Anyone using Claude Code? Is it better to switch to Claude Code directly or to the Pro + plan of GitHub?

I think Claude Code is a better product than Copilot. I use it in codespaces quite frequently, though it wouldn't be a good fit for my classroom usage since it does not have any easy way to get access for students. So I used to use copilot in my teaching as a result, but now I will have to switch.

BTW, getting Claude Code to login to a Cluade pro account on codespaces can be a bit cumbersome. I made a little in-browser JS script to do some of the work for me. (Or, more precisely, I had Claude Code build me a little js script.) It's at: https://pngwen.github.io/codespaces-claude-login/

This is a absolute nonsense from GitHub if you want to block, do it completely else got it to everyone

Seriously, no more Opus ? I'll take a look into Minimax Token plan ... ($10, good model, generous limits)

I just got Minimax M2.7 running locally, I'm tempted to try switching to that using Cline or Roo (not sure which is better yet). It's just kinda slow.

Claude models became unusually slow recently. They're running out of compute power.

so now if i want to use an opus model i have to use the x7.5 xd. big L github

Yeah, I upgraded to Pro+ for more requests for Opus, but that's just too much and I'll run out too quickly. I downgraded to Pro and will stick to Sonnet, but if Qwen3.6-122b is released soon and is good I might just run that locally. Or an 80b Coder model.

I think that's an Anthropic thing, which is a shame, as it looks like GH are really starting to set up to use mainly Anthropic, from looking at how the extension JS is changing.

https://platform.claude.com/docs/en/build-with-claude/extended-thinking.md

I literally moved from a student plan to a paid plan in order to keep using Anthropic's Sonnet 4.6 and Opus 4.6, look where it got us now...

Right now any of my student plan models are not Working

I was first getting constant crashing during prompts and now I won't even be able to use the specific models I paid for. Very disappointed.

my student plan copilot pro downgraded to free plan any one has this issues?

This is so bad. I didn't realize it also impacted people who already prepaid for a year subscription. This is a really bad look.

@artidones We understand this is frustrating. Calling this out from the post above with regards to paid subscriptions:

If you hit unexpected limits or these changes just don't work for you, you can cancel your Pro or Pro+ subscription and you will not be charged for April usage. Please reach out to GitHub support between April 20 – May 20 for a refund.

Imagine this is done on day one of the subscription. You are promises a month of X and then on the first day you only get Y. And then you get told "you can always cancel it if you don't like it". Yeah. That's the bare minimum and borders on false advertising. Oh well it was probably covered on page 461 of some TOS.

If you struggle then do a batter job at communicating changes. I don't need to ask an AI that tells me to go here to see its gone all tits up. You have the channels to let people know this better.

This is so bad. I didn't realize it also impacted people who already prepaid for a year subscription. This is a really bad look.

Agreed. I was also thinking that I probably made a mistake of not opting to purchase the yearly subscription before hand expecting the new policy shouldn't be affecting the agreements and contracts made before the change. Thats how it should be. This is purely a breach of contract, isn't it?

@artidones We understand this is frustrating. Calling this out from the post above with regards to paid subscriptions:

If you hit unexpected limits or these changes just don't work for you, you can cancel your Pro or Pro+ subscription and you will not be charged for April usage. Please reach out to GitHub support between April 20 – May 20 for a refund.

I agree that people can with their own free will (and money) opt-out if they are dissatisfied. That's how it should be.

The most difficult part and perhaps where frustration kicks in, is that there was not a lot of time to act. So if working with these models under the pre-existing constraints were the motor oil, you might go a little crazy seeing prices more than double, with limitations of how much oil you can put in your vehicle. Suddenly, trips that made the world spin for you halts or significantly alters short-term goals. The expectation put on users is the ability to immediately be able to reroute, and that seems a bit unreasonable to an extend.

It's tough, because Github Copilot has genuinely been a good and cheap product in comparison to other providers — and this move simply forces users to make "seemingly" unnecessary risks out of the blue.

A lot of us know that the demand has been extremely high on all providers, and as providers you navigate whale infested waters no doubt. But for perhaps your average Joe, it may seem sudden leaving them unprepared, naked and vulnerable.

So while the move IMO is more than welcome, a heads-up in the users IDE/CLI weeks prior seems to have been missed opportunities to be in the best interest of the users. Despite the chaos, thank you for your efforts and your services, I understand the reasoning and also the frustration in the other end.

