GitHub Copilot switches to AI Credits: end of flat-rate and developer anger
🔎 On June 1, 2026, 4.7 million developers discovered their new bill
On April 28, 2026, GitHub announced in an official blog post the end of the Premium Request Units (PRUs) model. Since June 1, every Copilot interaction is billed in tokens via a system called AI Credits. The change seemed technical. It is actually financial, massive, and deeply unpopular.
Within 72 hours, GitHub Community forums, Reddit, and X exploded with testimonials from developers discovering bill projections multiplied by 10, 20, sometimes 50. A case documented by Working Ref reports a projection going from $39.07 to $902.72 for identical usage.
This shift is not a simple pricing adjustment. It is a business model change that redefines the relationship between GitHub and its most engaged users — precisely those who adopt agentic workflows.
The key points
- Premium Request Units (PRUs) have been removed since June 1, 2026, replaced by AI Credits billed per token.
- Each model has its own cost per token: Claude Opus 4.7 costs 27 times more than GPT-5.4 in terms of credit consumption.
- Agentic users are the hardest hit: increases of 10x to 50x are documented, with an extreme case at $902 for a former $39 plan.
- 4.7 million paying subscribers are impacted, but agentic workloads concentrate the bulk of the increase.
- The market is reacting: Cursor, Claude Code, and OpenAI Codex are becoming credible alternatives, as confirmed by the Gartner MQ 2026.
- The risk of lock-in increases: purchased credits are non-refundable and non-transferable.
Recommended tools
| Tool | Main usage | Price (June 2026, check on site) | Ideal for |
|---|---|---|---|
| Cursor | AI-integrated IDE | 20$/month | Developers looking for a direct alternative to Copilot |
| Claude Code | Terminal coding agent | Direct API | Heavy agentic workflows without intermediary markup |
| OpenAI Codex | Cloud coding agent | API billing | Enterprise teams with existing infrastructure |
| Cline | Open-source VS Code agent | Free (bring your own API key) | Independent developers wanting to control their costs |
What exactly changes on June 1, 2026
The principle is simple in appearance, complex in practice. Before June 2026, Copilot operated with PRUs: each request consumed one unit, regardless of the chosen model or the context length. The system was predictable. A $39/month subscription entitled you to a monthly quota of requests, with a moderate additional cost beyond that.
Now, everything is measured in tokens. A token represents approximately 0.75 words in English (slightly less in French). Every Copilot interaction — inline completion, chat, agentic call in VS Code or GitHub Actions — consumes tokens on input (your prompt + context) and on output (the generated response).
The documentation officielle GitHub publishes a pricing grid by model. Each model has a "credit multiplier" that determines how many AI Credits a token from this model consumes.
The multiplier grid, the real problem
According to data compiled by SeptimLabs and CodePick, the differences between models are staggering:
| Model | Credit Multiplier | Relative cost per token |
|---|---|---|
| GPT-5.4 | 1x | Reference |
| GPT-5.5 | 3x | 3x more expensive |
| Claude Sonnet 4.6 | 8x | 8x more expensive |
| Claude Opus 4.7 (Adaptive) | 27x | 27x more expensive |
| Gemini 3 Pro Deep Think | 15x | 15x more expensive |
These multipliers concretely mean that a developer using Claude Opus 4.7 in agentic mode — exactly the type of workflow that GitHub had been actively pushing since late 2025 — sees their costs multiplied by 27 compared to basic usage on GPT-5.4.
The problem is not the price of the model itself. It is that GitHub did not communicate these multipliers transparently before the switch. Developers were using Claude Opus thinking it counted as just another "request".
Who loses what: the before/after comparison
To understand the real impact, we need to distinguish between three usage profiles. The full details are available on the SeptimLabs calculator, but here are the documented monthly projections.
Profile 1: occasional developer (inline completions + light chat)
This profile mainly uses code completions in the editor and a few chat queries per day on GPT-5.4. Before: $39/month fixed, never any overage. After: around $32 to $45/month depending on usage. Low impact, sometimes even a slight decrease.
Profile 2: intensive developer (frequent chat + GPT-5.5)
Regular use of chat with GPT-5.5 for code review, refactoring, and writing tests. Before: $39/month with a few occasional overages reaching $50-$60. After: $80 to $150/month depending on the volume of injected context. Moderate to high impact.
Profile 3: agentic user (Copilot Agent with Claude Opus 4.7)
This is the profile that explodes. A developer who lets Copilot Agent execute multi-file tasks with Claude Opus 4.7, several times a day. The context accumulated by the agent (files read, conversation history, intermediate results) inflates input consumption.
The case documented by Working Ref is revealing: $902.72 projected for a usage that cost $39.07 in PRUs. That's a 23x multiplier. According to How2Shout, some developers are warning of trajectories from $29 to $750/month.
Why are agents so resource-hungry?
A Copilot agent doesn't just make a single call. It reads files, analyzes context, plans steps, executes commands, and iterates. Each step sends input tokens (often 50,000 to 200,000 context tokens) and receives output tokens. With Claude Opus 4.7 and its 27x multiplier, a single complex agentic session can consume the equivalent of hundreds of "classic" queries.
