The end of the subsidized AI coding era
On June 1, GitHub switched every Copilot plan to usage-based billing. You spend AI Credits on tokens now instead of counting premium requests. The reckoning I’m talking about is the impending collapse of US market share in AI coding. For years, US frontier labs owned this space because their models were bundled into the tools we liked. Now that the tools are metering usage, we can see the price gap. US labs are charging a 10x premium for "frontier" performance that developers are realizing they don't need for 95% of their agentic loops.
GitHub made the bill visible
Copilot used to feel like a subscription with a soft cap. You paid $10 or $39, got a pile of premium requests, and learned to ration the expensive models. GitHub's April 27 update ended that. Chat, agents, code review, and the CLI now draw from monthly AI Credits priced by token use—input, output, and cached tokens at each model's API rate (GitHub, 2026).
Base prices stayed the same: Copilot Pro is $10 and Pro+ is $39. But the shape of the limit changed. Credits make the burn visible. A long agent session that spams tool calls used to feel like "four premium requests." Now it feels like spending an allowance.
GitHub also paused new sign-ups for Pro, Pro+, Student, and Max on April 20 while it reworked billing (GitHub changelog, 2026). That pause is still in place. If you aren't already in, you're waiting.
The student tier is what pushed me to write this. I graduated years ago, but I care about this lane because it used to mean verified students got Pro-class access for free. GitHub moved students to a dedicated plan in March 2026 (GitHub changelog, 2026). In practice, for the work students actually want—chat, agents, and multi-file edits—the Student plan is now just the Free plan with a different name.
GitHub’s own comparison table puts Free and Student side by side on "Copilot interactions": both are limited and use auto model selection (GitHub Docs, 2026). Free caps you at 2,000 code completions per month; Student says "unlimited." But completions were never the expensive part. The loop—reading the repo, planning, editing, and running the terminal—is what eats the margin. That loop eats credits whether you have a .edu email or not.
If you're in school and wanted Copilot as an always-on coding agent, you're on the same ration as everyone on Free. The marketing copy still says students can "access premium features at no cost," but that sounds like text left over from the old pack.
Flat-rate coding AI was priced for autocomplete. Agents rewrote the cost model.
The subsidy era was temporary
This shift isn't a surprise if you look at the margins. A coding agent isn't one completion. It's dozens of model calls and tool results stuffed back into context. The vendor pays for every hop. Users got trained on "unlimited" usage because companies were fighting for habit share.
GitHub said the sign-up pause was for reliability and performance (GitHub, 2026). Other reports were blunter: agent workflows were burning far more tokens than flat plans could fund. Every developer I know runs agents longer than they ran inline suggestions in 2023.
It’s the standard consumer tech playbook: hook with generosity, then meter when power users show up. Copilot is just the first major US tool to make the meter official. I expect the rest of the US stack to follow. The models didn't get worse; the bills just arrived.
Cursor is the best deal—and the canary
Cursor is my daily driver. I pay for the review loop—inline diffs and Composer for multi-file passes. I still think it's the best editor on the market.
I run two Pro accounts at $20 each and work most of the day without hitting a wall. I lean on Auto plus the Composer usage pool. Composer 2.5 is fast enough for refactors and test stubs. I save the heavy frontier models for the scary corners. On a normal Tuesday, I squeeze an embarrassing amount of agent time out of those subscriptions.
I doubt Cursor is breaking even on my usage. The math doesn't look close. Composer draws from a standalone pool on individual plans, while team tiers pay $0.50 per million input tokens and $2.50 per million output (Cursor, 2026). I'm one user with a strategy: default to Composer, batch the big thinking, and commit before long runs. Multiply that by every power user doing the same.
Cursor’s own launch post says Composer 2.5 builds on Moonshot’s open-source Kimi K2.5 checkpoint (Cursor, 2026). The value chain points at a Chinese open-weight lab, then an American editor company, then me at $20 a month. That's where the margin lives.