Agreed, as a solo dev, copilot pro was a very good, cheap, and useful product. Alongside the frustration of these new awsome rate limits, I've also lost the most useful model. (yay). I'm not paying microsoft $40 just because they could not be bothered to use some of their billions to keep the service going, and people happy. Nope. Wonder what their PR team is going to do about this. I suppose I might as well try and get a good gpu and go local.....

@artidones We understand this is frustrating. Calling this out from the post above with regards to paid subscriptions:

If you hit unexpected limits or these changes just don't work for you, you can cancel your Pro or Pro+ subscription and you will not be charged for April usage. Please reach out to GitHub support between April 20 – May 20 for a refund.

You understand? I doubt that. Typical microsoft business behaviour which translates to "we don't care".
Example - I have an issue open for like 3 weeks where github charged tokens for agent sessions that simply stopped "due to technical reasons". You just ignored that issue. the latest development unfolding now is only following the same pattern on a larger scale: you don't care.

dear github agent admin (doubt that real people follow this), I do the sentiment analysis for you: users are angry.

Yeah, and the replies from the GitHub community admin above are seriously tone deaf. 🙈

I really liked Copilot, and even though I have a paid subscription to Claude, I preferred to use Copilot as it is flexible enough and allows you to change models. This is a kick in the nuts.

Last week I notice the max plan in the model list request that Copilot makes. I'm guessing a $100-$200 "Unlimited with fair use" plan is coming soon, on par with Codex/Clause Code plans.

Personally I'd like to see it. I'd just rather they just rip the band aid off rather than the slow push out the door.

点了

nowhere near enough to keep paying. i have a year sub that i converted to pro+ and this bs about the weekly token cap is clearly BAIT-AND-SWITCH. why TF would I pay microsoft/github $40/month for something I get free or $10/month (minimax, glm, qwen, etc). minimax has been far more stable than anthropic or copilot. there are already many models that beat opus 4.6 on avg benchmarks, and as you all know opus is widely seen as king. (it isn't though. pull up pinchbench dot com and click to sort by avg score, which if you are intelligent it is immediately obvious why that matters more than the outlier "best ever" that happened on that one test one time. notice how many beat opus. (currently 3 but this varies))

nem de longe o suficiente para continuar pagando. Tenho um sub anual que converti para pro+ e esta melhor sobre o limite semanal de tokens é claramente BAIT-AND-SWITCH. Por que diabos eu pagaria à Microsoft/Github US$ 40/mês por algo que ganho de graça ou US$ 10/mês (Minimax, GLM, Qwen, etc). O Minimax tem sido muito mais estável do que o Antropic ou o Copilot. já existem muitos modelos que superam o opus 4.6 em benchmarks médios e, como todos sabem, o opus é amplamente visto como rei. (mas não é. Abra o pinchbench dot com e clique para classificar por pontuação média, o que, se você for inteligente, fica imediatamente óbvio por que é importante mais do que o "melhor de todos os tempos" atípico que aconteceu naquele teste uma vez. observe quantas beat opus. (atualmente 3, mas isso varia))

And where do you pay 10 dollars to access these Chinese models?

Also interested

nem de longe o suficiente para continuar pagando. Tenho um sub anual que converti para pro+ e esta melhor sobre o limite semanal de tokens é claramente BAIT-AND-SWITCH. Por que diabos eu pagaria à Microsoft/Github US$ 40/mês por algo que ganho de graça ou US$ 10/mês (Minimax, GLM, Qwen, etc). O Minimax tem sido muito mais estável do que o Antropic ou o Copilot. já existem muitos modelos que superam o opus 4.6 em benchmarks médios e, como todos sabem, o opus é amplamente visto como rei. (mas não é. Abra o pinchbench dot com e clique para classificar por pontuação média, o que, se você for inteligente, fica imediatamente óbvio por que é importante mais do que o "melhor de todos os tempos" atípico que aconteceu naquele teste uma vez. observe quantas beat opus. (atualmente 3, mas isso varia))

And where do you pay 10 dollars to access these Chinese models?

You can access some of those models through tools like Ollama and OpenCode Go, and if you use Z.ai’s GLM 5.1, the pricing can be very reasonable for the amount of work you get out of it.

In my case, with a heavy multi-subagent workflow, I can usually push it close to the limit and still get a lot of usable time before the 5-hour reset window. Most people will probably never hit that level of usage, so for average users it can stretch quite a bit further.

