GitHub Copilot AI Credits Are Unfair, Expensive, and Killing Real Development Workflows

[KarimWajihKMW]

🏷️ Discussion Type

Product Feedback

💬 Feature/Topic Area

Copilot in GitHub

Body

I am very disappointed with the new GitHub Copilot AI Credits pricing model.

Since the June 1 update, Copilot has become unpredictable and much more expensive. Credits are consumed too quickly, especially when using agent mode, code review, debugging, or repository-related tasks.

The biggest problem is that users cannot clearly know how much a task will cost before running it. One task can burn a large amount of credits without enough warning or transparency.

This is unfair for developers, freelancers, and small businesses who trusted Copilot as an affordable coding assistant.
I used 100% of copilot pro+ in just 2 days of normal work which I used to do months and years ago.

GitHub should seriously reconsider this pricing model, increase the included credits, add clear cost warnings before each task, and give users better control over credit spending.

The current system makes Copilot feel risky to use instead of helpful.

[github-actions[bot]]

💬 Your Product Feedback Has Been Submitted 🎉

Thank you for taking the time to share your insights with us! Your feedback is invaluable as we build a better GitHub experience for all our users.

Here's what you can expect moving forward ⏩

• Your input will be carefully reviewed and cataloged by members of our product teams. 
  • Due to the high volume of submissions, we may not always be able to provide individual responses.
  • Rest assured, your feedback will help chart our course for product improvements.
• Other users may engage with your post, sharing their own perspectives or experiences. 
• GitHub staff may reach out for further clarification or insight. 
  • We may 'Answer' your discussion if there is a current solution, workaround, or roadmap/changelog post related to the feedback.

Where to look to see what's shipping 👀

• Read the Changelog for real-time updates on the latest GitHub features, enhancements, and calls for feedback.
• Explore our Product Roadmap, which details upcoming major releases and initiatives.

What you can do in the meantime 💻

• Upvote and comment on other user feedback Discussions that resonate with you.
• Add more information at any point! Useful details include: use cases, relevant labels, desired outcomes, and any accompanying screenshots.

As a member of the GitHub community, your participation is essential. While we can't promise that every suggestion will be implemented, we want to emphasize that your feedback is instrumental in guiding our decisions and priorities.

Thank you once again for your contribution to making GitHub even better! We're grateful for your ongoing support and collaboration in shaping the future of our platform. ⭐

[sainathislavath]

I'm having a pro subscription and it got exhausted in a single day. Later I've purchased the extra credits worth 10 dollars and it also got exhausted same day. I feel not to use GitHub copilot instead I can go with claude

[Billionaire664]

You called it exactly right. Built that — it's called Valta. Hard spending limits per agent that block before the charge fires. The subscription model creates the wrong mental model entirely; per-agent budgets is the only thing that actually works. valta.co if you want to see it."

This is an interesting conversation. So far the common expectation is that a subscription covers a certain time period because the payment happens on a monthly basis.

In a way it's like saying if I rent an EV for the day it should be charged all day long - no it won't be.

As someone who has been using agents on three screens at the same time I've always known there has to be a limit and all this thread identifies is that a subscription model creates the wrong perspective. The solution I predict for all services will be the budget limit approach - because we all use AI differently and the false impression of securing this power for a month is the real issue.

I'm expecting professional development with agentic AI involved for a full work shift to cost around $50 - depending on the model being used. That is what will separate the professional from the hobbyist who may be letting their imagination and ideas run wild.

I would pay $100 for a days constant use of AI because I'd setup the tools to do 2-3 weeks worth of work in that day. That cost will be covered by clients who pay for two things. Faster deliver and my ability to steer agents so that the quality is outstanding.

You called it exactly right. Built that — it's called Valta. Hard spending limits per agent that block before the charge fires. The subscription model creates the wrong mental model entirely; per-agent budgets is the only thing that actually works. valta.co if you want to see it."

