Any fellow software engineers using paid GitHub copilot?
Much to my chagrin, the company I work for has done a lot in terms of steering/ pushing all software development be done through AI for some time now.
And what gives me much grin, GitHub changed their pricing structure for copilot. I'll skip the details the key fact is what used to be about $30/month per person + maybe few bucks in overages is now resulting in us hitting our usage cap on the 2nd day of the month. Overage costs this month will be hundreds of dollars per developer. I know this is an unexpected expense as I mentioned it casually to our CTO who had no idea.
I'm curious if this is going to force them to rethink the AI strategy. The incessant pushing to use more and more AI maybe will finally bite them on the ass so much they have to ask us to stop or pull back? Or maybe they'll just plunder our salaries, who knows.
I'm curious if anyone else is in the same situation.
Not exclusively copilot but Claude code. It was the opposite at the enterprise company I work at. They're still in the stage of 'don't worry about cost' and so every person is allocated around $300 in usage per month and any increases are automatically approved. They're chasing this dragon that developer velocity will proportionally lead to customer acquisition and retention which will lead to increased revenue. I'm not quite sure that's how it works.
Same here. It's more money to the bubble isn't it?
I mean, at some point either we are going to start to see more heavy usage restrictions or models will become more efficient and it won't matter.
I suspect it'll happen around the same time. Already Opus is capable of handling a lot of day to day tasks. I've been reading that Fable is a step up and sounds like it's actually overkill for most dev work.
If an Opus tier model was much cheaper to run, then most people would probably be fine?
I'm sure they'd continue to push the boundary and see how good at reasoning they can get it, but those models could become prohibitively expensive if it's pushing past "human reasoning" and they could probably get away with it I imagine.
At the enterprise level is Claude code all per token cost or is this for a subscription?
Enterprise pays per seat and per token.
(https://support.claude.com/en/articles/11526368-how-am-i-billed-for-my-enterprise-plan)
I mean it seams pretty easy to say “if you want me to use AI then you actually have to pay for it, otherwise I’m just gonna keep doing things how I was”, right?
I doubt they’ll pull back any time soon. The corporate class is smitten with AI and I don’t think that will change for now. I’ve personally seen CEOs moan about spend on other things while simultaneously saying they must spend as much as possible on AI or they’ll “be left behind”. Coworkers were actively tracked down and applauded for being top AI spenders.
As most SWEs seem to be, I’m also on the receiving end of the AI push by the executive class. Not loving it. BUT at least my company is willing to put their money where their mouth is. If tomorrow I hit a Claude limit that was obstructing me, the CEO would personally approve increasing that limit without hesitating. Is that healthy? No I don’t think so, but at least it’s not hypocritical.
This is the part I don’t understand about the current “AI push”: Isn’t awarding employees by most LLM (token) spend just “paid by LOC” all over again?
Has management (in the aggregate) really learned nothing, or alternatively, forgotten every lesson from that time?
I mean, even if you’re not familiar with the SWE side of this (hi)story, maybe you ought to have heard of Goodhart’s law instead if you’re in a position to green-light these massive budgets with absolutely no way of tracking, let alone predicting them?
I know I’d be uncomfortable if my company could just sack my role as the responsible person once things went south.
Management has been instructed by top level to make sure it's used.
Someone somewhere said "buy this product" and they're high enough up the chain to ensure that everyone in the company has to use the product, making sense or not.
That way in the short term while prices are low and effects are hard to prove, things look, at worst, like it's helping.
Management is mostly well aware of how stupid this all is, and it's going to take some sacrifices at the alter of stupidity before it gets rolled back. We did it with LOC, we did it with cloud, we're doing it with AI. So long as marketing/sales types wind up higher than engineering it'll keep happening.
Loc is exactly where my head went too. Awarded by token usage is bizarre but any hot new proxy for productivity gets attention I guess
We have a 20€/month allowance, which is now... useless. Claude Opus 4.6+ is now way too costly, and several colleagues were out of tokens in less than 2 days. Yesterday I heard someone say that the era of artisanal code was over. Lmao yeah, good luck with that and your 20€.
To be fair, depends on the model. I use Deepseek-v4-pro and 20€ could probably last me over a month of work. Not saying I agree with the "the era of artisanal code is over" take though, of course.
Never used Deepseek. How does that compare to Opus 4.6/4.7? Because I can probably last a month with the cheapest models, but I'm certainly not replacing my "artisanal" code with those. Anything older than Opus 4.6 for a remotely complex task will produce garbage and waste tokens.
(Not talking about auto-completion here, which is unlimited in our plans)
I haven't used Claude Code, so can't really compare, but benchmarks show Deepseek on the same level as Opus 4.6. Very much worth it IMO considering it's essentially free. Works pretty well for me, but of course depends on your tasks.
Deepseek v4 pro is like magic! I mean it does some stupid shit from time to time, but is easily corrected and costs practically no money.
But be careful, the 1/4 API cost it is offering is - entirely as expected - payed for by your data.