I bought the yearly Pro plan for around $145 back in December, and last I checked it looked closer to $180. I locked it in early because I suspected the price would go up later.

As for the model quality, I would compare GLM more to the Sonnet 4.5 style of model behavior. Earlier versions felt closer to that, and while they can sometimes get a bit verbose, they are still genuinely useful models to work with. I have not yet put 5.1 through my longest workflows, so I would not overstate that part yet, but the early results are promising.

OpenCode and Ollama also make these models easier to test, but the real answer is still the same: do the math for your own workflow. That is what matters most. Pricing only looks good or bad once you measure it against how many hours of real work you can actually get out of it.

• its gonna be priced more, its a promo price lol

Deleting 4.5/4.6 opus models, its because they actually working, and now they want you to pay more and more money for something that actually working, I guess me and most of others dont really need the 4.7 model, because 4.6 doing the work fine, so why they deleted the 4.5/4.6 opus models, because they want force you pay more and more money for something that actually working 4.7
this is insane from my perspective, now i know how they thinking i cant trust this github anymore

Deleting 4.5/4.6 opus models, its because they actually working, and now they want you to pay more and more money for something that actually working, I guess me and most of others dont really need the 4.7 model, because 4.6 doing the work fine, so why they deleted the 4.5/4.6 opus models, because they want force you pay more and more money for something that actually working 4.7 this is insane from my perspective, now i know how they thinking i cant trust this github anymore

Great analysis, That's exactly how I would put it.

Remover o Opus é um tiro no pé...

Ao menos para mim, pois com a exclusão deles para meu plano (PRO +), me vem à pergunta, vale a pena continuar usando o github, ou vale a pena migrar diretamente pra Claude?

This is tremendously terrible behaviour GH. Nothing more than a money grab once again, we are already paying 3X the tokens for the model and do so sparingly. Why would you make this an incremental cost on your side? We know its possible because we were already using it.

Big Thumbs Down from an org based on enablement for all.

An organisation owned by one of the most evil companies on earth, none other than everybody's favourite, Microsoft.

It feels like the plane is being repaired while burning on the ground...

it feels like the plane was fine, but the staff decided to sabotage it anyway

Feels like the "copilot" just doesn't feel like it and opens the door mid-flight.

If there isn't a pro-rata refund option available for annual subscribers, within the UK/EU you can get your credit card to refund under a chargeback. It's classified as "bait and switch" under "Unfair and blacklisted commercial practices".
https://europa.eu/youreurope/citizens/consumers/unfair-treatment/unfair-commercial-practices

Quite simply you cannot charge someone a subscription then change the service to the customer's detriment half way through the period without stating this in the initial agreement.

It's like booking a suite in a hotel, then when you get there being told you have to sleep in the dumpster; but you'll be charged the same price. Not ethical, not moral, not legal.

This forum feels like you're screaming into the wind and no one at GitHub / Microsoft is listening. I wrote a breakdown on exactly how this 'rug pull' put profits above our productivity. Microsoft doesn't have a compute problem; they have a trust problem that they manufactured. Here's why we need to move to a BYOK (Bring Your Own Key) model to protect our productivity: https://cabeza.cc/the-copilot-rug-pull-when-tech-giants-lobotomize-your-workflow/

@anjeleno do the math first. even with the updated pricing MS/GH are definitely not getting any profit off of copilot and still subsidize it. but hey - go for it. cancel your Pro, try BYOK and tell us how much you saved

@psimsa I have... Before I make any final decision, I'm waiting to see if Microsoft reverses course. If I do rip off the band-aid, I'll gladly share the BYOK math once the migration is done, but you're missing the point. It isn't just about trying to save $10 a month; it's about predictable performance. Right now, losing the unthrottled reasoning of Opus 4.6 and being forced to burn tokens 2.5x faster (jumping from a 3x to a 7.5x multiplier) for a degraded, "Medium thinking" model just isn't worth the price of the subscription. Whether Microsoft is taking a loss on compute or not, silently rug-pulling your paying users isn't the solution. Honesty and transparency should still matter. In the meantime, considering reactivating my Cursor subscription and finding other workarounds to get reliable access to the models I've built workflows around and have come to depend on.

I'm not worried about Microsoft's profitability.

Financially, it was a year of record performance. Revenue was $281.7 billion, up 15 percent. Operating income grew 17 percent to $128.5 billion. And Azure surpassed $75 billion in revenue for the first time, up 34 percent. These results reflect the growing demand for our platform and the trust customers are placing in us. We take neither for granted.

