GitHub Copilot is changing its pricing model and moving to usage-based billing from June 1, 2026. Previously, pricing was based on “premium requests”, basically the number of prompts you sent while using premium models. On the Pro plan, that was 300 per month. From June 1, this changes. Instead, every plan will include a monthly amount of “AI credits” (1 credit = $0.01). Usage will then consume credits based on the number of input, output, and cached tokens used by the model. For Business plans (Copilot Business and Enterprise), credits are pooled. Example: an enterprise with 100 Copilot Business users gets a shared pool of ~190,000 AI credits instead of 100 individual limits. Heavy users can consume more, lighter users balance it out. My personal take: Copilot will likely become noticeably more expensive with this model. And I think this is another sign that the era of flat subscription-based coding agents is ending. If you pay per use anyway, and unused credits do not roll over from month to month, then token efficiency suddenly matters much more. Especially for people using coding agents heavily or integrating them deeply into their workflow. Good news: there are already approaches to reduce token usage without losing quality. Example: https://lnkd.in/dNh2-kb2 Also: review your AGENTS.md and keep it short and precise. Disable MCP servers you don’t need now. Official sources: Press release: https://lnkd.in/deN6nPQe Pricing details: https://lnkd.in/dQ7cd97V For OpenAI models, GitHub’s listed model rates appear aligned with token-based API-style pricing: https://lnkd.in/dMbsM5d3
Soon, IT departments will setup GPU farms to run LLM endpoints for software engineers.
Does anyone know how this affects Teams/Enterprise plans? I have been subtly nudging my org to move up plans so we could gain benefits of copilot agents / reviews. We aren't a tech company so there is minimal investment in AI infra. This would be a blow to my proposal.
Life as a service. Now we need a subscription to work.
The massive subsidy copilot was providing was too good to last. Comparing with colleague using Claude and API key billing I've continually been blown away as I spend like $5 to their $50
Felix Müller
3d
Interesting that the worst service is now raising the prices. I think that strategically, this is a bad decision, because it will drive even more people to Claude & Codex. We have just cancelled our subscription, because even with the old prices we didn’t get enough value in comparison
Andreas Nigg
4d
It will be interesting to see how this plays out. One thing I currently find interesting: OpenAI is constantly resetting rate limits just for some social media gains. They don't have to do that. But they do. I really think OpenAI has some super efficient models while Anthropic has some enormous giants which are basically not servicable.
Matthew Hook
1d
It'll end up like the old days of high speed Internet where people start building and then get a shock bill at the end of the month. It's at that point they cancel their subscription.
Sounds like… Copilot just got a new NFR: cost optimisation… Soon devs be like… "Bug is still there… but I’m waiting for next month’s credits to fix it..."
I heard there's a new super-duper trend gaining popularity. You can actually write code without AI. Just open a notepad and write. The token savings are insane.
I did a rough set of calculations based on my observed usage for my personal projects. I used the sonnet 4.6 model mostly, and on average I would have a token count of aroun100k to 125k per premium request. Since the context caching only lasts 5 minutes from the start of your request and there's no realistic way to type a follow up that fast, I excluded cache hits from the calculations since they effectively don't get used. Based on my usage alone, my 300 monthly requests will translate to something around 30 to 50 requests of equal value based on how I've used copilot. And with the fact that they're also charging for every model, even the current free ones, I'm sad to say that I'll be moving my entire setup to windsurf. Also a question none of the paid plans from any of the service providers address, is the one of what happens to hallucinated output? Some of these model like opus and gpt 5.5 are being billed at $25 to $150 dollars per million output tokens, so if the model hallucinates a nice big bowl of spaghetti code, then where is the option to flag that output as slop and get a refund for that usage? If we now have to token manage every workflow, then I expect the providers to do the same on their end