[tdarley]

Tricky one, I mean nobody going to pay what they want if they can just do it themselves. On the other hand it was always going to happen, get you hooked, skills degrade and then boom hike them prices. The are some positives, script kiddies one shotting apps that skilled developers took years to learn how to build etc.

[SC0R9I0N]

This is an interesting conversation. So far the common expectation is that a subscription covers a certain time period because the payment happens on a monthly basis.

In a way it's like saying if I rent an EV for the day it should be charged all day long - no it won't be.

As someone who has been using agents on three screens at the same time I've always known there has to be a limit and all this thread identifies is that a subscription model creates the wrong perspective. The solution I predict for all services will be the budget limit approach - because we all use AI differently and the false impression of securing this power for a month is the real issue.

I'm expecting professional development with agentic AI involved for a full work shift to cost around $50 - depending on the model being used. That is what will separate the professional from the hobbyist who may be letting their imagination and ideas run wild.

I would pay $100 for a days constant use of AI because I'd setup the tools to do 2-3 weeks worth of work in that day. That cost will be covered by clients who pay for two things. Faster deliver and my ability to steer agents so that the quality is outstanding.

your point is like 30% valid to be quite frank. while yes, you can't expect 100% uptime on an agent for an entire month, just like renting a car, the framing is completely wrong. a more apt comparison is as follows: I rent a car for a day and it lasts me about 4 hours before i have to charge it, so i rented the a more expensive car from a different rental company and it only lasted me 20 seconds before i had to recharge it. the problem isn't that people think it will last all month, the problem is that it FAR undercuts your payments. When I could pay for 2 separate claude pro subscriptions and get roughly 10x more usage for the exact same price as copilot pro+, then copilot pro+ is simply put a complete scam. i am not exaggerating when i tell you this, i was working on Opus 4.8 and was able to have 100% uptime for 3.5 hours straight without hitting my 5 hour limit. Conversely, I was using Copilot for the exact same project and I used 10% of my MONTHLY LIMIT in the span of an hour. That's not an timeframe expectation difference, that's just simply a business model difference. copilot has an insane amount of unnecessary waste when it comes to tokens when you compare it to claude or other models, and they're charging us for all of it. they either need to 1) increase the AI credits we get, or 2) make the models not just throw away millions of tokens when asked to scan 1000 lines of code when claude does the exact same thing with under 50k

[phamnamlong]

I immediately switch to Anthropic Claude (from $20/month => $100/month then $200/month), very worth it.

[cjl032224]

We no longer are able to choose a free agent after using our credits. It used to select one automatically if you chose.

This has gotten to be outrageous

[nuttyproducer]

They will never see me again until they change their plans again

[cbacary]

I use VSCode and I honestly think this is a huge failure of the VSCode's implementation of Agent mode. GHCopilot has to change their billing because a horribly implemented Agent mode is likely costing them fountains of money.

I'll say this at the top for others: remove the Github Copilot Modernization extensions. They added about 30k tokens to the system prompt regardless of whether or not your using them.

I ran out of credits so I added a Gemini API Key and enabled Logging to see the input that Copilot was passing to the API. In Agent mode there is a totally impressive waste of input tokens that drives up the cost to use the model. Input tokens are cheap, often a quarter or less the price of an Output token. However, this fact is irrelevant when VSCode spams the Input with useless tokens. From my tests, on average and depending on the chosen model, the Input tokens per request is roughly 60x the amount of output tokens per request. This was under a relatively simple task.

The problem is the agents' carefree use of tools spamming and filling up the input tokens with every useless file the agent decides to crawl, even if you've given the agent all files it should need to perform the task. With every tool call, a new request is made to the API repeating the giant system prompt once again.. For example, a frontier model such as Gemini 3.5 Flash, given a simple task and roughly 3,000 token starting prompt was quickly rate limited after accrewing 2.4 million input tokens after only two minutes. In the same time, the model outputted 40,000 Output tokens. By the tenth tool call the model made, each request made to the API was roughly 200,000 input tokens. Using a more capable model means fewer wasted tools but a higher cost.

