GitHub Copilot transitions to token-based billing from Monday, marking the end of the corporate all-you-can-eat model
The migration to AI Credits, effective from this Monday, transforms the predictability of the monthly seat fee into a consumption meter that could multiply engineering teams' bills by more than 25 times.
From this Monday, 1st June, GitHub Copilot moves from a fixed monthly subscription to a consumption-based billing model. Each plan comes with a quota of GitHub AI Credits, where 1 credit is equivalent to US$ 0.01, and anything exceeding this quota will be billed based on the input, output, and cache tokens of the model selected by the developer. The Pro subscriptions at US$ 10 per month provide US$ 10 in credits, the Pro+ at US$ 39 offers US$ 39, the Business plan at US$ 19 per user also includes a pool of US$ 19, and the Enterprise plan maintains US$ 39 per user under the same corporate pool logic.
The official communication came from Mario Rodriguez, Chief Product Officer at GitHub. "Today, a quick question in the chat and an autonomous coding session lasting hours can cost the same for the user," Rodriguez stated on the company's blog, justifying that the premium requests model "is no longer sustainable." The interpretation from CEO Thomas Dohmke is equally straightforward: the flat rate, he claims, no longer covers the inference costs that Copilot embeds in each agentic session.
Who foots the bill for the agentic revolution
The change does not impact developers using line autocomplete; code completions and Next Edit Suggestions remain unlimited across all paid plans. Those incurring costs are those using Copilot as an agent: code review, chat agents, and autonomous coding sessions. Discussion #192948 on GitHub's official forum surpassed 400 comments and accumulated around 900 downvotes since the announcement. Developers report projections of costs jumping from US$ 29 to somewhere between US$ 750 and US$ 3,000 per month for intensive agent profiles, according to a report by TechCrunch.
The pricing for different models varies widely. The Claude Opus 4.7 costs US$ 25 per million tokens, the Claude Sonnet 4.6 is priced at US$ 3 per million for input, the Gemini 3.1 Pro is US$ 12, and the GPT-5.4 nano is US$ 0.20. The effective cost of a feature in the pipeline depends on which model the team has standardised and how much context each agent carries per request. For a bank with 5,000 engineers using Claude Opus in agentic mode, the arithmetic shifts from predictable budgeting to a variation curve that seems more like a cloud than SaaS.
What will happen inside consultancies and banks
Accenture, which according to Satya Nadella brought the Microsoft 365 Copilot to 743,000 employees in April, serves as a bellwether for the industry. When a delivery team in Bangalore or Buenos Aires runs review agents over client legacy repositories, the token ends up in a finance ops spreadsheet rather than in a closed seat budget. Capgemini, with prospective orders worth US$ 12 billion around agentic AI, as noted by Aiman Ezzat at the Capital Markets Day on 27th May, is precisely selling this kind of intensive usage.
In Brazil, the immediate effect is felt in IT consulting delivery centres and within the internal platforms of banks like Itaú, which in May formalised the Iara platform to centralise models and governance, and Bradesco, which has been operating AI agents in software engineering since 2022. The unit cost now depends less on the number of authorised seats and more on the discipline of prompt engineering and retrieval. In India, where TCS and Infosys have thousands of engineers serving clients worldwide, this topic dominated conversations for the quarter: Copilot transitions from a fixed line to an input margin.
The machinery that the change exposes
GitHub is not alone. Between April and May, Microsoft, Workday, SAP, and ServiceNow had already pushed the pricing of agents towards a consumption model. As Microsoft's own finance team indicated in their quarterly results, the booking growth of the AI portfolio now relies on a combination of seats and consumption because the cost of serving scales with usage intensity, not with the number of logins.
Existing Business and Enterprise customers will receive an inflated pool for three months, from June to August: US$ 30 in credits for Business and US$ 70 for Enterprise per user. In September, these figures revert to US$ 19 and US$ 39, respectively. The calibration period is the window during which each CIO discovers whether their developers are using Copilot as an expensive agent or as an expensive autocomplete. For those who underestimate the changes, the Q4 budget arrives as a surprise, and the CFO finds AI within the utilities line.