[vmbm]

Yeah, I think it is pretty obvious at this point that users have gotten over their skis a bit on the fixed rate plans. I suspect a lot of folks were playing around with agentic workflows on a $10-200/month plan and then started implementing them at their companies under enterprise accounts not realizing that API based billing would result in easily 10-100x the costs. Running hundreds of agents 24/7 is all fun and games when you can do it for beer money. Not as fun when it is super yacht money.

As an anecdote, Github is changing their copilot plan to usage based billing next month. They released a tool that allows their users to estimate what their bills will look like under the new plan based on their past usage. There are some screenshots online from users showing their plans will go from $40/month to $3-5k/month. I imagine this is happening everywhere. These tools absolutely can do more than they were capable of just six months ago. But if the true costs are as high as it appears, folks are going to be much more judicious with how they use them moving forward.

[256BitChris]

GitHub doesn't have access to its own models so it has to pay the prices of the model it uses.

People confuse price with cost all the time. The price of Opus has dropped from $75/1M output tokens to $25. That's the price. The cost is much lower and according to Dario, about a month ago, they had about a 73% margin.

I don't understand how anyone would use GitHub copilot...it's basically running a custom harness and using close to API pricing for Opus. This is why Microsoft is cooked in this game.

But yeah, I don't understand why people switch from subscription to API prices for Claude. They're way higher, but again that's price and not necessarily the cost to Anthropic to serve.

[vmbm]

For the longest time Copilot was the best deal in town. For $10 a month you would get ~1300 requests. A single prompt was counted as a request, and it didn't matter how many tokens were used or how many loops the agent did. It was a spectacular deal. Of course, now that they are moving to API based billing the plan is a not really a deal at all. The monthly plan is essentially pre-paying for API credits, which are use it or lose it and they are not discounted in any way.

And FYI, most enterprise accounts were forced to switch to a hybrid monthly seat license plus API based usage earlier this year. So that is why we are probably seeing so much alarm over Ai bills at the enterprise level. Companies went whole hog on agentic workflows not fully appreciating the costs structures of their new plans. Didn't help that pretty much every VC and board was probably breathing down their neck that if they didn't jump on the AI bandwagon they would get left behind.

[riffraff]

> according to Dario, about a month ago, they had about a 73% margin.

Do you have a source for this? I missed this and it does not match my impression.

[SatvikBeri]

Not OP, but Dario said on Dwarkesh's podcast 3 months ago that their gross margins are "significantly higher than 50%"

[Incipient]

I'm really annoyed at github for only allowing export of a single month of usage.

There were a lot of people (here included) that were absolutely abusing github - getting Opus to generate its own subagents for days on end with a single premium request. If THAT'S $3k/mo, I'm honestly not worried.

I was just asking sonnet, 5.4, opus, in single agent session to fix a problem for 20 minutes...if THAT'S $3k/mo, then AI is truly cooked.

[simonw]

I went digging and found this one - https://www.reddit.com/r/GithubCopilot/comments/1tbfkui/ill_... - billing $39.00, usage-based billing $5,851.77

[vmbm]

Yeah, I don't doubt that there was some abuse going on. But keep in mind there are plenty of reports out there that companies were setting up tokenmaxing leaderboards essentially encouraging this type of behavior from their employees. So I think the broader point here is that the insane growth that Anthropic has experienced over the last six months might be a temporary blip due to companies scrambling to take advantage of these more powerful agentic capabilities while not fully understanding the resulting costs.

I think for your use case, the most likely outcome is that you are going to need to be on a $100-200 month plan if you want access to the cutting edge models. But on the other end of the spectrum, you could probably get closer to $10-20 month using Chinese models. My usage was closer to yours, with the occasional tokenmaxxing sprint just to experiment a bit and test out the limits of the plan. I am not quite sure what I will be doing next month once it moves to API billing but I suspect I will move to openrouter and one of the open source CLI harnesses.