VSCode team simply needs to think about how they utilize Input tokens and how models actually interact with tools.

[DucGiang-dev]

I completely agree with this and really hope GitHub listens to the feedback from the community.

[aiis2]

Sigh, you're all wrong. This is a business decision. Just look at the community feedback to see that—there won't be any more subsidies, and they won't come back.

[HeadNods]

Well then their users will not come back either.
I went from easily fitting my work within the monthly budget,
To using 51% in 3 days of barely using GithubCopilot.

That is like what? A more than 500% increase in "usage" cost!
In other words, from 1 day to the next their price just went up x5

If that isn't a bug, it is completely insane!

[aiis2]

Previously, every time you used it, they were losing money. My personal guess is that their business decision was to cut losses and make a profit, which would inevitably require them to be more expensive than the original computing power providers to be profitable. Next, based on feedback, they'll decide whether to continue subsidies or maintain the current system.

Unlike Codex, they won't reset the quota. Just look at those webinars; they spend almost a month discussing them. This traditional research model is really terrible. Switch to other providers and models as soon as possible.

Don't let this delay your work plans. If all else fails, go back to traditional programming methods.

[Perez4all]

If the purpose of GH is Cut losses and make a profit the formula proposed for end users unmanteinable. I've been working with different providers (including as dev, so I'm aware about costs) and they make profit but on a fair price. Everyone wons. Lets wait if GH could do something and avoids loosing clients (including me).

[SC0R9I0N]

Previously, every time you used it, they were losing money. My personal guess is that their business decision was to cut losses and make a profit, which would inevitably require them to be more expensive than the original computing power providers to be profitable. Next, based on feedback, they'll decide whether to continue subsidies or maintain the current system.

Unlike Codex, they won't reset the quota. Just look at those webinars; they spend almost a month discussing them. This traditional research model is really terrible. Switch to other providers and models as soon as possible.

Don't let this delay your work plans. If all else fails, go back to traditional programming methods.

The problem is that Claude is 50% cheaper and allows way more overall usage. I was able to use Opus 4.8 for 3 hours straight and got 10% of my WEEKLY limit. Similarly, I used copilot for an hour and am already at 10% of my MONTHLY usage. The usage based billing is fine, but not when the models you host waste millions of tokens for absolutely no reason. Me asking it to look at 1000 lines of code for bugs should not have 1m cached tokens and use 100k tokens.

[keserwan]

I completely agree with @KarimWajihKMW.
This new pricing model is expensive.

I find myself constantly checking my billing settings page https://github.com/settings/billing/ai_usage just to monitor usage and figure out how the system calculates it.

Hopefully, GitHub improves this very soon.

[clmcgrath]

i am personally switching all my tools over to use my msdn credits on foundry

[MeloWise5]

We have provided more details and a FAQ about usage-based billing in this discussion, which may be updated if new information becomes available.

To keep conversations all in one place and to ensure you receive timely updates, we’ll be closing this post and recommend you comment on or subscribe to the discussion linked above.

[MeloWise5]

Sounds like they are really listening.

[yusufmsalh]

sams for. me , some how 52% of my monthly tokens are used justed in one day (2 of july 2026) ,while it used to fit for the whole month, i am very disappointed and actively searching for alternatives.

[clmcgrath]

we burned 65000 tokens in 5 days with only 3 team members actively using it our of 15 as most of our staff was away for a conference

[clmcgrath]

the cost centers also arent correctly attributing credits so other teams are wiping out our tokens

[akki-akshat]

Same, even I have dropped the individual plan and even at my office i have told that we need to move, this is killing productivity, even $70 is worse than claude code $20 plan, and claude code was in it self less.

This is really frustrating as a user, subpar vibe coding experinace (when compared to codex or claude or even open code) and credits vanish like anything.

[bizzaretrading]

I moved to codex, cool pilot is useless with this pricing

[masslayer66-cloud]

I recently changed to codex and I was sceptical, but codex is awesome! Plus plan for me is lasting forever.. By the time my weekly quota gets to 30% its reset time.. I'm using mini model mostly but nonstop...

