Three months ago, almost every AI coding tool had the same deal: pay a flat monthly fee, use it until you hit a wall. That era is over. On June 1, 2026, GitHub flipped Copilot from counting requests to metering tokens, and the reaction was immediate and ugly. The GitHub community discussion announcing the change has 24 thumbs up and 958 thumbs down. Some developers reported their effective costs jumping 10x to 50x overnight because of which model they had selected.
Meanwhile Claude Code never left the flat-rate model. It just quietly throttles you instead: a five-hour rolling usage window, a weekly cap, and since March, reduced limits during weekday peak hours.
So which billing model actually costs you less? I spent an evening with GitHub’s published rate sheets, Anthropic’s plan limits, and the third-party calculators that have emerged to track both. Before we start, one disclosure: I run my own AI agent infrastructure on a rented VPS and pay per token through APIs, so I have lived with metered AI billing for two years. I have not personally subscribed to Copilot’s new credit system. Every number below comes from primary sources and published rate cards, linked so you can check me.
What Copilot actually changed on June 1
The new unit is the GitHub AI Credit, and the conversion is simple: one credit equals $0.01 of metered model usage. Every chat message, agent task, and code review now consumes tokens, the tokens are priced at each model’s published API rate, and the dollar total converts to credits drawn from your monthly pool.
What each plan includes is where sources differ slightly, because GitHub structures it as a base plus a flex allotment:
- Copilot Pro ($10/month): $10 in base credits (1,000) plus $5 in flex credits, so 1,500 credits total, about $15 of metered usage.
- Copilot Pro+ ($39/month): 3,900 base plus 3,100 flex, 7,000 credits total, about $70 of usage.
- Copilot Max ($100/month): 10,000 base plus 10,000 flex, 20,000 credits, $200 of usage. That is effectively a 2x match on your subscription price.
- Business and Enterprise ($19 and $39 per seat): credits pooled across the organization, currently boosted to 3,000 and 7,000 per user in a promotional window that ends September 1, 2026.
Three details matter more than the headline numbers:
Completions are still free. Inline code completions and Next Edit suggestions remain unlimited on paid plans and never touch credits. If Copilot is mostly autocomplete for you, nothing changed.
The model choice is the bill. Twenty-plus models live in the Copilot catalog, and their rates vary enormously. Claude Sonnet 4.6 runs $3 per million input tokens and $15 per million output. Opus-class models run $5 and $25. Some lightweight models, like GPT-5 mini, are included at zero credits. One community user calculated that Opus went from affordable to roughly 27x more expensive under the new math, enough for about 140 Opus requests a month from a Pro+ allowance.
Overage just runs a tab. The old system cut you off when you hit your request cap. The new one keeps going and bills the excess at $0.01 per credit. Predictable ceiling, gone. In its place, a variable bill that depends on your model discipline.
And starting September 1, code review consumes GitHub Actions minutes on top of AI Credits, a detail buried in the community manager’s announcement that teams will feel in two places on the same invoice.
How Claude Code’s model works instead
Claude Code takes the opposite bet: a flat subscription with a rationing system rather than a meter.
- Claude Pro, $20/month: full Claude Code access, roughly 10 to 45 prompts per five-hour rolling window, with a weekly cap measured in compute hours. Sonnet-class models only.
- Claude Max 5x, $100/month: five times the per-session usage, access to Opus-class models, roughly 50 to 225 prompts per window.
- Claude Max 20x, $200/month: twenty times Pro usage, for full-time agent users.
When you hit the wall, you wait. The window resets on a rolling five-hour cycle, and the weekly cap resets Monday. No overage, no surprise invoice, but also no guarantee you can finish your task today. Anthropic has also reduced five-hour limits during weekday peak hours, 5 AM to 11 AM Pacific, since March 2026, which stings if you are a heavy user in a timezone that overlaps US mornings.
The philosophical difference is real. Copilot prices your usage and lets you work unlimited. Claude Code caps your usage and prices you nothing extra. One behaves like a cloud provider, the other like a gym membership.
