Phil Rentier Digital

Posted on May 12 • Originally published at rentierdigital.xyz

The AI Coding Subsidy Era Just Ended. Cursor and Bolt Are Next.

#ai #programming #aicoding #developertools

For 3 weeks, I left GitHub Community open in a tab while I worked. The same threads came back every morning. "GOOD BYE GITHUB COPILOT". "evaluating GITLAB". "Claude Teams it is." Tens of thousands of words of anger in 6 weeks, and half the posts cited the same scribbled calculation (3900 divided by 27 makes 144 Opus requests per month on a Pro+ plan). Reddit fury doesn't translate mechanically into exodus. What's actually happening is quieter. And bigger.

TLDR: GitHub multiplied the price of Opus by 3.6 in 6 weeks. That's the boring side. The interesting side, the one that changes how you pick an AI coding tool for the next 12 months, is what forced Microsoft to move. The same equation is coming for the others.

GitHub just did what the 4 other major AI coding distributors will have to do in the next 12 months. Not out of malice. Because at some point, the math catches up to everyone. I'm going to name the mechanism, explain why Microsoft cracked first, and give you the order in which the rest will follow.

What a Subsidy Cascade Looks Like

Before any of this makes sense, the mechanism needs a name. A subsidy cascade goes in 3 steps.

Step 1: a distributor absorbs the gap between the price it announces and the real cost of what it delivers. Grabbing market share is worth more than margin right now, so the gap stays hidden.

Step 2: one player cracks. Usually the biggest, or the one most accountable to public quarterly numbers. The cost gets too visible to keep absorbing.

Step 3: the others follow within 6 to 12 months. The math becomes defendable now that the first one broke the taboo.

You've seen this movie before. Uber 2018 (the end of driver bonuses). MoviePass the same year (collapse). Netflix 2022 cracked down on password sharing. Spotify squeezed royalty payouts in 2024. WeWork 2019 when the lease subsidies stopped making sense to anyone. Every time, the same arc. Someone subsidizes to win the market, then someone else runs the numbers in public. The rest of the field has 6 to 12 months to figure out a new plan.

One caveat before I keep going. Anthropic might be subsidizing too. Nobody really knows if the Max plans are profitable for them. Today's transparency isn't a guarantee for tomorrow. But Anthropic is a producer, not a distributor. If they're subsidizing, they finance it differently (data acquisition, training compute amortization), not by eating the inference cost of someone else's customer. The math is structurally different.

That distinction is the whole article.

3.6x in Six Weeks: GitHub's Receipt

Now the receipt. Same model throughout, to avoid the Opus 4.6 vs 4.7 confusion that's poisoning half the takes online.

April 16, 2026: GitHub announces Opus 4.7 GA on Copilot with a promotional 7.5x multiplier valid until April 30. April 30: automatic shift to 15x (the regular Pro+ multiplier). June 1: jump to 27x for anyone still on the annual request-based plan. That's 3.6x on the same model in 6 weeks, sourceable line by line on github.docs and github.blog.

In parallel, on April 24: GitHub paused new sign-ups for Pro, Pro+, and Student. They pulled Opus from Pro. They pulled Opus 4.5 and 4.6 from Pro+. Three moves at once. Price going up, access shrinking, forced migration to usage-based billing.

Mario Rodriguez (CPO at GitHub) put the official line in the blog post: "the current premium request model is no longer sustainable."

Translation in plain English: we can't keep paying for what you're using. Even when "we" is Microsoft.

One thing worth flagging. The 27x might come back down if GitHub backpedals under pressure. The exact number moves. The mechanism stays.

144 Requests a Month: What Pro Plus Actually Buys You

Before June 1, you ran agents without watching the meter much. Pro+ at $39 a month, the "premium request" counter in some opaque unit, the gap between you and the cap big enough that nobody bothered tracking it day-to-day.

After June 1, here is what you actually get on Opus 4.7.

3900 AI credits per month divided by 27x equals 144 Opus requests maximum per month. That's 4.8 per day. A typical Claude Code agentic session (2 hours, real feature, the kind where you let the agent loop and self-correct) burns somewhere between 50 and 200 inferences depending on how much it has to plan, rewrite, and retry. One serious session eats a week of your Opus quota.

Compare with the producer's own pricing page (claude.ai/upgrade). Max $100 gives you 5x the rate limits of Pro. Max $200 gives you 20x the rate limits of Pro. For an agentic long-running workload, the ratio flips without ambiguity.

One caveat. GitHub Copilot is still excellent for classic autocomplete and targeted completions on the standard completion model (still at 1x). The problem is specific to Opus in long-running agentic mode, not to all of Copilot. If your day is mostly inline completions in VS Code, you're not the target of this article. If you launch agents and let them run for 2 hours, you redo the math or you walk.

Walking, it turns out, is the rational option.

Why Microsoft Cracked First

Three reasons, stacked.

Volume. With hundreds of thousands of Copilot Pro+ subscribers, the absorbed cost on Opus in long-running agentic mode reached a number Microsoft couldn't pretend not to see at quarterly review. No other AI coding distributor has that scale. When the absolute amount gets that big, the meeting tone changes.

