GitHub Copilot switches to token billing: the end of the flat rate and the shock for devs
🔎 On June 1, 2026, GitHub killed Copilot's fixed subscription
On June 1, 2026, Microsoft switched GitHub Copilot to a pay-as-you-go billing system called "AI Credits". The end of the fixed-price monthly plan, the end of unlimited. Every autocomplete, every chat message, every code review is now measured in tokens and deducted from a credit.
The backlash was immediate. On Reddit, X, and internal forums, developers are talking about "deception", "cost trap", and the "end of the golden age". Reports cite bills multiplied by 10, 25, or even 50 for the most active users.
This change is not trivial. It signals an industry shift: AI vendors have realized that the flat rate doesn't hold up economically against the real-world usage of developers. Copilot is the first major tool to cross the rubicon. It probably won't be the last.
The key points
- Everything is billed per token: autocomplete, chat, code review, PR summary — nothing is "included" in the subscription anymore.
- Copilot Pro at $10/month only includes $10 of credits: a moderate user exhausts them in a few days (source: Bodega One).
- Documented increases from $29 to $750/month for devs using Copilot in an agentic way, according to wowhow.cloud.
- Code review consumes GitHub Actions minutes on top of AI Credits, creating double billing (source: abhs.in).
- A $0 spend cap exists but must be configured manually — it is not enabled by default.
- Flat-rate alternatives like Cursor and Windsurf are seeing an influx of Copilot refugees (source: DEV Community).
Recommended tools
| Tool | Main usage | Price (June 2026, check website) | Ideal for |
|---|---|---|---|
| Cursor | Active dev with built-in AI | ~$20/month flat rate | Developers who want a predictable plan |
| Claude Code | Agentic work in CLI | Pay-per-token (Anthropic) | Senior developers, complex tasks |
| Windsurf | Alternative IDE with AI | ~$15/month flat rate | Those who want to vary models |
| Copilot Free | Basic autocomplete | $0 (limited credits) | Casual devs, discovery |
| Cline | Autonomous agent in VS Code | Model cost only | Advanced users, total control |
What exactly changes on June 1, 2026
GitHub is no longer charging for access. It is charging for each interaction.
The old model was simple: $10/month for Copilot Pro, $39/month for Copilot Business, $19/month per user for Copilot Enterprise. You paid, you used it, regardless of the quantity.
The new model introduces AI Credits. Each credit is worth a certain amount in dollars, and each Copilot operation consumes a variable amount depending on the chosen model, the context length, and the type of interaction.
According to Bodega One, a $10/month Copilot Pro subscription now only includes $10 in credits. In other words, you pay $10 to get $10 in value. The subscription no longer subsidizes anything.
Luca Berton details that model selection plays a central role in consumption. Using GPT-5.5 in chat costs significantly more tokens than Claude Sonnet 4.6 in autocomplete. The developer must now constantly evaluate which model to use based on the task.
Classmethod explains that GitHub has introduced a Billing Preview in the settings to estimate costs. But the tool is only available after activating the new billing system.
The shock in numbers: from $29 to $750/month
Initial real-world feedback is brutal.
Wowhow.cloud has compiled user reports showing spectacular cost jumps. A solo developer going from $29/month (former Copilot Business) to over $750/month under the new regime. The cause? Intensive use in agentic mode: Copilot would modify entire files, generating thousands of tokens per iteration.
Peremptory confirms these orders of magnitude with a mathematical analysis. "Agentic" users — those who let Copilot execute multi-file tasks — see their bills multiplied by 10 to 50x. A development pattern that was economically viable under a flat rate becomes ruinous under pay-per-use.
byteiota summarizes the situation with a clear formula: token billing kills the very concept of AI subscriptions. When the cost is directly proportional to usage, the monthly subscription no longer protects against anything. It's an à la carte payment model disguised as a subscription.
Newsbreak highlights the asymmetry of the impact. Large enterprises with negotiated Enterprise contracts are better able to absorb the shock. Small users and freelancers are the hardest hit.
TechCrunch captured the immediate reactions: "What a joke", "disguised price cut", "I can no longer recommend Copilot to my students". The dominant sentiment is one of betrayal: the product was sold as an accessible tool, it is becoming a premium product by the back door.
To understand what this change represents in the broader ecosystem, check out our article on GitHub Copilot passe au token billing : fin de l'abonnement, début de la facture à l'usage.
How AI Credits work in practice
An AI Credit is an internal GitHub currency unit. Its dollar value depends on your plan.
For Copilot Pro ($10/month), you receive $10 in credits. For Copilot Business ($39/month), $39 in credits. The ratio is 1:1, which means the subscription provides no advantage over pure pay-as-you-go. You are simply paying an access fee.
Consumption varies greatly depending on the type of interaction:
- Autocomplete: low cost per request, but high frequency (dozens per hour of coding).
- Chat: average cost, depends on the context length and the chosen model.
