I had 300 Premium requests included in the university teacher program (enough for 14 days of vibing), now it's Credits. Noticet after 50 % usage warning after 2 hours of vibing.

It eats tokens like a Dodge Ram gas.

"Good job" MSFT!

I just used up 30% of my monthly quota on a simple renaming task on a small python application. This is price gouging at its finest. I for one will not be upgrading my account to get more credits and instead will be cancelling it. This used to be a great tool that I looked forward to using but now has deteriorated to a baseless money grab. Thanks, Microsoft and GitHub, I will not be recommended your products going forward.

after 6 days of light use

Shocked by this too, did a tiny bit of messing around and I'm at 44%. What I did today would have been <5% historically. This is wild. Claude code I guess here I come. Or maybe an API key? This is so annoying.

API prices became usage based for Claude.

Eu também fiz isso no passado e ocorreu o mesmo, eles nem perguntam, te cancelam na hora. Muito louco isso.

Hi @Fu-Jie,

Thank you for writing in and explaining the situation.

We understand from what you shared that clicking the refund button processed immediately without a confirmation step, and that this was not your intent. We appreciate you flagging this as a UX concern. I've noted this feedback and shared with internal teams for investigation.

Regarding restoring your previous subscription: this falls outside what I'm able to action directly. We cannot resolve account related issues in this public forum, and I'm sharing this information with the appropriate team to review whether your original subscription and expiration date can be reinstated.

I'll follow up once I have more information. Thank you for your patience.

Yes here same 2 days work 7K credits burned.. god bye GitHub.. you are not more payable, and its close to robbery! besides Copilot keeps you busy after restart to fix the VS again.. +its a shame thins so many developers here put things for free with all there know how, which get robed too. Years a go GIT was relay good now its just a massive rib-off...very sad!

I was a Github Copilot beta tester, then one of the first paying for the service. I was an evangelist of Github Copilot. Today I canceled my subscription. I am switching for Antigravity and Claude, and I am also considering testing how good the new Cursor is.

I upgraded to MAX and started to restructure some docs in my project and after that I did a few task for some implementations - I ended up spending all my MAX credits under 24 hours, this is crazy....

I was working heavily almost every day last month on Pro+ and a budget on $100 - no problem...

Same here the costs are exploding in just a couple of days. It is expected that some raises are necessary but these exceptional higher use of the Agents models and costs are way out of balance

The new pricing model is outright robbery!!! I've already canceled my GitHub Copilot plan subscription!!

did the same

The new pricing model is outright robbery!!! I've already canceled my GitHub Copilot plan subscription!!

did the same

It is disappointing to see the response to these concerns shift into a 'blame game' where users are essentially told they have a 'skill issue' for not optimizing tokens, rather than GitHub providing a transparent, predictable estimation model before requests are run.As it stands, the 'integration' advantage doesn't justify the cost or the lack of predictability.

For my personal work, I’m moving to Codex and exploring other providers immediately.
For my team, the economics of a $0.5M internal deployment for 100 seats now far outweigh the headache of usage-based billing, privacy concerns, and the nightmare of renegotiating third-party NDAs.

It is disappointing to see the response to these concerns shift into a 'blame game' where users are essentially told they have a 'skill issue' for not optimizing tokens, rather than GitHub providing a transparent, predictable estimation model before requests are run.As it stands, the 'integration' advantage doesn't justify the cost or the lack of predictability.

For my personal work, I’m moving to Codex and exploring other providers immediately. For my team, the economics of a $0.5M internal deployment for 100 seats now far outweigh the headache of usage-based billing, privacy concerns, and the nightmare of renegotiating third-party NDAs.

I hope you realize that Codex will do the same as well if they get the huge amount of transferred users. The old rates were simply massively underwater financially and MS has drawn a line.

It is disappointing to see the response to these concerns shift into a 'blame game' where users are essentially told they have a 'skill issue' for not optimizing tokens, rather than GitHub providing a transparent, predictable estimation model before requests are run.As it stands, the 'integration' advantage doesn't justify the cost or the lack of predictability.
For my personal work, I’m moving to Codex and exploring other providers immediately. For my team, the economics of a $0.5M internal deployment for 100 seats now far outweigh the headache of usage-based billing, privacy concerns, and the nightmare of renegotiating third-party NDAs.

I hope you realize that Codex will do the same as well if they get the huge amount of transferred users. The old rates were simply massively underwater financially and MS has drawn a line.

That is a "them" issue. They should not have made a product with high value, low cost, tie it into systems that a majority of developers use, push it at every turn, then yank the proverbial rug out saying "Oops! Too expensive."

