From 1 June, GitHub Copilot bills by what you use. Every plan now draws down a monthly allowance of GitHub AI Credits, consumed by token usage — input, output and cached — at published per-model rates. The change, set out on the GitHub blog, replaces the older system of premium request units. The autocomplete most engineers lean on dozens of times a day now runs on a meter.

What actually changes

Headline prices hold. Copilot Pro+ stays at US$39 a month and includes US$39 of AI Credits; Business is US$19 a user with US$19 of credits; Enterprise is US$39 with US$39. Instead of the old premium request units, usage is now measured in credits drawn down per token at each model's API rate. Heavier models cost more per call. A developer who lived comfortably inside a flat plan can now run the meter down faster than expected.

Why developers pushed back

The official announcement thread drew more than 400 comments and close to 900 downvotes, per the GitHub community discussion. The worry is predictability. A flat fee is easy to budget; a token meter is not, especially for engineers who keep an AI assistant running through the day. There is also a quieter change: from 1 June, code-review workflows consume GitHub Actions minutes too, so the cost of automation creeps in from more than one direction.

What it signals

This is the AI-tools business model catching up with the cost of running the models. Flat-rate AI was always a subsidy. As usage grows and inference stays expensive, vendors are passing the cost back to users. GitHub is large enough to set the tone for the market. Expect more coding tools to follow Copilot onto a meter, per Enterprise DNA.

What to do before the meter bites

Check which model your team defaults to, since the gap between a light and a heavy model is where credits disappear. Watch the first month's consumption before assuming the included allowance is enough, and set spending limits where the plan allows. The tools work the way they did before. What changed is how the bill is arrived at.

Three months on, the meter spread and the rates fell

The forecast here was that more coding tools would follow Copilot onto a meter, because flat-rate AI was a subsidy that inference costs would eventually end. Metering did spread over the following quarter. The bills did not behave the way that reasoning implied.

Metering did spread, and it moved up a layer. Stripe agreed in August to buy OpenRouter, a gateway that routes enterprise work across roughly 400 models and meters the token spend across all of them. Billing by consumption is becoming infrastructure rather than a pricing choice each vendor makes on its own.

What did not happen is the bill going up. OpenAI cut its cheapest model by 80 per cent on 1 August, taking Luna from one dollar to twenty cents per million input tokens three weeks after launch, with Terra down 20 per cent and the flagship unchanged. Anthropic had announced a Sonnet 5 rise for 1 September and withdrew it on 10 August, making the US$2 and US$10 rates permanent. Across the large Western labs the movement over the quarter was downward or flat.

DeepSeek is the exception, and it went further than a price rise. V4 now charges by the clock, with peak and off-peak rates, and the increase lands hardest on the cheapest token category. The peak window sits inside ASEAN working hours.

The pushback was about predictability, and cheaper units do not fix that

The 400 comments and 900 downvotes on GitHub's announcement thread were about budgeting rather than price level. Falling rates do not answer that objection.

Cheaper units do not restore a fixed monthly number. A team that could previously budget a flat US$19 per developer now budgets a distribution, and the shape of that distribution depends on which model each engineer's editor defaults to. Cheapest-model price cuts help only if the cheap model is the default, and the gap between the top and the bottom of OpenAI's own ladder widened to 25 to 1 over the same period. That makes the default editor setting more consequential than it was in May.

DeepSeek's time-of-day pricing is the sharper version of the same problem. A flat per-token rate can at least be multiplied out in advance. A rate that changes with the clock means an identical task costs different amounts depending on when an engineer runs it, and for teams working in ASEAN hours the expensive window is the working day.

The advice held, for a reason the piece did not anticipate

Checking which model your team defaults to was the practical suggestion here, and it turned out to matter for a second reason.

Microsoft confirmed at Build 2026 that Copilot would move to Project Polaris, its own mixture-of-experts coding model, replacing GPT-4 Turbo as the default engine across all subscribers by August. A vendor that owns the default model sits on both sides of the meter, setting the rate and supplying the service that consumes the credits. That is a different position from a vendor reselling another company's inference, and you should know which one is billing you.

One thing belongs alongside the original advice. GitHub was degraded for about three hours on 17 August, with archive downloads failing at roughly half, and code-review workflows have been drawing on Actions minutes since 1 June. Automation billed by consumption keeps consuming while the pipeline underneath it is failing, and retries against a degraded dependency are billed like any other run. Set spending limits before you test that combination, not after.