- Code review: high cost because it analyzes entire PR diffs with extended context.
- Agent mode: very high cost because it generates iterative multi-file changes.
abhs.in adds a crucial detail: code review via Copilot consumes not only AI Credits, but also GitHub Actions minutes. This is a double billing that escaped many users during the announcement. You pay for the AI tokens, and for the compute minutes for execution.
The Gist by maltbae provides projections by usage tier. A "light" developer (a few autocompletes a day) will likely stay within their credits. An "average" developer (regular chat, a few reviews) exceeds them by 2-3x. An "agentic" developer exceeds them by 10-50x.
For details on the credit mechanics and initial feedback, see GitHub Copilot passe aux AI Credits : fin du flat-rate et colère des développeurs.
The spending cap: a misleading protection
GitHub introduced a configurable spending cap. You can set it to $0 to block any usage beyond your included credits. In theory, this protects against surprises.
In practice, it is insufficient for three reasons.
First, the cap is not enabled by default. If you don't configure it, your card is charged without limit. This is a design choice that encourages overconsumption.
Second, a $0 cap means that Copilot stops working as soon as your credits are exhausted. No degraded mode, no simplified autocomplete. An abrupt cutoff right in the middle of a dev session. For a tool integrated into your workflow, it's brutal.
Third, the cap does not cover the GitHub Actions minutes consumed by code review. You can have a tokens cap at $0 and still receive an Actions bill.
Dataconomy notes that this complexity is precisely what frustrates developers. A simple billing system would have been understandable. A dual-meter system (credits + minutes) with a partial cap is a budget predictability nightmare.
Alternatives benefiting from the movement
The backlash is creating a massive migration to competitors. DEV Community has published a migration guide that maps out the options.
Cursor : the refuge for the active dev
Cursor maintains a flat rate at ~$20/month (June 2026, check on cursor.com). For a developer coding 6-8h/day with AI, it's immediately more predictable than Copilot's token billing. Cursor offers autocomplete, chat, and an agent mode. The underlying model can be GPT-5.4 or Claude Sonnet 4.6 depending on the user's choice.
This is the most natural option for leaving Copilot: same paradigm (IDE with integrated AI), same workflow, radically different pricing. For a full comparison, check out our page on the best AI tools for code.
Claude Code : for serious agentic work
Anthropic's Claude Code runs in the CLI and is purely billed by the token via the Anthropic account. No fake subscription, no internal credits: you pay for what Claude Opus 4.7 or Claude Sonnet 4.6 actually consumes.
Anthropic's pricing is transparent and competitive. For heavy agentic use, Claude Code with Claude Sonnet 4.6 often costs less than Copilot with GPT-5.5, because Anthropic doesn't add an intermediary margin. The downside: no integrated IDE, it must be used in the terminal or via VS Code with an extension.
Windsurf : model variety
Windsurf offers a flat rate similar to Cursor but stands out with access to a wider variety of models: GPT-5.4, Claude Sonnet 4.6, Gemini 3.1 Pro, DeepSeek V4 Pro. This is interesting for teams that want to adapt the model to the task without managing multiple subscriptions.
Copilot Free : the zero plan for light users
GitHub maintains a free tier with limited credits. For a developer doing 2-3 autocompletes a day and occasional chat, it might be enough. But as soon as the usage becomes regular, the cap is reached quickly.
The asymmetric impact: who loses what
This change doesn't affect everyone equally. Newsbreak identifies three categories of impact.
Small users win (in theory). If you did 5 autocompletes a day, you were paying $10/month for almost nothing. Under the new system, your credits are enough and you might even switch to the free tier. Except that the value of Copilot for such light usage is marginal.
Average users lose. This is the largest category. Developers who use the chat several times a day, do a few reviews a week, occasionally use the agent. Their $10 or $39 in credits are insufficient. They now pay $20-60/month instead of $10-39.
Agentic users are crushed. Those who have integrated Copilot as a true peer programmer — asking it to refactor entire modules, generate tests across multiple files, analyze 500-line PRs. Their costs explode by 10x to 50x. This is the category fueling the reports of $750/month bills.
The paradox is cruel: the more deeply you integrated Copilot into your workflow, the more you are penalized. The early adopters, those who pushed the tool to its limits, are the first to be punished.
Double billing: tokens + Actions minutes
One technical aspect deserves particular attention. Code review via Copilot doesn't just consume AI Credits. It also consumes GitHub Actions minutes.
abhs.in was one of the first to document this mechanism. When Copilot analyzes a PR, it launches a GitHub Actions workflow in the background. This workflow consumes minutes from your Actions quota — a quota that many teams thought was reserved for their own CI/CD.
The result: a team with 2000 Actions minutes/month can see this quota reduced by Copilot reviews. Fewer minutes for your tests, your deployments, your custom workflows.