It is disappointing to see the response to these concerns shift into a 'blame game' where users are essentially told they have a 'skill issue' for not optimizing tokens, rather than GitHub providing a transparent, predictable estimation model before requests are run.As it stands, the 'integration' advantage doesn't justify the cost or the lack of predictability.
For my personal work, I’m moving to Codex and exploring other providers immediately. For my team, the economics of a $0.5M internal deployment for 100 seats now far outweigh the headache of usage-based billing, privacy concerns, and the nightmare of renegotiating third-party NDAs.

I hope you realize that Codex will do the same as well if they get the huge amount of transferred users. The old rates were simply massively underwater financially and MS has drawn a line.

That is a "them" issue. They should not have made a product with high value, low cost, tie it into systems that a majority of developers use, push it at every turn, then yank the proverbial rug out saying "Oops! Too expensive."

I mean, it's immoral but also entirely predictable and a classic business strategy. You were effectively all on a discounted preview. If you look at API costs on any frontier model, it is obvious MS was eating over 90% subsidy here. MS may live on a mountain of money but the board would never approve of this forever.

Guess what? It's called effective management, hah

I upgraded to GitHub Pro ($48/year) but downgraded it back to Free because I cannot subscribe to Copilot at this time.
I am requesting a prorated refund for the unused portion of the GitHub Pro subscription.
I would like to subscribe to Copilot Pro+ at $39 per month.
What should I do?

I upgraded to GitHub Pro ($48/year) but downgraded it back to Free because I cannot subscribe to Copilot at this time. I am requesting a prorated refund for the unused portion of the GitHub Pro subscription. I would like to subscribe to Copilot Pro+ at $39 per month. What should I do?

Switch to claude code

but at the new consumption rate, that should give you about two days worth !!!

@DanielRuf 👋🏻 Thanks for the question. Our team here supports GitHub Copilot specifically, and we don't have visibility into billing plans for Microsoft 365 Copilot or other Microsoft products.

For questions about M365 Copilot, Microsoft's product and support channels would be the best resource.

I am assuming this was posted in jest, but answering with added context for you or any other users where the info may be helpful.

I regret not taking the refund offer you made

I hear you.. Sound like a trap indeed. So sad and uneven

The new pricing model is outright robbery!!! I've already canceled my GitHub Copilot plan subscription!!

I think it is more of a pendulum; yes, intially a lot of people will leave, but I have a feeling they will come back... eventually.

I think it is more of a pendulum; yes, intially a lot of people will leave, but I have a feeling they will come back... eventually.

most likely they won't for a fairly simple reason... you have set your whole workflow on copilot, adjusted every single part to your needs and you had to, forcefully, to change to another app, another workflow, another settings and so on.... once you got used to that new one, you won't be coming back even if the older one is better just because you are tired to going back and forth... you most likely won't be doing any of this. you will stick to the new one and start adjusting yourself

I'm currently using SONET 4.6 and it's charging me for GPT 5.4 mini, are you kidding me?

@jakteixeiraf 👋🏻 Hi there,

Thank you for reporting this — I want to make sure we understand what's happening.

To clarify: you selected GPT 5.4 Mini in VS Code Chat, but your usage report shows charges attributed to Gemini 3.5 Flash. Is that correct?

A couple of things to be aware of:

• Some Copilot features (e.g., code completions, agent sub-tasks, or background operations) may use different models than the one selected in your Chat panel. These are separate from your explicit Chat model choice.
• Your detailed usage CSV will show exactly which model was used for each request.

Could you please:

1. Download your usage CSV from Settings → Billing and licensing → Metered usage → Copilot
2. Confirm whether the Gemini 3.5 Flash charges correspond to Chat requests (where you selected GPT 5.4 Mini) or to other features like completions or background operations

If the CSV confirms you're being charged for a model you didn't select for those specific interactions, please share that detail and we'll investigate further.

the admin itself is a f*****g bot :)

Yes one small UI bug fix 1,50$ lmao

Imagine , 2 calls , 550 credits so if i'm in Pro plan = 3 commands -_-

The new pricing model is outright robbery!!! I've already canceled my GitHub Copilot plan subscription!!

Check out the input/output costs for some of the agents, it's insane. Claude Sonnet is 300 input, 1500 output, 30 for cache.

👋🏻 @SuperManito Hi there, Thank you for flagging this.

Regarding the usage limit change:

The billing model has shifted from counting flat requests to measuring actual token consumption (expressed as AI credits). These are different units, so a direct numerical comparison between the old request count and the new credit balance isn't apples-to-apples.

Under the previous model, every request counted as "1" regardless of complexity. Under usage-based billing, a simple one-line completion costs very little, while a large multi-file generation costs more, reflecting actual compute used.

Importantly:

• Code completions (autocomplete) remain unlimited and are not charged per token
• Your per-seat subscription price has not increased
• You can monitor exactly where credits are going in Settings → Billing and licensing → Metered usage → Copilot

Regarding model quality:

Model behavior feedback is best surfaced through the 👍/👎 buttons on Copilot responses in VS Code - this goes directly to the team responsible for model quality. If you're seeing a consistent degradation, that feedback is valuable.

