Devin and SWE-2 Are How I'm Keeping My AI Bill Sane

It’s no secret that AI costs start stacking up once you’re using it for more than the odd question. A $20 plan is fine for casual use. Lean on agents all day and you’re soon paying for the bigger plans, topping up with credits, or both.

The plans haven’t got any more generous lately either. Anthropic wound back the boosted Claude Code weekly limits on the 14th. At DevDay on the 29th, OpenAI reopened its $200 Pro plan with a smaller allowance than before and added a $500 tier for people who want the fast stuff. Copilot has gone to dollar-denominated AI credits too. The sticker prices mostly sit still while the allowance underneath them shrinks, and they’re all hoping you don’t do the maths.

I do the maths. I run agents for my day job, for contract work and for the side projects I keep shipping this year. If I pointed a frontier model at all of that, I’d be paying more for AI each month than I pay for my car.

So for the last month Devin has been my daily driver, and that bill looks a lot less scary.

If you last looked at Devin back when it was $500 a month and a viral demo video, it’s a different product now. There’s Devin Desktop, a CLI and the web version, and the self-serve plans have a weekly quota like everything else. You pick your model from a list that has most of the big names on it. I live in the CLI because that’s where I already spend my day.

SWE-2 is the reason I stayed.

Cognition released it on 10 September. It’s their own coding model, post-trained on top of Moonshot’s Kimi K3, and it comes in Medium, High and Max effort levels. Cognition claims it’s close to the frontier on their evals. I’m not going to pretend it beats Fable 5.1 or GPT-6 Astra on a hard benchmark, because it doesn’t. On Terminal-Bench 4 it’s about half their score.

I don’t care much, because most of my day isn’t Terminal-Bench 4.

Most of my day is “add a settings screen that saves these five fields”, “this test started failing after the dependency bump, work out why”, “wire this webhook into the existing queue”. Unglamorous, well-scoped work. SWE-2 eats it for breakfast. It reads the code before it touches anything, it runs the tests without being told, and it keeps going until the thing works. It feels like a hard-working junior who’s had three coffees and doesn’t need to be told twice.

It’s also efficient in a way I didn’t expect. Cognition says the medium effort mode takes 58% fewer turns than SWE-1.7 and costs 81% less on average. I can’t verify their numbers, but I can see my own. It doesn’t wander off on a ten-file exploration to change one line. It doesn’t write me a little essay before every edit. It does the job and stops.

Fusion is where I started having fun. It pairs two models. A lead model does the planning, the thinking through the nasty bits and the review at the end. A sidekick does the implementation, runs the builds and checks the tests pass. You talk to one Devin and the handoff happens behind the scenes, with each model keeping its own cached context so you’re not paying to re-read everything every time they swap.

My pairing is Opus 5.5 as the lead and SWE-2 as the sidekick, and it’s a monster.

Opus 5.5 is the model I trust most with judgement calls. It’s the one I want deciding how a migration should be staged, or noticing that a one-line change to a billing calculation touches three other places. It’s also the one that chews through quota the fastest when you let it do the typing. Fusion means it doesn’t have to do the typing. Opus works out what should happen, SWE-2 goes and does it, and Opus checks the result before I see it.

Last week I used it to pull a client’s job processing out of a cron script and into a proper queue with retries. Opus caught that two of the jobs weren’t idempotent before a single line was written, which is exactly the kind of thing I’d have found at 11pm after a duplicate email went out. SWE-2 wrote the workers, the migrations and about forty tests. I read the diff over lunch and changed two things.

Cognition’s own numbers say Fusion runs around 36% cheaper than the same lead model on its own in Claude Code, for close to the same score. Treat a vendor’s benchmark with the usual suspicion. My own /stats output agrees with the direction of it, if not the exact percentage.

What I didn’t see coming was my weekly usage barely moving. I’ll get to Thursday and I’ve used a sliver of the allowance, after a week where I’d normally have been watching the meter and rationing prompts. SWE-2 does so much of the work so cheaply that the quota just sits there.

And because I’m paying for it, that bothers me. Unused quota on a weekly reset is money I’ve set on fire. So by Friday I’m finding excuses to switch to the expensive models. I’ll throw Opus 5.5 at a full refactor on its own, or ask a frontier model to review a codebase I already know is fine, purely so the week’s allowance doesn’t go to waste.

I’m aware that’s a ridiculous problem to have. A year ago I was hitting limits on Wednesday and sulking until the reset. Now I’m the guy at the all-you-can-eat buffet going back for dessert he doesn’t want because he paid for it.

Fair warning on timing. SWE-2 has been free on Devin’s self-serve plans since launch, and that free run ends tomorrow, 15 October. After that it’s list price with a 75% discount until the end of December. Some of that untouched quota is down to the promo, and I’ll be watching what my week looks like once SWE-2 starts counting.

I’m not too worried. Even at full price SWE-2 costs a fraction of what the frontier models do per token, and it uses fewer tokens to get there. The worst case is that I stop topping up my week with Opus on a Friday afternoon, which my wallet would be fine with.

I’ve also stopped thinking of one model as the answer. Using the most expensive model for everything made sense when the plans were generous and the companies were burning cash to win you over. That period’s ending. The setup that makes sense now is a cheap, capable model doing the bulk of the work, with a smart one making the calls, and Devin has made that the default instead of something I’d have to build myself.

If your AI spend has crept up this year and you haven’t noticed, have a look at your last three invoices. Then try SWE-2 on a week of normal work and see how much of your allowance is left by Friday.