This is a problematic design. GitHub charges for the service (AI Credits) and the execution infrastructure (Actions minutes). Imagine paying Uber for the ride and also paying for the gas. Transparency is nowhere to be found: this double consumption is only mentioned in the fine print of the documentation.
What this change means for the dev AI industry
Copilot's shift to token billing is not an isolated event. It's a signal.
Current models are expensive to serve. GPT-5.5, Claude Opus 4.7, Gemini 3 Pro Deep Think — these cutting-edge agentic models consume tens of thousands of tokens per complex interaction. The illusion of the flat rate only held up because average usage was still low. The more sophisticated devs became, the more the flat-rate business model collapsed.
What is happening with Copilot will happen again. Other dev AI tool vendors will have to choose: maintain a flat rate and cap usage (frustrating), or switch to usage-based billing (unpopular). The liste des meilleurs outils IA will likely see its pricing models shift in the coming months.
It is also striking to note that open-source models like Kimi K2.6 or GLM-5 are becoming more attractive in this context. A self-hosted model has no marginal cost per token beyond compute. For a team that controls its infra, it is an increasingly serious alternative. The topic is actually central to our analysis of Subquadratic sort du stealth avec SubQ : 12 millions de tokens de contexte, fin de l'attention quadratique ?, which shows how architectural innovations reduce the cost of processing long contexts.
How to estimate your costs before switching
If you are still on the old system or considering migrating, here is a concrete method.
Classmethod recommends using the Billing Preview in the GitHub settings. This tool estimates your monthly credit consumption based on the last 30 days of activity. It is the most reliable data you have.
The Gist de maltbae offers a simple rule of thumb:
- Less than 50 interactions/day: you will probably stay within your Pro credits.
- 50-200 interactions/day: expect a 2-3x overage.
- More than 200 interactions/day or agentic usage: 10-50x overage, consider an alternative.
One interaction = one accepted autocomplete, one chat message, or one PR review. Count yours during a typical work hour, multiply by your monthly hours, and you will have a ballpark figure.
❌ Common mistakes
Mistake 1: Thinking the $0 cap is enabled by default
The spending cap must be configured manually in the billing settings. Without this step, your card is charged without a limit as soon as your credits are exhausted. Go to Settings > Billing > Copilot > Spending Cap and set it immediately.
Mistake 2: Ignoring the consumption of GitHub Actions minutes
Copilot code review eats up your Actions minutes. If you have a limited quota, your own CI/CD workflows will suffer. Monitor the Actions tab to see the share consumed by Copilot.
Mistake 3: Choosing GPT-5.5 for everything
GPT-5.5 is the most expensive model in the Copilot stack. Using it for simple autocompletes or trivial chat questions is pure waste. Claude Sonnet 4.6 or GPT-5.4 are sufficient for 80% of daily tasks and cost significantly less in tokens.
Mistake 4: Not testing an alternative before dropping Copilot
Migrating from an AI tool integrated into the IDE is not instantaneous. Your prompts, your habits, your shortcuts — everything must be relearned. Test Cursor or Claude Code in parallel for a week before cutting off Copilot.
Mistake 5: Comparing the price without comparing the model
Saying that "Cursor at $20 is cheaper than Copilot" is incomplete if you are comparing the Copilot plan with GPT-5.5 against Cursor with Claude Sonnet 4.6. Compare using the same model, otherwise the comparison makes no sense.
❓ Frequently Asked Questions
Does token billing also apply to Copilot Free?
Yes, but with a very low credit cap. Once exhausted, Copilot Free stops working until the monthly renewal. No overage is possible, but there is also no real utility beyond minimal usage.
Can I choose which model to use in Copilot?
Yes, GitHub allows you to select the model by interaction type. For example, you can use Claude Sonnet 4.6 for autocomplete and GPT-5.5 for chat. This is the main lever for controlling your costs.
Does code review consume both credits AND Actions minutes?
Yes. This is the double billing documented by abhs.in. Each review triggers an Actions workflow that consumes minutes from your quota, on top of the AI Credits for the model.
Will Cursor also switch to token billing?
No announcement to that effect (June 2026). But the same economic pressure applies. If agentic usage explodes on Cursor, their flat-rate plan will become unsustainable. Follow our page of the best AI tools for code for updates.
Is there a viable economic model for dev AI?
A flat-rate plan with a generous cap (Cursor/Windsurf model) seems the most viable in the short term. In the medium term, falling compute costs and architectural optimizations (like subquadratic attention) could make token billing acceptable. But in June 2026, the flat rate remains the most predictable choice for an individual developer.
✅ Conclusion
GitHub Copilot's switch to token billing on June 1, 2026, marks the end of an era: the one where dev AI was a flat-rate subscription like any other tool. The numbers speak for themselves — increases from $29 to $750/month documented by wowhow.cloud — and developers have gotten the message. If you're still on Copilot, activate your spending cap immediately and test Cursor ou Claude Code in parallel. The time to decide is now.