If GitHub had to pause sign-ups over agent economics, Cursor’s Composer pool is the canary in my wallet. Something will give: the pool will shrink, prices will rise, or defaults will get slower. I'm enjoying the window, but I'm not planning on it lasting forever.
Anthropic wins quality until the invoice wins
I don't use Claude Code much, but I accept the consensus: Anthropic’s models are the ones people trust when the task is hard or subtle. That reputation is a real moat.
Moats don't change the math. When every harness can swap models, "best" becomes a line item you invoke on purpose. For boilerplate edits, I reach for Composer. For small local tasks, I use models that would have been jokes on my GPU two years ago. They aren't Claude, but they're good enough to clear my queue of boring work.
Leaderboard heroics don't pay your token bill. A model can score well on a chart and still lose on cost per merged pull request when you run it eight times through an agent loop. Anthropic can keep winning quality headlines while losing default placement in daily coding. The invoice usually picks the model, not the benchmark.
DeepSeek, Kimi, and the harness pattern
The price cuts I'm watching now aren't coming from US frontier labs.
DeepSeek’s V4-Pro is $0.435 per million input tokens on a cache miss and $0.87 per million output (DeepSeek, 2026). GitHub made a 75 percent promo permanent rather than let list prices snap back. I used to pipe DeepSeek API calls through Trae, and the workflow felt close to what I get from Cursor now: a harness in the IDE, a cheap model behind it, and me reviewing diffs.
Moonshot’s Kimi line keeps showing up in my workflow. Composer 2.5 is a retuned K2.5. Kimi K2.5 and the K2.6 previews are strong for code in community testing. They might not beat Opus on everything, but the gap narrowed faster than US pricing relaxed.
Local models improved on a different curve. They won't carry a huge refactor, but they can rename variables and scaffold tests on hardware you already bought. One upfront GPU bill, no per-token meter.
The pattern that ties this together is the harness. Cursor, Claude Code, and Copilot are all harnesses. US vendors want to own the shell, but they can't own every model if the shell lets you point at DeepSeek, Kimi, or a local endpoint. The product is the agent loop and the policy layer; the model is now a swappable plug-in. Qodo already assumes you keep Copilot or Cursor for writing and adds its own review layer on top. That stack is the natural endpoint when the "best" model costs ten times more than the "good enough" one.
This is the reckoning: US frontier labs are about to lose the coding market. For years, we used OpenAI or Anthropic because they were the only engines bundled into the tools we liked. Now that the tools are metering usage, we’re realizing we’ve been paying a massive premium for a brand name. If a $0.40 model from an international lab clears your ticket as well as a $4.00 model from a US lab, the US lab loses that market share the second the developer sees the invoice. The "all-you-can-eat" era of American frontier models is over.
What happens next
The market is repricing itself in real time. Here is how the shift looks from my desk:
- US labs lose the default slot. Developers will bring their own API keys for cheaper, international models and only call a US frontier model for the hardest 5% of tasks.
- The "Harness" wins, the "Engine" commoditizes. You will pay for your editor (Cursor, VS Code) and treat the model as a swappable utility.
- The middle tier vanishes. "Pro" plans that try to bundle expensive US models will either double in price or quietly swap in cheaper, retuned open-source weights to stay solvent.
- Local routing becomes the standard. For simple renames and test stubs, the latency and cost of a frontier call make no sense.
What I’m doing to prepare:
- Fund a non-US API account. I keep DeepSeek and Kimi credits ready for when my US-based allowances hit the meter.
- Audit every agent loop. If a tool doesn't show me the token burn, I assume I'm being overcharged.
- Master the local stack. I'm routing rote tasks to local models to save my "frontier spend" for merges that actually matter.
- Stay flexible. I avoid annual subscriptions for any tool that locks me into a single model provider.
GitHub made the bill visible on June 1. Copilot Student matching Free on agents is the receipt for an era that couldn't last. The builders who stay productive will be the ones who treat the editor as their permanent home and the subsidized US model bundle as a temporary gift.