An interesting framing of this I saw for coding application is "why would I be against having more code in the dataset that matches my coding style and problem area?"
(But of course I probably wouldn't use the official Deepseek for any kind of personal advice)
Thanks for the info, I'll pass the word and see if we can use it.
Putting aside the benchmarks, since it varies widely depending on which ones you look at, Deepseek is most definitely not on par with Opus 4.6.
Unless you factor cost in, then Deepseek is lightyears ahead of Opus.
I'm confused by your comment. Did you mean the opposite?
Meaning that Opus is dramatically better than Deepseek for complex coding tasks, but if you include cost in the calculation, Deepseek looks a lot better.
"unless you factor cost in; in that case deepseek is lightyears ahead of Opus" is the intended reading I suppose
Is that 20€ in API credits or a 20€ subscription? The latter is quite a bit more usage and the limits reset in 5 hour windows. Both options are available in enterprise and team setups.
An alternative you might suggest is Open AI at 20/month subscriptions. For Claude you need 100/month subscriptions for serious usage but with GPT 5.4 you can get a lot farther on 20/month.
But no matter how much they pay, artisanal code is only mostly dead. Miracle Max's pill will come in the form of everyone realizing that having no one left that can actually code isn't working out as planned!
Those subscriptions are not available to enterprise customers. They are wildly worth the cost, but are basically bait and switch.
You're right, I shouldn't have included enterprise... The team plans offer subscription billing (as opposed to API prices)
Team plan also caps out at 150 seats.
I'm on the $10 plan and have pretty much resorted to GPT-5 mini on low reasoning, with caveman on ultra mode.
For what I need it to do (analyze a large and obscure database, match some binary data with some C data structures and help me write parsers) it does well, but I do miss abusing Claude Sonnet 4.6 and not worrying about limits.
We stopped using Copilot as soon as Cursor hit the scene. Then we stopped using that as soon as IDE-less workflows via Claude became common. And now everyone's all hyped about Codex. It's frankly really annoying. Everything but Cursor is available via company accounts still, though, so I usually start the month on Claude, switch to Codex midway, then finish month limping along with Copilot.
Remember when our jobs were rewarding? That was a fun nearly 2 decades of my career.
I'm literally in mourning of this profession. It was perfect for me, now it'll never be the same and I think I'll never get the same joy out of it.
Has Codex gotten dramatically better recently? When I use it for work, it seems to make dumb decisions very confidently, and it takes forever to deliberate stupid things before it actually does what you want it to do.
I really don't get it either. A few of our devs were hyping up this hybrid approach of planning in codex, then executing in Claude, but I've yet to see a benefit to that? At best Codex is as good as Claude is, but certainly not better.
It's substantially better than copilot though!
So in 2021-2023 I was kind of drifting into being a manager and I want enjoying it, so I did a purposeful reset back to a more hands on role…. Just in time for that role to turn into basically the same focus just directed at AI but without the personal connections or advancement opportunities…
We dropped GH copilot because it just didn't keep up with other options capabilities. Github cant help but fumble the ball.
No - we didn't get good results with Copilot. My team is using Devin extensively, and it actually really does seem to be hugely beneficial for us.
I forgot that was a thing! That was the first major one to hit the scene and it really flopped from what I saw. Makes sense it's more capable now, just needed some time for the models to catch up
So with Devin it's basically leveraging the current LLMS, I'm, using it with Claud Opus 4.6. The way the AI knowledge base is built up within their web environment, using hints and the way it "remembers" flows and instructions on how to do things like compilation tests, runtime tests and what have you is great. As I say, we're getting a lot out of it. Me especially as one of the most senior technical engineers I can drive it well, and get high quality outputs, verified by me, in rapid time. It's cut literal years off a project we're doing.
After this project comes to an end though - who knows. Back to cutting code, I guess. A task for which I've had increasingly less enthusiasm for as the years roll on. I want to be designing and implementing solutions, and using my knowledge to design, validate and verify, not writing screeds of
MOVE 10 TO FACTOR
COMPUTE NEW-BALANCE = BALANCE * FACTOR
PERFORM PRINT-INVOICE
for another 10 years.
Right now, cursor is the new hotness still but just because we’re in the sweetheart “first taste is extremely reduced” phase. Most SWEs use cursor since our limits are higher and can use the advanced models plus MAX mode.
Our copilot usage has a bucket each month that resets and some of the more expensive models are disabled.
I speculate like the 2010s “race to the cloud” with vendor lock-in, it’ll be this way with AI tools soon enough. Do we rehire some of those we used AI as an excuse for quarterly cuts? Do we just eat the higher token costs?
Ironically, you can’t outsource AI, aside from building your own infra with open source models and that’s a whole different can of worms.
So between outsourcing SWE to Poland and India, using cheap(for now) AI and retaining minimal talent usually at the mid-senior levels, some interesting business decisions will have to be made. Especially with junior pipelines drying up. I suspect the gap and issues will only get worse before they get better.