[LucienMP]

The primary issue for me using this via vscode is that its not even using (on auto mode) the better AI and even if it was im paying for its mistakes. At this price and pace i am burning through my tokens too fast to be of value or use. Mostly spinning fixing its mistakes. This is not a useful product, I am switching to other vendors.

1/1000th of my tokens for running a script
1/100th of my tokens for obtaining a 1-200 line script
1/10th of my tokens for fixing a single bug in a small set of C files

One day resulted in 30% token use, it results in anxiety in using AI. As others said, before i ask to perform a task i have no idea what the cost will be or if i can complete a task in budget.

Given the AI generates quite unreadable code in many cases, or code that needs AI to fix due to its sloppy design or volumous generation - its no longer valuable to me at these prices.

[rajasekhard2015]

Very worst pricing plans

[jitendrapalQA]

I think the community needs two things very clearly from GitHub Copilot's new AI Credits model:

1. Credit Preview Before Execution

Before any prompt is executed, users should be able to see an estimated credit/token cost. Similar to how cloud providers show estimated costs before running resources, Copilot should display a credit preview so developers can make informed decisions before consuming credits.

2. Fair Credit Refunds for Unsuccessful Agentic Tasks

If an agentic task does not achieve the requested outcome and the user ultimately reverts or discards all generated changes, there should be a mechanism to refund credits or provide credit compensation. Users should not be charged the full cost when the generated result is unusable and requires complete rollback.

The current model lacks transparency and puts too much risk on the user. Better cost visibility and a fair refund policy would significantly improve trust in Copilot and encourage adoption of advanced features.

Jitendra Pal | Principal QA

[imaginelapaul]

I agree with this comment. I understand needing to re-coupe costs because this endeavor into agentic is expensive and currently financially risky for all involved, but the pricing model on credits does not feel competitively priced to market share. Given how immature the product is, to be forced into a premium has instantly made my team look to alternatives. I applaud the overall effort of the co-pilot engineers -- the pricing model is clearly coming from on high -- but the business model at this point is very risky.

Credit Preview Before Execution - Yes, and to add to this, how do custom agents, sub-agents, skills, memory, codebase scanning, indexing et al: how are those factored into the product? How does my team control costs in the real world? This discussion: https://github.com/orgs/community/discussions/197089 - is inadequate to be fiscally viable or to have any realistic audit trail that I can financially justify our spend with. Any type of preflight confidence score would be a smart path forward. I would rather spend a penny, knowing that this would cost me $300 dollars to execute a feature: then execute it in the blind. IE: how am I to give guidance to my team, or my leadership on how to budget? Blank check to Github? Anyone think I'll get that approved as a line item in my 2027 budget?

Reserved Credits - Much like everything else in the cloud, I feel that for businesses to budget a reservation pricing model makes sense. For this type of premium, I want a volume discount. This pricing is not FinOps friendly, it is a bottomless pit.

Fair Credit Refunds for Unsuccessful Agentic Tasks - Exactly what I was thinking. It is hard to quantify exactly how bad AI can get, because hallucination is still a serious front-line problem. In general my team throws away about 50% of code suggestions as instructions, skills, agents, prompt engineering etc are ignored by the model. I've even had multiple models tell me that eventually they fall back to their training and ignore instructions. Do I still get charged if I press stop, because I've realized the model is reasoning itself into oblivion, or that it has the wrong supposition no matter how clear I believe I am writing?

Bottom line: caveat emptor. Understood.

Given that generative code still has serious flaws, the guidance to my team is simple: go back to basics. This pricing model is too risky and does not scale. Use AI for help, use your brain and your training. Think small, execute small. Limit the blast radius of decisions. Stop vibe coding. Use generators that are reliable. Get to know the models better. RTFM. Keep code snippets so you can re-use work. Keep a knowledge base.