This is precisely the paradox: GitHub massively pushed agentic workflows as the future of development, then priced these same workflows in a retroactive and punitive manner.
Decoding the Microsoft/GitHub Monetization Strategy
Shifting from a per-request system to a token-based system is no technical accident. It is a strategic decision whose lock-in mechanisms SmartScope analyzes.
From buffet to à la carte: capturing the value of advanced usage
The PRU model was a fixed-price buffet. Light users subsidized heavy users. This model works when usage is homogeneous. It becomes untenable when 10% of users (the agentic ones) consume 80% of the resources.
The shift to tokens solves this problem for GitHub: each user pays for exactly what they consume, model by model. But above all, the token system makes it possible to capture the value of premium models (Claude Opus, Gemini Deep Think) whose infrastructure cost is significantly higher for GitHub.
Opacity as a revenue lever
The PRU system was transparent: one request = one unit. The AI Credits system is opaque by design. The developer does not know exactly how many tokens their next agentic request will consume, as this depends on the accumulated context, the length of the open files, and the selected model.
This opacity is not a bug. It makes budget forecasting almost impossible for teams, which naturally pushes them to buy extra credits "just in case" — credits which are, according to GitHub's terms, non-refundable.
The parallel with cloud computing
The strategy is reminiscent of the AWS/Azure model: an attractive entry price, usage-based billing that seems reasonable at first, and then increasing complexity that makes cost estimation impossible without dedicated monitoring tools. GitHub does not offer such a real-time forecasting tool in the IDE. The developer discovers their consumption after the fact.
As noted by l'article de Défense-Sud-Est, it is the end of an era: the one where the individual developer could use an AI tool almost unlimitedly for a predictable fixed price.
The backlash: reactions from the developer community
The anger is not an overreaction from a few stingy users. It is structural and based on three main grievances documented by NotebookCheck and community forums.
Grief 1: the retroactive change
Developers adopted agentic workflows based on a flat-rate pricing. Changing the rules of the game after habits have been formed creates a feeling of being trapped. "I structured my entire workflow around Copilot Agent with Claude Opus. Now I'm being told it costs $900/month. I would never have started if I had known" — a recurring comment on GitHub Community.
Grief 2: the lack of safeguards
No spending cap is enabled by default. A developer could theoretically consume thousands of dollars in a single day without a real-time alert in their IDE. GitHub added a consumption dashboard, but it is not proactive. It is up to the developer to check it.
Grief 3: the feeling of double billing
GitHub charges for AI Credits, but the underlying models (GPT-5.5, Claude Opus 4.7) are provided by OpenAI and Anthropic, who already charge GitHub per token. The developer feels they are paying a significant margin on every token, without proportional added value compared to a direct API call. SeptimLabs calculated that Copilot applies a 2x to 5x markup compared to the direct API cost for non-OpenAI models.
The exodus has already begun
According to nxcode, a growing number of developers are migrating to alternatives. The movement has accelerated since the April 28 announcement, peaking on June 1st when the first real invoices arrived.
The alternatives: who wins in this chaos?
The backlash directly benefits competitors. The Gartner MQ 2026 positions Cursor, OpenAI Codex, and GitHub Copilot as the three leaders in enterprise coding agents, but the momentum is shifting.
Cursor: the immediate big winner
Cursor is probably the most natural alternative for a Copilot user. The VS Code-based IDE offers comparable features (completions, chat, multi-file agents) with a more predictable pricing model. At $20/month, Cursor's Pro plan includes a generous quota that covers most agentic usages without any surprises.
Migration is simple: Cursor is a VS Code fork, so extensions and shortcuts are compatible. Our guide to the best AI tools for code details the differences in how they work.
Claude Code: for developers who want control
Anthropic's Claude Code is a command-line coding agent. It doesn't go through a middleman that adds a markup. You pay for Anthropic tokens directly via your API key. For intensive agentic usage with Claude Opus 4.7 (scored 94.3 in agentic according to benchmarks), it is often 2 to 5 times cheaper than the same usage via Copilot.
The trade-off: no integrated IDE, manual configuration, and you have to manage your API key yourself. This is the option for the senior developer who masters their own infrastructure.
OpenAI Codex: the native cloud agent
OpenAI Codex, scored 80 in agentic, is OpenAI's cloud coding agent. It runs in a dedicated sandbox and is billed directly via the OpenAI API. For teams already in the OpenAI ecosystem, it's a natural transition from Copilot (which already used the underlying GPT models).
Cline and open-source agents: the radical option
Cline, featured in our comparison of the best AI tools for code, is an open-source agent for VS Code that works with any API key. You choose your model (GPT-5.4, Claude Sonnet 4.6, DeepSeek V4 Pro), you pay for what you consume, with no middleman. It is the definition of anti-lock-in.
For teams that want to go further in multi-agent orchestration, platforms like ruflo are gaining traction on GitHub precisely because of this need for control and flexibility.