Forced transparency. The move from premium requests to usage-based billing makes every inference traceable down to the token. As long as you charge in opaque "premium requests", you can hide the gap inside the global margin. Once you go token-based, the unit economics show up on a spreadsheet that someone at Microsoft has to sign.

Internal compute arbitrage. Microsoft compute resources are finite. The direct Azure margin on OpenAI inference is better than the Copilot margin relayed through Anthropic. Internal pressure to reallocate compute is structurally one-way.

None of the other distributors have any of those 3 levers. Cursor doesn't operate at Microsoft scale on subscribers. Bolt has no cloud of its own to arbitrage compute against. And Lovable runs everything through external APIs at retail rates. They have one trajectory.

Follow.

Cursor, Bolt, Lovable, v0: The Order Matters

Nominative predictions now. With reasoning. Pushback welcome.

Cursor. Heavy Claude Opus and GPT-5 user. Raised serious money recently, but inference subsidy eats fundraises faster than people realize. Current pricing is $20 Pro with most features unlimited. That's the textbook subsidy reveal scenario. Best guess: repricing in Q3 2026.

Bolt and Lovable. Both positioned as "build apps from a prompt" for non-devs. Which means per-user inference is even higher, because their users don't know how to optimize a context window. Flat $20 to $30 a month is structurally insolvent. Best guess: repricing Q3 or Q4 2026.

v0 (Vercel). Backed by Vercel's own compute infra, which gives them more runway than the others. Possibly later, Q4 2026 or Q1 2027.

To be clear about what I'm doing here. These are predictions drawn from a pattern, not certainties with a date stamp. Either these tools reprice, or they consolidate toward the producer, or they find a new business model. No scenario maintains current pricing on an 18-month horizon. Pick the one you think is most likely. I picked repricing.

The distributors will tell you I'm wrong. The math will tell you who is.

The Producer Wins by Math, Not Marketing

The flip is structural. While distributors adjust their pricing, the direct producer (Anthropic for Opus, OpenAI for GPT, Google for Gemini) sees its relative position improve mechanically, without lifting a finger.

Anthropic didn't attack GitHub. No press release, no counter-promotion. They didn't even cut Max plan prices to capitalize on it. The math did it for them.

That's a structural win for producers over distributors, not a commercial win for Anthropic over GitHub. There's a difference, and it matters for what you do next.

You've seen this pattern in other markets. Spotify versus the labels. Netflix versus cable. Each time, when the distributor stopped being able to subsidize, the producer took back the customer relationship.

The App Store versus indie developers tells the same story on a longer timeline. Apple subsidized distribution for years (curation, payment rails, refund handling, search), built the empire on the 30% cut, then started squeezing developers when the alternatives had already shrunk. The producers (the devs themselves) couldn't escape because there was nowhere to go. The producers in AI coding can. Anthropic and OpenAI sell direct to you with 2 clicks. There is no app store moat between you and them. The only thing between you and the producer today is a monthly habit you haven't reviewed.

For the end user, this means one thing. In 12 to 18 months, you're going to end up at the direct producer, whether it's called Anthropic, OpenAI, or Google. The choice isn't "which distributor" anymore. It's "which producer".

It's the same kind of structural call I made when I moved my agent stack off MCP and onto CLIs. Different layer, same logic. Pick the architecture that survives the repricing of the layer above.

How to Pick Your AI Plan Without Being Stupid in 2026

3 questions. Answer them honestly.

Q1: How much do you pay per month across all AI coding tools combined?

Under $20: stay on the free tier or the cheapest Pro of whatever distributor is most convenient. You're not the target of the repricing.

$20 to $50: move to the direct producer. Anthropic Max $100 or ChatGPT Plus depending on which model dominates your workflow.

Over $50: Max $200 or direct API access becomes almost mandatory. Also, track that budget as a monthly metric, not a recurring auto-charge you forgot about.

Q2: How many hours per week of long-running agentic work?

Over 5 hours: get out of request-based pricing. The math doesn't forgive at that volume anymore.

Q3: Is your main tool today a distributor or a producer?

If distributor, plan your migration in the next 12 months even if nothing has visibly moved on their side yet. The criterion that survives every repricing wave is simple. Reason in unit cost per inference, not in monthly subscription headline.

One caveat. For enterprises on Azure or Microsoft volume contracts, the arithmetic changes with discount tiers nobody publishes. This article is written for the solopreneur who pays from their own card.

A quick personal note, since someone is going to email about it. In February, I walked away from a $200-a-month OpenClaw setup and rebuilt the same workflow for $15. Today, $200 a month at Anthropic direct becomes defendable where it didn't 3 months ago. The math, itself, has flipped. The framework that flipped it is the one above.

Accept that your favorite tool today isn't necessarily your favorite tool in 12 months.

Microsoft eventually handed over the invoice it had been keeping in its pocket for 2 years. Everyone looked at the number, did their little Reddit outrage, and kept using the same tool with slightly different pricing.

The mechanism, itself, hasn't moved.

Pick your producer before your distributor picks one for you.