For managing your usage:

• https://docs.github.com/en/copilot/tutorials/budgets/getting-started-with-budget-controls

You should have tested this on people instead of using AI. Most of us are willing to pay $10/month but not $100 or more per month. The mystery billing is problematic because we can use up a quarter of our budget on the 3rd and end up with nothing left by the 10th. Then we are stuck trying to decipher the zillion lines of AI generated code in order to debug problems created with the assistance of AI.

I hope you will consider how this impacts your customers.

You should have tested this on people instead of using AI. Most of us are willing to pay $10/month but not $100 or more per month. The mystery billing is problematic because we can use up a quarter of our budget on the 3rd and end up with nothing left by the 10th. Then we are stuck trying to decipher the zillion lines of AI generated code in order to debug problems created with the assistance of AI.

I hope you will consider how this impacts your customers.

Brother you can't even run hundreds of thousands of tokens on a middling, non-Agentic model for $10/month.

As someone who has used independent providers and hosted my own locally, frontier models are staggeringly expensive. Power users were eating up thousands and thousands of dollars in model use.

You should have tested this on people instead of using AI. Most of us are willing to pay $10/month but not $100 or more per month. The mystery billing is problematic because we can use up a quarter of our budget on the 3rd and end up with nothing left by the 10th. Then we are stuck trying to decipher the zillion lines of AI generated code in order to debug problems created with the assistance of AI.

I hope you will consider how this impacts your customers.

Brother you can't even run hundreds of thousands of tokens on a middling, non-Agentic model for $10/month.

As someone who has used independent providers and hosted my own locally, frontier models are staggeringly expensive. Power users were eating up thousands and thousands of dollars in model use.

Except they did do that
And they took people's money up front for a year and then backed out on it

You should have tested this on people instead of using AI. Most of us are willing to pay $10/month but not $100 or more per month. The mystery billing is problematic because we can use up a quarter of our budget on the 3rd and end up with nothing left by the 10th. Then we are stuck trying to decipher the zillion lines of AI generated code in order to debug problems created with the assistance of AI.
I hope you will consider how this impacts your customers.

Brother you can't even run hundreds of thousands of tokens on a middling, non-Agentic model for $10/month.
As someone who has used independent providers and hosted my own locally, frontier models are staggeringly expensive. Power users were eating up thousands and thousands of dollars in model use.

Except they did do that And they took people's money up front for a year and then backed out on it

You still got the year you paid for, they're just opting out of offering it going forward. I'm not trying to defend MS' actions here, just the logic behind them and how predictable it was. Have you seen what OAI or Anthropic API key rates are?

@CCarrMcMahon 👋🏻

You're right that the units have changed: the old model counted "requests" regardless of complexity, while usage-based billing measures actual token consumption. For users who frequently use higher-cost models like GPT-5.4 with larger context, the per-interaction cost is higher than a flat request count implied.

A few things that may help manage your credits:

• Model choice matters significantly. GPT-5 mini ($0.25/1M input, $2.00/1M output) is substantially cheaper than higher-tier models. Choosing a model appropriate to the task complexity can stretch your credits considerably.
• Code completions remain unlimited and are not charged per token on all paid plans.
• Caching helps over time. Repeated context (e.g., the same files across a session) gets cached automatically at up to 90% lower cost.
• Shorter, focused prompts with only necessary context reduce input tokens.

You can see exactly what each request costs in your usage CSV: Settings → Billing and licensing → Metered usage → Copilot.

For budget controls and notifications:
https://docs.github.com/en/copilot/tutorials/budgets/getting-started-with-budget-controls

@CCarrMcMahon 👋🏻

You're right that the units have changed: the old model counted "requests" regardless of complexity, while usage-based billing measures actual token consumption. For users who frequently use higher-cost models like GPT-5.4 with larger context, the per-interaction cost is higher than a flat request count implied.

A few things that may help manage your credits:

• Model choice matters significantly. GPT-5 mini ($0.25/1M input, $2.00/1M output) is substantially cheaper than higher-tier models. Choosing a model appropriate to the task complexity can stretch your credits considerably.
• Code completions remain unlimited and are not charged per token on all paid plans.
• Caching helps over time. Repeated context (e.g., the same files across a session) gets cached automatically at up to 90% lower cost.
• Shorter, focused prompts with only necessary context reduce input tokens.

You can see exactly what each request costs in your usage CSV: Settings → Billing and licensing → Metered usage → Copilot.

