neither got answered. That is what makes this the moment worth writing down.
Comments