Impact on teams and businesses
The shift to AI Credits doesn't just affect individual developers. It disrupts the budget management of engineering teams.
The budget manager's nightmare
Before June 2026, the Copilot cost for a team of 20 developers was predictable: 20 × $39 = $780/month, plus a small overage. Today, that same cost can range from $780 to $5,000+ depending on the workflows adopted, the models chosen, and the complexity of the month's tasks.
No budgeting tool allows a CTO to say "we allocate $X per developer per month for Copilot" with a margin of error under 50%. It's unmanageable.
Non-refundable credits: a financial risk
GitHub sells packs of additional credits. These credits have no expiration date, but they are non-refundable. A team that buys $500 worth of credits in anticipation of an intensive sprint, then sees the sprint canceled or postponed, has $500 locked in the Copilot ecosystem. It's a subtle but real financial lock-in.
The enterprise reaction: negotiate or flee
Large accounts with GitHub enterprise contracts have levers of negotiation that individual developers do not. But even there, the signal sent is negative: if the terms can change overnight with a 10x to 50x impact on the bill, contractual trust is eroded.
What this shift reveals about the AI for code market
Beyond the Copilot case, this change illustrates three structural trends in the AI for code market in 2026.
The end of the AI "all-you-can-eat" illusion
The era when AI tools could offer unlimited access for a fixed price is coming to an end everywhere. The reality of infrastructure costs (GPUs, inference scaling, premium models) is catching up with the aggressive pricing models adopted for market capture. Copilot is the first major tool to cross the Rubicon, but it won't be the last.
The battle shifts from the IDE to the agent
Inline completions have become a commodity. The real value lies in agents capable of executing multi-step tasks, reading entire codebases, and modifying multiple files. This is exactly what Copilot has pushed, and it is exactly what it is now pricing at a premium. The paradox is complete.
Transparency will become a competitive advantage
Tools that provide real-time visibility into token consumption, configurable limits per user and per model, and predictable billing will gain market share. Cursor is starting to do this. Copilot is not doing it sufficiently yet.
❌ Common mistakes
Mistake 1: continuing to use Claude Opus 4.7 in Copilot without monitoring
Claude Opus 4.7 has a 27x multiplier. Using it in agentic mode without monitoring its consumption is like letting a taxi meter run without looking. The solution: switch to GPT-5.4 for everyday tasks, reserve Claude Opus for cases where its superiority (94.3 in agentic) is proven, or migrate directly to Claude Code.
Mistake 2: not setting a spending cap
GitHub allows you to configure consumption alerts, but this is not enabled by default. Failing to do so by June 1st is a budgeting mistake. Set a cap in your GitHub organization settings immediately.
Mistake 3: believing the Enterprise plan protects you
The $79/month Enterprise plan includes more credits, but the billing model is identical. An enterprise team aggressively using Claude Opus 4.7 can just as easily see their bill explode. The Enterprise plan does not protect you from usage-based billing; it just pushes the threshold higher.
Mistake 4: ignoring the cost of input context
Many developers focus on the length of the generated response. But in an agentic workflow, the input token cost (files read, history, agent plan) often represents 70 to 80% of the bill. Reducing the injected context (closing unnecessary tabs, targeting relevant files) has a greater impact than reducing the length of queries.
❓ Frequently Asked Questions
Do AI Credits completely replace PRUs?
Yes. Since June 1, 2026, PRUs no longer exist. Everything is converted into AI Credits billed per token based on the model used. There is no transition period or double counting.
Can I still use Copilot for a predictable price?
Partially. If you exclusively use GPT-5.4 for inline completions and light chat, your bill will remain close to $39/month. As soon as you switch to a premium model or an agentic workflow, predictability disappears.
How much does Claude Opus 4.7 really cost via Copilot vs the direct API?
According to SeptimLabs' calculations, the markup is 2x to 5x compared to a direct Anthropic API call. For the same agentic usage, going through Copilot costs significantly more than using Claude Code with your own API key.
Are excess purchased credits refundable?
No. Additional AI credits purchased on GitHub are neither refundable nor transferable. It is a locked-in investment within the Copilot ecosystem, which reinforces the lock-in analyzed by SmartScope.
Is Cursor really a viable alternative?
Yes, for the majority of use cases. Cursor offers completions, chat, and multi-file agents in a VS Code-compatible IDE, for $20/month with a quota that covers most needs. Our guide to the best AI tools for code compares the two in detail.
✅ Conclusion
The shift from GitHub Copilot to AI Credits is a turning point: it marks the end of predictable AI subscriptions for developers and establishes an opaque model where agentic users pay a steep price for the workflows GitHub sold them as the future. Faced with documented price increases of 10x to 50x and a lock-in reinforced by non-refundable credits, migrating to alternatives like Cursor, Claude Code, or Cline is no longer a matter of convenience — it's a matter of budgetary survival. To understand all the stakes of this shift, read our detailed analysis: GitHub Copilot passe au token billing : fin de l'abonnement, début de la facture à l'usage.