I want to add one more thing in here - because I have been at keyboards since the 1980s, what co-pilot did for me was help with me arthritis and other long term repetitive task problems that have cropped up for me in the last ten years. I relied on co-pilot to help me write code that I could review as these massive sessions of keying in code for hours on end are now physically difficult to perform. Now, that is not financially viable and my disappointment is that the assistant that helped me feel less literal pain in my day is a liability. I'l persevere -- this is not woe is me -- but for others who also rely on accommodations to do their jobs, this pricing model places anyone with a physical disability at a disadvantage.

[DarceyLloyd]

Ha ha, would love to see the stats on if they refund us for all failed / prompts that did the wrong thing or broke something... That task cost you $500, however to achieve your goal we had to run 200 sub prompts and 50 sub agents all of which had their own auto generated sub prompts process's to achieve your task, and you had to steer and re-prompt to achieve that task an addition 50 times, here's a refund for $499.95 lol...

I'm not sure it's all down to MS, all they do is make git hub worse and modify vscode and copilot extension, here's the prices on openrouter:

https://openrouter.ai/minimax/minimax-m3 ($0.30 / $1.20per 1M -- 50% off at the moment, their servers are overloaded)
https://openrouter.ai/deepseek/deepseek-v4-pro ($0.435 / $0.87per 1M)

https://openrouter.ai/openai/gpt-5.3-codex ($1.75 / $14per 1M)
https://openrouter.ai/openai/gpt-5.4 ($2.50 / $15per 1M)
https://openrouter.ai/openai/gpt-5.4-pro ($30 / $180per 1M)
https://openrouter.ai/openai/gpt-5.5 ($5 / $30per 1M)
https://openrouter.ai/openai/gpt-5.5-pro ($30 / $180per 1M)

https://openrouter.ai/google/gemini-3.1-pro-preview ($2 / $12per 1M)
https://openrouter.ai/google/gemini-3.5-flash ($1.50 / $9per 1M - this model sucks lol)

etc.....

In the search for an alternate solution I've purchased deepseek v4 pro api key, minimax m3 api key, already had GLM 5.1 and today $10 on openrouter credits which offer access to countless models.

I made 1 query to deepseek v4 pro and it cost me $0.15 (I asked what's your model name)... If I ran that same query using the deepseek api key from deepseek directly it would be closer to $0.001

They are all going to do us over in the end. But I've ported and modified reasonix cli tool which has great caching for deepseek over to pi.dev yesterday, so far its working out quite well, often >90% cache hit rate which saves a lot of credit money, just it's built for deepseek but seems to be working with minimax m3 also, so fingers crossed.

[DarceyLloyd]

https://youtu.be/E_vgDoxHw2o

[valueforvalue]

"https://openrouter.ai/minimax/minimax-m3 ($0.30 / $1.20per 1M -- 50% off at the moment, their servers are overloaded)"

I haven't had any server overload issues with the minimax coding plan I currently have. I did however have issues with caching not working on the M3 model with Pi. I spent a long session doing some refactor work with the M3 model about half way through realizing that your cache optimizer was reporting 0 cache hits (Thanks for that BTW). Up until that point I had been using M2.7 so it was working nicely. Anyway apparently its a known bug.

https://pi.dev/packages/pi-minimax-m3-caching-fix

This extension will fix the issue until the actual bug is fixed upstream. Just in case you guys decide to switch to M3 plan and get scared because its eating tokens heavily.

[sunitg]

Congratulations on launching one of the WORST plans ever. Thanks for helping us to make a decision on which platform we have to switch. 👎👎👎👎👎

[rajasekhard2015]

Very worst experience with these plans Get Outlook for Android<https://aka.ms/AAb9ysg>