But don't take my word for it. Take theirs!

https://www.microsoft.com/investor/reports/ar25/index.html#:~:text=Financially%2C%20it%20was%20a%20year,We%20take%20neither%20for%20granted.

Microsoft has a trust problem, not a financial problem.

And yet, Microsoft is just removing a popular, well used service. Great idea.

What are the exact limits for session and weekly usage?

They will not tell you because else its hard to "tweak" them, aka tighten the limits even more without you noticing ...

Unless I was unlucky with recent 2-3 requests (although it was gpt5.4 medium so should eat like 1-2% max), or github being slow (both possible) I think they reduced the total number of requests as well. I have been at ~60-70% and im at 80%.

MAYBE I AM INCORRECT. BUT. Without any number from GitHub I WILL assume I am correct and pretend to be angry, while looking for alternative.

Google at least has balls to show you all the graphs of your usage. Here? I can only trust that they are not doing anything, while staring at "%"

My case (after 3 weeks of 1-5 hours per day), this started happening yesterday:

this is literally what happend to me
So dissapointing
I felt like I got catched into scam lol

Yeah this is bullshit, been paying $200/month in overage charges on my Pro account. And they just up and kill it. Sure hope DeepSeek 4 is as good as they are promising.

Because CC is trash.. but you wanna use Opus model

I am on the same boat. What other options do you guys have? This is s***

How big is the context window for Opus 4.7?

192k Tokens. Same as 4.6

What other alternatives do we have? Do you think is worth it to pay the max plan from Claude? I am afraid that the usage limits there will be tricky as well, since a lot of users there are just complaining. The thing is that is cost $100, far more than copilot pro+.
Using anthropic API could be another alternative, but no idea about the pricing per token.

Any thoughts?

I am in a simliar situation like you are. First shocked and angry. Now I also consider other options. My first try was to use other models in Copilot. It seemed to be like in early times. I wanted to change coding logic (Flutter) and gave the same instructions to other models. Nothing felt like Opus 4.6. But I cannot compare anymore. I just find Gemini 3.1 Pro nice to use. Additionally I installed Codex today.

May be more people can write their ways too.

Yeah, see this, straight from the horse's mouth: https://messagebase.net/Read/6346/wow-visual-studio-202x-is-a-dead-product

They're doing better than ever financially

GitHub Copilot’s team is basically saying: ‘Go use Codex or ClaudeCode.’ Fine by me—I’ll walk away with you.

Yes, that's how I said goodbye to windsurf forever. And I don't plan on going back there. GitHub Copilot, followed the same path.

Hi @annasalova92-netizen,

Your Copilot plan includes premium requests, which use more advanced models or features and count against your monthly premium request allowance.

The premium request quota renews on the first day of the month, regardless of the subscription renewal date.

For additional information, visit our docs: About billing for individual GitHub Copilot plans.

Sure, they take a bad decision about this. We need a new solution for example OPENCODE with another MODEL

Aren't they all trimmed down in github? Like opus isnt it meant to have 1M context but we get 192K.

GPT‑5.4 in the GitHub Copilot Pro subscription and in OpenAI’s native services are, technically, different products built on the same foundation. The native OpenAI version (Codex/ChatGPT Pro) provides full access to the model’s capabilities, including a true million‑token context window, maximum xhigh reasoning levels, and autonomous agent functions. The implementation in GitHub Copilot Pro is a ‘compromise’ version: it is optimized for low subscription cost and high speed, which inevitably leads to behavior that professional developers perceive as a degradation of intelligence.

The difference in programming effectiveness is not caused by poor GPT‑5.4 quality, but by its operation under severe compute constraints within the Azure/Copilot ecosystem. For everyday boilerplate generation and simple functions, Copilot remains a leader thanks to its IDE integration. However, for senior‑level architectural tasks, deep refactoring, and complex bug analysis, the native OpenAI subscription or direct access to the Codex API is indispensable, because only they allow the model to ‘think’ as long as the complexity of the task requires.

DeepSeek 3.2 and Kimi k2.6 turned out to be smarter than GPT‑5.4 (at its maximum reasoning level) in GitHub Copilot

DeepSeek 3.2 and Kimi k2.6 turned out to be smarter than GPT‑5.4 (at its maximum reasoning level) in GitHub Copilot

hmmm I don't have those