For budget controls and notifications: https://docs.github.com/en/copilot/tutorials/budgets/getting-started-with-budget-controls

@github Community Admin
I also see a major issue with the output cost, which is significantly higher than the input cost: users have no control over the number of tokens the model generates. While users can keep their input short and focused, every LLM I’ve seen so far tends to be far too verbose, which consumes a large number of tokens.

And what if the output is completely useless? Users still have to pay for it—and on top of that, they then have to spend additional time fixing the mistakes in an often lengthy and expensive response.

Furthermore, the creativity is completely lost, if you constantly have to monitor your token consumption and then realize, that your tokens are gone, before you even reach a viable result.

You really should overthink that, otherwise you will loose a lot of customers...

Join the $10Club :) I am not going to change it at all. I just want to see where it goes. I hope they change their way of billing and continue to support the developers.

GPT-5 mini lol

Thank you for the feedback.

To confirm: included AI credits reset each billing cycle and do not roll over to the following month. This applies to all plans.

Regarding the reasoning behind this decision: I don't have product strategy context to share on that front, but feature requests and feedback like this are across the GitHub Community Discussions are tracked by the product teams, and we report on what we're hearing from users regularly, so continue to share here.

We appreciate you sharing the comparison and your workflow. That kind of feedback is valuable for the team to understand how users are adapting.

Let them know I cancelled my office365 too because of your greed

Open Router can't change and edit files, it is useless for me, as I am not a programmer, I am a systems admin.

Open Router can't change and edit files, it is useless for me, as I am not a programmer, I am a systems admin.

Try opencode (has a web ui and vscode plugin) with openrouter as provider. Add some specialized agents and you're off. You can even mix and match providers as you wish. Basically supports all of them since if not supported directly you can input any openAI standard API as a custom provider.
There's also a plugin to use existing copilot instructions inside opencode

Open Router can't change and edit files, it is useless for me, as I am not a programmer, I am a systems admin.

That may have been true some time ago. I can say with absolute certainty that using OpenRouter within VS Code is as seamless as using Copilot’s own models.

If you ask Google today, the AI response will likely give you some outdated info about requiring an extension for it to work. So I understand the skepticism.

Creating an account is free. Why not give it a try using one of their free models? That’s what I did last month when I was weighing out my options.

Open Router can't change and edit files

what do you mean? get token. connect in VSCode. all AIs that support tools connected throuhg standard VSCode can use 100% of tools there. vision, files, everything. from gemini class models to deepseek or anything, local or cloud, does not matter. as long as they speak OpenAI API.

I use the same model GPT-5.4 before and after, so please don't blame the model selection. And now in 2 days, I'm almost out of credits meaning value of Pro+ license is pretty much 0.

gpt5.4 is previously 1X, now 6X.

this is fraustrating , i notice in 4 prompt of sample question it used 64% of my budget , what is going on? i tried to contact some human support of github more than 10 times, but nothing

What?? Github, but you just said to use short question, one topic per question to keep small context and now you recommended to repeat content. You mean your BIG but cached context. Great. Just continue spent money.

This looks much better Opus 4.8 - but this is not Copilot, with Copilot it would have been 2 min. on the same credits...

We are not the main target customer anymore. If you read the news, you can see that large companies are paying millions for this type of service, they don't need users like us.

You're right Microsoft is trusted by a lot of companies but I don't think they realize when you do customers like this they remember it

Time to ditch Microslop then and go with linux!

Well, my IT lead already looks where to limit Github Usage and try to figure out if it was a necessary usage. Actually you need good models like Opus in Enterprise context, but if time isnt crucial or you havent a customer who spends tons of money for development, I see no reason to use GitHub Copilot anymore. Its too expensive now

How can I request a refund? I need to get my fucking money back.

Yepp.. I went from being able to vibecode for a month. Some months i would hit the limit but the $10 over budget limit was more than enough to finish up until next month..

Now i consumed my whole budget in just two days from the absolut lightest work. I havent vibe coded in months.. I just asked it kubernetes questions and some linux questions...
This has turned into a outright scam.

Same. im at 110% somehow... I got two workdays out of this...
Before i would get 20 workdays and barely hit 40% of my montly limit... WTF is this?
I managed to consume an entire months worth of request in two days?!

I have a $10 over budget limit.. I hit that in 3h maybe with 7-10 prompts.. WHAT?!
At this point, im thinking of investing in my own hardware and host a LLM myself.. jeezzz
Canceling my subscription and going to some competitor meanwhile...

Of course, it doesn't make any sense otherwise.

Microsoft/GitHub don't think it's profitable with plebs and their Pro/Pro+ subs eating huge amounts of tokens with agentic runs. They'd rather have us use autocomplete and maybe a few agent sessions per month (if we're ok with using like GPT 5.1).

Loyal customers are easier/cheaper to kick off the ship when it starts slowing down, a (disgusting) tale as old as time. Good think it never backfired.