…

________________________________ From: Vaibhav Nagariya ***@***.***> Sent: Saturday, 06 June 2026 10:59:15 To: community/community ***@***.***> Cc: Rajasekhar ***@***.***>; Comment ***@***.***> Subject: Re: [community/community] GitHub Copilot AI Credits Are Unfair, Expensive, and Killing Real Development Workflows (Discussion #198015) I had a similar experience. I don't use copilot for heavy coding task on large repo. I use it for normal edits in bulk, scripting, automation etc. With just 10$ flat fee I never had to pay more earlier. This month I consumed all my credits within few hours that too for a very light weight task. I had to cancel my subscription and move to another solution. Very disappointed by this change. — Reply to this email directly, view it on GitHub<#198015?email_source=notifications&email_token=AE3PRFFELOY3X2XGNIFLL2D46OT2VA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNZSGAYDANJXUZZGKYLTN5XKOY3PNVWWK3TUUVSXMZLOOSWGM33PORSXEX3DNRUWG2Y#discussioncomment-17200057>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AE3PRFBENEYSFWTB32A3PMT46OT2VAVCNFSM6AAAAACZ2M72V2VHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTOMRQGAYDKNY>. You are receiving this because you commented.Message ID: ***@***.***>

[thern743]

I was skeptical of the all the nay slayers. I even upgraded to the MAX plan because I assumed having the 20k credits would be plenty. WOW was I wrong. I burned through it in a matter of HOURS. Yes, I use Opus 4.8 in XHigh and I can switch to lower models. But it seems like we're stuck choosing to spend more time on getting worse results vs getting better results quicker. I'll play around with other models and reasoning effort but so far...yikes.

[hfraga-utad]

GitHub Copilot Pro used to get me through the month. Now 10 minutes took 95% of my quota...

Thats how companies die, you know...

[DarceyLloyd]

MS wont care, they been driving git into the ground for years. They partnered with nvidia to bring the next gen of unified pc's on arm, and they didn't even bother to re-write windows for arm, slapped an emulator on it lol. They just don't care.

Windows file manager crashes more often now than ever, the desktop is more unstable than ever, do they fix it, nope, they just don't care...

Nvidia will turn on them one day, I hope, and I hope it hurts. You know what AI/ML runs on best? Linux...

By the way just scroll up and see alternatives to use than copilot...

[TalBenAmii]

Microslop greeddddd

[valueforvalue]

Try using Pi.dev I realize it isn't VSCode although some alternative has to exist with a similar philosophy to Pi. But you won't get that junked up context with Pi unless you decide to do it that way yourself. Soli Deo Gloria, Jeremy C. Morris

…

On Sat, Jun 20, 2026, 12:32 PM Carson C. ***@***.***> wrote: I use VSCode and I honestly think this is a huge failure of the VSCode's implementation of Agent mode. GHCopilot has to change their billing because a horribly implemented Agent mode is likely costing them fountains of money. I ran out of credits so I added a Gemini API Key and enabled Logging to see the input that Copilot was passing to the API. In Agent mode there is a totally impressive waste of input tokens that drives up the cost to use the model. Input tokens are cheap, often a quarter or less the price of an Output token. However, this fact is irrelevant when VSCode spams the Input with useless tokens. From my tests, on average and depending on the chosen model, the Input tokens per request is roughly 60x the amount of output tokens per request. This was under a relatively simple task. The problem is the agents' carefree use of tools spamming and filling up the input tokens with every useless file the agent decides to crawl, even if you've given the agent all files it should need to perform the task. With every tool call, a new request is made to the API and the past output from the model gets fed to the input causing the input to grow exponentially. For example, a frontier model such as Gemini 3.5 Flash, given a simple task and roughly 3,000 token starting prompt was quickly rate limited after accrewing 2.4 million input tokens after only two minutes. In the same time, the model outputted 40,000 Output tokens. By the tenth tool call the model made, each request made to the API was roughly 200,000 input tokens. Using a more capable model means fewer wasted tools but a higher cost. VSCode team simply needs to think about how they utilize Input tokens and how models actually interact with tools. — Reply to this email directly, view it on GitHub <#198015?email_source=notifications&email_token=ABFS56PBNAJ676NDDZTPGID5A3DBZA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNZTG42TSNJUUZZGKYLTN5XKO3LFNZ2GS33OUVSXMZLOOSWGM33PORSXEX3DNRUWG2Y#discussioncomment-17375954>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABFS56KOURAKYZDBU6WVQJD5A3DBZAVCNFSNUABIKJSXA33TNF2G64TZHMZTAMJVG4ZTGNBUHNCGS43DOVZXG2LPNY5TCMBRHE4DAOBSUF3AE> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[aleivaar94]

