I purposefully left it a little while after the re-launch before writing this just to see if it would settle down, or if it would all change again.

It seems like they are still having issues with their new highly specialised, unique, flexible commercial model. (If you're read my blog before you know that over-use of hype-words means it's gonna be bullshit).

Before the punchline

Personally, I'm dying to let you have it up front, but let's take a look at why the old plans were a liability and had to be axed.

I should also add, at times during this article I will be talking to you in a way that I assume the people who were responsible for the new plans must think about you. Except that I'm doing it for comic effect. I actually kind of like you, well, the vast majority of you.

Old Pricing Model

Where the old model was based around 'requests' being either normal or premium, with premium requests having the notion of a model multiplier - different models might use more or less premium requests per premium request - so easy everybody can understand it.

500 premium requests in the GitHub Copilot Pro plan would be 500 actual messages to something like GPT-5.4 or Sonnet 4.6 that had a 1x multiplier, but only 166 messages to Opus 4.6 or just 66 messages to GPT 5.5, with the most 'expensive' being Opus 4.7 that had a 15x multiplier meaning your 500 premium requests would be used up in just 33 messages.

Smart pedants out there might be screaming at me right now, saying "500 divided by 15 is 33.33333 recurring"... yeah, well what are you going to do with a 1/3rd of a message huh? talk like caveman, save words, ugg? Oy, don't laugh, its' a real thing - check it out the Caveman Compression repo on GitHub.

Agents Killed The Subscription Star

It's the same with all 'bundled' products for anything (mobile minute or data bundles are a good example), the seller assumes a spread of usage, some you win, some you lose, overall it gives you an attractive and simplistic fixed price offer to go to market with - ultimately gaining you a larger customer base. You size the bundle so that most people simply don't reach their limits every month. People's eyes are bigger than their belly's - I might use 5Gb, so I'll buy 10Gb to be safe, but for a little bit more I could get 25G, but I don't want to be throttled or cut off if I end up using extra out of the blue just when I need it, so I best get 100Gb, ah to hell with it, I'll treat myself, top package, unlimited data, sold.

As soon as you sell something to customers that has a different cost-base to what you're charged by your supplier, you have risk that needs to be absorbed. Price it right and you're laughing, price it wrong (or people's usage changes across the board ← foreboding) then you're bleeding out.

And that's where I think we are, it's down to a fundamental change in how people were vibe coding across the board... Microsoft say so themselves.

With the improvements in AI models' ability to work longer without needing human input, larger context length to fill up with more of your codebase and extra thinking.. Combined with gimmicks like 'ultrawork' or 'RalphLoops' that became popular (where the prompts were prepended with specific instructions telling the AI model to keep on working, never stop until the task has been completed, no pauses, no checkpoints, no questions, JFDI (just fucking do it)).

It changed people's habits from what used to be a regular back and forth conversation with the AI, that suddenly evolved into a single prompt easily triggering 20+ minutes of full-on autonomous work. People were taking the piss without even knowing it, just because the newer models and coding apps let them work that way.

I personally blame OpenClaw, just because I can and it sound reasonable without needing to provide evidence... People were authenticating OpenCode against their GitHub Copilot accounts in droves (no evidence), and then both due to their lobsters running autonomously 24/7, and the trend leaning towards using longer and more detailed prompts to kick off long-horizon tasks, burning through tokens before finishing simply broke their 'request-based' plan (still no evidence).

Suddenly people realised they were on to a good thing, and inadvertently costing Microsoft far more in tokens, than they were paying for in those 'Premium Requests' bundled in to their plan, even at the excessive multipliers for those frontier models.

So things had to change... (assuming that is what happened, since I have no evidence for it either way).

New Pricing Model

It's obvious Microsoft have well and truly learnt their lesson, because they are now being very transparent - you pay for what you use.

Sorry, did I say transparent? Microsoft being who they are just couldn't let it be that simple.. money for tokens, bah, no chance, that's a mug's game... so they invented a new 'usage-based' model where you use money to buy AI Credits, and those AI Credits can be exchanged for tokens by way of... wait for it... converting back into money.

If that sounds like everybody else's API-plan but with extra steps, let me explain in shorter simple words you'll have a better chance to understand...

Yes, you are correct.

Totally not to be confused with the well established pay-as-you-go API models where you get a million tokens for a set price. Microsoft have introduced a much needed abstraction layer called AI Credits!

The full cost of your GitHub Copilot plan is converted directly into AI Credits.. So Pro would cost you $10 per month, you get 1000 AI Credits to spend. Pro+ is $39 per month, so you get... can you guess... yeah! 3900 AI Credits - well done math(s) nerd, you can move a decimal point!

