GitHub Copilot's Billing Problem Is Bigger Than

There's a particular kind of chaos that emerges when a developer with too much curiosity and a flat-rate subscription decides to find the limits of the deal. Theo—the developer and streamer behind t3.gg—spent a recent livestream doing exactly that: systematically trying to burn through as many GitHub Copilot tokens as possible on a $40/month plan, before GitHub closes that particular loophole in June.

It's the kind of stunt that's funny on its face. But underneath the bit is a genuinely interesting argument about how AI inference gets priced, who absorbs the cost when the math doesn't work, and what the coming shift to token-based billing means for the developers who've been quietly getting a spectacular deal.

The Arithmetic Problem Microsoft Created for Itself

The core of Theo's experiment wasn't really about trolling Microsoft. It was about demonstrating something structurally broken.

"Billing on messages is stupid," he said during the stream. "You cannot sell a number of messages because each message can rotate between costing like a few cents and costing like $50."

That's not hyperbole. Modern AI models—especially the premium ones now embedded in tools like Copilot—can consume wildly different amounts of compute depending on what you ask them to do. A quick autocomplete suggestion costs almost nothing. A multi-step agentic task that keeps a context window open for 40 minutes, runs cryptography puzzles, and accumulates 8 million input tokens? That's a different animal entirely.

Theo demonstrated the latter in real time. A single prompt he kicked off at the start of stream was still running 40 minutes later when he checked back in. That one request, on a plan that includes 1,500 such requests per month, had already accumulated 8 million tokens of input. He had used 0.2% of his monthly allowance.

The implication is uncomfortable: if even a fraction of Copilot's users figured out how to run agentic, long-context workloads at scale, Microsoft would be writing enormous checks to the underlying model providers—OpenAI, Anthropic, Google—on behalf of customers who paid $40 for the privilege. That's not a sustainable unit economics story. It's barely a unit economics story at all.

GitHub's response—shifting to token-based billing starting June 1st—is the industry's admission that flat-rate AI subscriptions were, in retrospect, a remarkably expensive user acquisition strategy.

Who Actually Benefits From This Transition?

The shift to token-based billing is rational from Microsoft's perspective. Whether it's good for developers is a different question, and one that doesn't have a clean answer yet.

Heavy users—the ones running agentic workflows, long context chains, complex multi-step prompts—will almost certainly pay more. That's the point. The current pricing essentially subsidizes power users at the expense of Microsoft's margins, and the new model ends that subsidy.

Casual users who lean on Copilot for basic autocomplete will probably see little change, or might even benefit if tiered pricing lets them pay less than $40 for limited usage.

The middle group—developers who've started integrating AI into serious workflows but haven't yet hit the extreme usage patterns Theo was stress-testing—face the most uncertainty. Their bills could go up meaningfully depending on how GitHub calibrates the token allowances, which Theo noted have already been added to the dashboard even before the official June switch.

There's a precedent worth looking at here. Cloud computing went through a similar maturation. The early days of AWS featured pricing so opaque that companies regularly received bill shock—hundreds of thousands of dollars for infrastructure they didn't realize was running. The industry eventually developed better tooling, better alerts, and better mental models for cost management. AI billing is in roughly the same adolescent phase right now: the pricing structures are shifting, the tooling is nascent (Theo's chat literally built him a token-monitoring script mid-stream because nothing adequate existed), and most developers don't have a clear intuition for what things actually cost.

The Bigger Picture: Subsidization Wars

One moment in the stream that didn't get much attention but deserves some: when asked whether T3 Chat would build its own AI inference subscription layer to compete with products like OpenCode, Theo declined.

"I don't want to compete on the subsidization wars," he said. "I want you to play with the subsidization wars however you want and we just provide a really good interface for it."

This is a more interesting strategic read than it might appear. Right now, several AI tool companies are effectively competing by losing money—offering inference at below-cost rates to grab market share, betting they'll figure out monetization once users are locked in. It's a pattern with a long history in tech (see: ride-sharing, cloud storage, streaming services), and it tends to end in one of three ways: consolidation, price normalization, or a dominant player with enough scale to make the math work that couldn't work for anyone else.

Theo's instinct to stay out of that fight and focus on interface quality is a reasonable hedge. But it also means T3 Chat's value proposition is entirely dependent on the health of the underlying inference ecosystem—which is currently being propped up by venture capital and big tech balance sheets.

The Open Source Angle Nobody's Building Yet

The most genuinely novel idea in the stream was something Theo floated and immediately disclaimed any ability to build: a "Folding@home for AI compute."

For context, Folding@home is a distributed computing project that lets people donate their idle GPU cycles to protein-folding research. Theo's proposal takes that concept and applies it to subscription limits: a bot that, near the end of the month, donates your unused AI allowance to open source maintainers who need inference budget to run automated tasks on their projects.

The mechanics he described—clone the repo locally, run the prompts on your machine, file a PR—are technically feasible. The interesting wrinkle is his stated goal: he wants to build something "optically damaging" for Anthropic to ban, specifically to highlight what he sees as the tension between AI companies' open source rhetoric and their actual policies.

That's a provocation more than a product roadmap. But the underlying idea—that flat-rate AI subscriptions create a new kind of underutilized resource that could be pooled and redistributed—is worth taking seriously on its own merits, separate from whatever corporate gotcha game he's playing.

On YC Applications and What Actually Gets Funded

Somewhat tangentially, Theo offered advice on Y Combinator applications that's worth preserving outside the stream context.

"The two things that they really index on are that you see something others don't see and you have drive," he said. "How far you are with the product matters a hell of a lot less than how deeply you understand the thing you're trying to build for."

This tracks with what Paul Graham has written publicly, and with how YC has described its own process. The implication for founders is that a half-built product with a sharp, specific insight into an underserved problem is more fundable than a polished product built on a generic observation. YC is betting on the person's model of the world, not just their current execution.

His advice about VC more broadly was considerably more cynical: "I think most VCs are stupid and the role of analysts at most VCs is to help them justify the stupid thing they were already going to do." Whether you agree with that or not, it reflects a view shared quietly by a lot of founders who've been through the fundraising process—that the elaborate due diligence theater often post-rationalizes decisions that were made on vibes and pattern matching.

The GitHub Copilot billing transition is small news by industry standards—a subscription product adjusting its pricing model. But it surfaces something real about the current moment in AI: the introductory pricing that made these tools feel like magic is ending, and we're entering the phase where everyone has to reckon with what inference actually costs.

For most developers, the answer is: probably more than you've been paying. The question is whether the tools are productive enough to justify it once the subsidy goes away.

— Marcus Chen-Ramirez