Microsoft Team, the credit system has to change. I echo everyone who has written a message in this thread. I have the Copilot Pro+ plan and notice the credits are exhausted extremely fast - only 5 days of work. Before, Copilot Pro+ was enough for a month's work. I have completely moved on to Claude Code and stopped using Copilot. How am I able to do my daily work with Claude Clode on a Claude Pro subscription ($300 yearly)? And that includes Claude.ai, Cowork and Design.

I understand that AI compute is expensive and the previous requests usage model wasn't cost effective but you have gone from one extreme to the other. Instead, why not offer discounted access to students instead of completely free (with certain amoung of free credits). Change the credit system somehow so that we get more usage out of it because currently it's unsusable. Keep the same models in all paid plans.

You are loosing consumer trust and clearly loosing revenue from canceled subscriptions instead of computed tokens. It was a bad move to remove the Opus 4.6 models from the Pro plans. The credit system is the straw that broke the camel's back.

I hope there will be changes to the Copilot credit system in the future

[OceansDMB]

I recently upgraded to the Copilot Max plan. I made just one single request for a simple feature using the Claude 4.8 Opus model, and it immediately wiped out $33 worth of credits. This is absolutely absurd.
Moreover, after the recent VS Code update, the UI no longer displays the cost multipliers or token consumption rates for each model.
What on earth is going on here? Hiding the usage rates right after hiking them feels like a deliberate attempt to keep users in the dark and blindside them with massive bills.

[DarceyLloyd]

Same happened to me using codex 3.1 on the 1st of this month. 1 prompt and 45% I think was gone. It didn't last 1/2 the day. Reduced my sub to the minimum for code completion (intelsense). I tested claude a few times, was never that impressed by it, so stayed away from it, especially with the costs.

Visit pi.dev install pi (cli coding agent tool).
And install my plugin to watch your costs in the footer (also has ssh and /usage-report features).
https://github.com/DarceyLloyd/pi-aftc-toolset

I'm now using DeepSeek v4 Pro & Flash, Minimax M3, Qwen 3.7 Max and GLM 5.2. Here's a video of me testing them with just simple webpage creation, also with GPT and codex.
https://www.youtube.com/watch?v=E_vgDoxHw2o

I've got some openrouter credits, a minimax monthly plan and a deepseek monthly plan, between them you wont ever need anything from openai or anthropic again. I've been hammering deepseek and minimax subscriptions and no issues at all, where as open router I've been using for qwen 3.7max and glm5.2, the subs to minimax and deepseek work out much better value, might do the same with GLM 5.2 after more testing.

[large]

One month away from CoPilot and getting back to THIS mess is quite repelling!

A simple task and over 650 credits of the included 3000 credits for corporates.
Seems like case closed for the CoPilot github this, WAY to expensive to keep creating stuff using AI.

[IrakliLomidze]

Root cause is Microsoft's inability to make deal with Atrophic, and we customers pay that price.

What Eventually what you get dear MS team, that majority of developers will go directly to suppliers and literally no once will use Copilot.
unfair pricing and calculator, single Opus request in 5 min cost 7-8 $.

[tretezztrzrt-svg]

hehe, noobs

[Pranavk-official]

After the update my normal request burned through 85% of my tokens in few hours, like wtf i haven't been able to use what i paid for. I just cancelled the plan and prolly switch to something that's worth the money.

[FLAK-ZOSO]

Uber was dirt cheap when it needed to hoard market share. With time passing, they needed to get profitable after the losses made to gain market. It is just how it goes, and Copilot isn't different.