The session-by-session math
Plan tables hide the decision, so let me price an actual unit of work. Say a typical agentic session: the model reads your repo context, plans, and writes a change. Tracked breakdowns of Copilot billing put a realistic Sonnet 4.6 agent session at around 50,000 input tokens and 20,000 output tokens.
At Sonnet 4.6 rates, that is $0.15 of input plus $0.30 of output, so $0.45 per session, about 45 credits. Notice that output is only 40 percent of the tokens but 60 percent of the cost, because output tokens cost 5x input on Anthropic models. That ratio drives everything.
- Light agent use, 4 sessions a week: about $7.80 a month. Fits inside Copilot Pro’s 1,500 credits with room for chat on top.
- Daily agent use, 30 sessions a month: about $13.50. Blows through Pro, fits comfortably in Pro+ at $39.
- Heavy Opus work, 30 sessions a month: about $22.50 at Opus rates. Still fine on Pro+, but only because Pro+ overprovisions flex credits.
Now the same user on Claude Code. Thirty agent sessions a month is roughly one or two solid working sessions a day. Claude Pro at $20 would plausibly survive that, if the sessions are not enormous, but one deep afternoon of agent work can eat the five-hour window and put you into cooldown. Max 5x at $100 removes the anxiety entirely.
So the crossover is somewhere around daily agent usage. Below it, Copilot’s meter is cheaper because you simply spend less than the subscription price. Above it, Claude’s flat rate wins because the marginal session costs zero, up to the throttle.
Which profile are you?
I reduced this to four working styles, and each has a clear winner:
- Autocomplete plus occasional questions. Copilot Free or Pro. Completions are unlimited, light chat fits in the free credit allowance. Claude Code is the wrong tool at this usage level.
- Daily chat, weekly agent tasks. Copilot Pro at $10, if you keep the model picker disciplined. GPT-5 mini and other included models cost zero credits. The developers who got 10x-50x bill shocks were almost all running frontier models for work that did not need them.
- Daily agent loops on the best models. This is the contested middle. Copilot Max gives you $200 of usage for $100. Claude Max 5x gives unlimited-with-throttle for the same $100. If your work comes in unpredictable bursts, Claude. If it is steady all month, the math favors Copilot Max.
- Team budgets that cannot surprise anyone. Neither is great, but Copilot Business pooled credits with per-user budget controls give finance a lever to pull. Claude’s flat $20 or $100 per seat is predictable too, as long as nobody complains about hitting cooldowns. Note the September 1 step-down: team credits drop from the promotional 3,000 and 7,000 per user to the standard 1,900 and 3,900. Size your budget for the step-down now, not in September.
The checklist I would run before picking
- Audit one real week of usage. Count agent sessions and their rough token sizes before trusting anyone’s averages, including mine.
- Check which models your workflow actually needs. If a lightweight included model handles 80 percent of your tasks, Copilot’s meter barely registers.
- Decide what failure feels worse: a bigger invoice, or a locked-out afternoon. Copilot fails you with money, Claude Code fails you with time.
- Set a spend cap either way. Copilot exposes user-level budget controls. For API-billed setups, hard-limit the card.
- Re-check in September. The Business and Enterprise step-down lands September 1, and both vendors adjust rates often enough that this article’s numbers have a shelf life.
What I would do
My own setup is the third path nobody markets: no subscription at all, just an API key, a small pre-funded balance, and a personal rule that any task over a few cents of output tokens has to justify itself first. Metered billing disciplined me years before GitHub made it mandatory. But that only works because my usage is spiky and I write the harness myself. For most working developers, Copilot Pro plus model discipline is the cheapest sane default in 2026, and Claude Max is worth it the day you catch yourself waiting on a cooldown instead of shipping.
The deeper shift is worth naming. AI coding tools have moved from subscription economics to cloud economics: metered units, tiered rates, promo windows, step-downs. The developers who thrive in that world are the ones who read the rate card. You are now the person who reads the rate card.
I write about AI tooling, engineering economics, and building with agents every week. Subscribe, it is free, and it helps me keep doing the unglamorous math so you do not have to.
Have you felt the June 1 change in your Copilot bill, or hit a Claude Code cooldown at the worst moment? Which billing model suits the way you actually work? Tell me in the comments.