The exact same simple formula applies to the business plans too.. Business costs $19 and that gives you 1900 AI Credits and Enterprise costs $39 and also gets you 3900 AI Credits.

AI Credits, AI Credits, AI Credits...

How much is an AI Credit worth?

To find out how much actual usage you'd get, rather than tell you how many tokens you can use per AI Credit, the models are all priced in the very well known and already established rates of "$ per million tokens".

Or if you listen to Microsoft - Money gets you AI Credits. And AI Credits gets you tokens... but don't you dare for one second think that money gets you tokens.

So it's very simple to work out how many AI Credits you need, buy converting them back into dollars! 3900 AI Credits divided by 1000 gives you $39 to spend on actual AI model tokens (hmmm "$39" I'm sure I saw that figure somewhere else earlier). Amazing huh! Look at your confused face, your little mind blown away by the incredible value that this should sound like! 3900 - thirty nine thousand AI credits, that's easily 2-3 new apps per month!! You're gonna be rich!

GPT-5.5 is $5.00 per 1M input + $30 per 1M output tokens.

Wonderful, so all you have to do is follow this simple set of steps... Luckily most of you can't count this high to notice the absurdity.

1. spend $39.00 of your money to buy the Pro+ or Enterprise plan
2. The value of the plan is converted to AI Credits by multiplying the dollars by 1000.
3. $39 x 1000 = 3900 AI Credits
4. Check the price list to see how many dollars it costs per 1M of each model's input / cached / output tokens.
5. Take your 3900 AI Credits and convert back to dollars give's you $39.00
6. Divide $39.00 by the cost of tokens
7. You have now successfully calculated the usage-based model.

You'd be able to use a little bit more than the full context window size (of the un-restricted GPT-5.5 model - that you can't actually use via GitHub Copilot because it's limited to 400k instead) before you use all of your $39 (if seeing the figure in dollars upsets you, and you're missing the silky goodness of AI Credits, this is for you, thicky! It's 3900!!).

But I have an annual plan!

Well you're still screwed, even with the 3 options they give you.

• Stay as you are, but it will not be able to renew, instead you'll drop down to the 'free' plan.
• Cancel and get a refund for the remaining time.
• Upgrade to a Pro+ monthly plan, using the pro-rated credit to use instead of getting it as a refund.

Sweet, I'll stay!

Even if you're on the annual plan, where you'd think you get what you paid for until it expires, well you'd be wrong.

Assuming you've decided not to get a refund, and you fancy staying for the remainder of your annual plan to ride out the shake-up with your existing 'request-based' model in your ivory tower lording it over the peasants having to pay per token.. erm. ahem.. erm pay to get AI Credits that are exchanged for tokens, sorry... You are hit with a change to your model multipliers.

The most notable uplifts are...

• Claude Opus 4.6 goes from 3x to... 27x - yes, 27x
• Claude Opus 4.7 goes from 15x also up to 27x
• Claude Sonnet 4.6 goes from 1x up to 9x
• Google Gemini 3 & 3.1 Pro go from 1x up to 6x
• OpenAI GPT-5.4 also from 1x to 6x
• OpenAI GPT-5.5 goes from 7.5x up to TBD!!

And at the other end of the scale, there used to be free access (zero multiplier) to several of the smaller or older models.

• GPT-4o / 4o mini / 5 mini / and Raptor mini go from 0x to 0.33x
• And GPT 4.1 goes from 0x to 1x

New Limits!

And just to help you control yourself you greedy twat, Microsoft are also introducing two different limits to slow you down, you'll get a session limit, and a weekly limit...

How long is a session? How dare you ask that, but for those out there that have never heard of a "week"... Don't worry, they've got you fam 😉

How long is a 'session', I've no idea, based on my college days, a drinking session could end up being almost as long as a week! But not these days, not unless you count having an addiction to Pepsi Max and bulk buying cans when there's an offer on, and then opening one after another until they're all gone as a session.

What do I think?

There are still some nice features for free, like the code completions and next edit suggestions, and, erm.. the warm fuzzy feeling you get from being a Microsoft fanboi and refusing to see the harsh reality that is Microsoft let their monopoly-play slip away because they actually underestimated how popular their product would be... now that's a turn up for the books.

It used to be one of the few, or only AI-Coding subscription plans that gives you access to a selection of models from multiple vendors per 'message', with a small selection of unlimited models thrown in.

But after this change, you can get a even more from pre-paying OpenRouter / OpenCode Zen / Vercel AI Gateway or any other AI proxy / token broker. Just buy $39 worth of credit on the account, set your monthly limit to the same, and crack on.

And maybe that's just the push you needed to try the awesome OpenCode.