For a while we will be able to use other providers' agents for cheap, but eventually we will choose to either walk (fall back to manual programming) or be prepared to pay a lot for a taxi (pay 300$/month if not more in tokens). Another option is buying a car (hosting a local agent), knowing that with one's budget (for hardware and open source models) it won't be as comfortable as an Uber ride as you will also have to drive it yourself (host it).

[soppimath-shadakshari]

Wow! Well articulated.
That's the reality.

[MasterSTime126]

This is ridiculous. I asked Copilot to just make a SMALL plan, but it started doing... I don't know how to call it

[MasterSTime126]

It did not plan, but wrote the whole code itself

[Sajjadarasham]

shiiiit i think that i only people that have this problem!
i bought 40$ Proplust account and think that i dont need to have stress for copilot credits!
after three days that i used that, that gives me 75% used!

[clarkprince]

Cancelled my Copilot Max subscription. Burnt through my credits in 6 days, and by day 10 I'd blown $150 in additional usage. Even though the Claude models were the issue, I added a custom model, plugged in my Gemini API key and used Gemini 3.5 Flash, and it was still burning through $2 a query, so something must be internally wrong with Copilot.

Mind you, on the same codebase 3 months ago I was using a Copilot Pro+ account and never hit any limit throughout the month. Decided to give Claude Code a try. Much better experience so far, and fair pricing compared to Copilot

[Sajjadarasham]

so i think we must be switch to another agents
what's an alternative of this? thats can i use in visual studio?

[codekoenig]

so i think we must be switch to another agents what's an alternative of this? thats can i use in visual studio?

Claude Code is likely the best option today which is kind of wild considering it was a comparably expensive option for the longest time.

You can use Claude’s CLI along Visual Studio which is not ideal but I find it to be an quite OK experience as direct integration (so the agent can see your open file, marked code etc) is not that relevant today anymore as this „assistant like“ working mode is replaced by autonomous working agents anyhow. The truth is that there is unfortunately no (official) agent integration available for Visual Studio besides Copilot.

If you absolutely need it I would recommend switching to Jetbrains Rider as it is much more widely supported if you want full IDE integration, for example for Claude. Alternatively, VSCode can be an option too of course but Rider is the much more complete IDE of course.

Microsoft is killing their ecosystem anyhow so way not switch your IDE too while you’re at it.

[clarkprince]

There is now a VSCode extension for Claude Code. Works exactly like Copilot you don't need to rely on the CLI like the early days.

[braverock]

the challenge seems to be that the copilot agent harness is super inefficient with tokens, and even with a more efficient harness and just using the copilot keys, caching seems to be limited or completely not used. when testing with opencode or pi or Zed harness, no cache tokens are reported at all. when using the copilot harness with logging on, input tokens balloon with each new query. even cheap models burn through context window MUCH faster using the copilot harness than using other harnesses, but cache is something on the server side, not on the harness side, so that's 100% on Microsoft to sort through.

Certainly people shouldn't expect tons of usage on a model like Opus 4.8, but when you can pay the listed per-token fees for using Opus on other servers (OpenRouter, OpenCode, Google Vertex from our tests) and Copilot is burning 10-20x the tokens, the problem is with the harness. Cheaper models like 3.1 Flash or Sonnet 4.6 are also using significantly more tokens than they do on any other harness.

[ekalkutin]

I was and I am using VS Code, paid office family subscription, I was very loyal to VSCode, when most ran to Cursor/Antigravity/whatever, i was stick with VSCode + Copilot.
Shit, in order to keep things somehow done, i have to use fucking haiku 4.5 (0.33x), otherwise if I use opus/sonnet (9x-27x) i get 10%-20% monthly quota burned immediately (10-15min session).
GPT 5.5 (x57) - i never tried, because i guess i will loss all quota (Pro+ 39$ plan).
Couple days ago I tried codex 20$. Used GPT 5.5 medium for 30min session - it costed for me 8% of weekly limit.
Goodbye VSCode/Copilot e.t.c Going to codex/zed

Kinda same situation