SWE-2 Is My Daily Driver Now, and It's Free Until October 15

Here’s how I’ve been working lately. SWE-2 builds the first implementation, the whole thing, and then I get Fable 5.1 or GPT-6 Astra to run over the top of it and clean things up. The two most expensive models I pay for have been demoted to tidying up after a model that currently costs me nothing, and I’m completely fine with that.

Cognition’s last model, SWE-1.7, scored 42% on FrontierCode, which is the sort of number that gets you used for renaming variables. SWE-2 came out on the 10th of September, scores 50% on the same benchmark, and it’s become my daily driver in under two weeks.

If you haven’t been following along, it’s built on top of Moonshot’s Kimi K3, which is a 2.8 trillion parameter model, and then Cognition have thrown a mountain of their own reinforcement learning at it. You get three effort levels (medium, high and max) and it lives inside Devin Desktop and the Devin CLI. There’s no API, which I’ll moan about further down.

I’ll just say it: the quality of what comes out of this thing is incredible. It’s a very intelligent model, but plenty of models are intelligent and still manage to be annoying to work with because they wander off, second guess themselves and re-read the same file five times. SWE-2 is proficient, which I think is the better word. It gets on with the job.

That’s why the build first, clean up second routine works so well for me. What SWE-2 hands over is already a proper implementation, so when Fable or Astra come in behind it they’re polishing, catching the odd thing it missed and tightening up the bits I’m fussy about. If they had to rewrite it there’d be no point and I’d go back to giving them the whole job.

Cognition have a number that lines up with this. On FrontierCode, SWE-1.7 took 127 steps on average to get through a task, and SWE-2 on medium takes 53 and gets a better score doing it. That’s less than half the faffing about, and you can feel it when you’re sitting there watching it work.

The benchmarks, quickly, because everyone wants them. FrontierCode 1.1 has SWE-2 on 50.0%, Fable 5.1 on 50.9%, Astra on 53.3% and GPT-5.6 Sol on 47.5%. DeepSWE 1.1 has it on 73.0%, which is ahead of Fable 5.1 (67.4%) and Sol (72.7%) and about a point behind Astra. It’s top of the table on Terminal-Bench 2.1 with 92.8%.

And then Terminal-Bench 4 happens and it gets 27.3% while Fable gets 55.8% and Astra gets 57.9%. Ouch. Credit to Cognition for putting that row in their own announcement, because I wouldn’t have. So if your day is long, messy, open-ended terminal jobs that run for hours, the flagships are still the flagships, and I think that gap is a big part of why having one of them go over SWE-2’s work afterwards pays off.

Right, money. At the moment there isn’t any to talk about, because SWE-2 is free on all the self-serve plans until October 15. That’s the $20 Pro plan, the $200 Max plan and Teams. Devin bills your quota based on the token price of whichever model you’ve picked, and SWE-2’s token price right now is zero, so it doesn’t eat into your quota at all. Twenty bucks gets you a frontier-ish coding model to absolutely hammer for the next three and a bit weeks. That’s silly. Go and do it.

After the 15th you pay list price, and Cognition have tucked that away in a little note in the Devin docs where nobody’s going to see it: $3 per million input tokens, $15 per million output and 30 cents per million for cached input. Fable 5.1 and Astra both charge $10 and $50. So token for token SWE-2 is 70% cheaper than either of them, and that’ll be where the up to 70% line in their launch tweet came from.

Their blog post says 64% cheaper though, and I wanted to know why the two numbers don’t match, so I went digging in the footnotes. The 64% is per task. They list Fable 5.1 on medium as costing $3.28 per FrontierCode task for its 50.9%. Take 64% off $3.28 and you get roughly $1.18 per task for SWE-2. They don’t print that figure anywhere, I’ve just done the sum for them.

What that tells you is a bit interesting. If every token is 70% cheaper but the finished task only comes out 64% cheaper, then SWE-2 is burning through about 20% more tokens than Fable to get the same job done (0.36 over 0.30 is 1.2, if you want to check me). So it’s a little chattier, or it reads a little more. When the tokens cost this little, who cares?

I also wanted to know what $1.18 looks like in tokens, and this bit is me guessing, so take it as a rough sketch. Agent runs are mostly cached input, because the model re-reads the same growing pile of context on every step. Something like 1.5 million cached tokens (45 cents), 100K of fresh input (30 cents) and 28K of output (42 cents) gets you to $1.17. You can shuffle those around a fair bit and still land pretty close to it.

They also claim SWE-2 gets within a few points of Astra for a quarter of the cost. Flip that around and Astra is running at something like $4.70 a task, for 3.3 extra points. I know which side of that trade I’m on.

So say you’re doing ten tasks a day that are about as big as a FrontierCode problem, and you work 22 days a month. That’s 220 tasks. SWE-2 does that for about $260. Fable 5.1 on medium wants about $720 and Astra wants over $1,000. Most of what I ask a model to do on a normal day is smaller than a benchmark task, so think of those as the top end, but the gap between them stays the same whatever size your work is.

The maths backs up my routine too. The big expensive chunk of any task is building it, all those steps and all that context being re-read over and over. That’s the part I’ve moved to the $3 model. The flagship comes in at the end, reads a finished diff and makes some edits, which is a fraction of the tokens at the $10 rate.

(If you’re on an enterprise ACU plan it’s even sillier. You get 75% off list until December 31, which is 75 cents in and $3.75 out, and that same $1.18 task costs you about 30 cents.)

Cognition have a built-in version of roughly what I’m doing, called Fusion, and it shipped the day after SWE-2. You pick a lead model and a sidekick. The lead does the planning, makes the judgement calls when something’s ambiguous, and reviews the work at the end. The sidekick does the typing: writes the code, runs the builds and the tests, reports back. They each keep their own context and their own prompt cache and only pass short briefs and results between them, so you aren’t paying to push one enormous conversation through two models.

With Fable 5.1 or Astra as the lead and SWE-2 as the sidekick, it’s a beast. You get the flagship’s judgement on the hard calls and SWE-2’s speed and price on everything else, and you don’t have to babysit the handover like I do when I’m doing it by hand.

Artificial Analysis ran it through their Coding Agent Index, so these aren’t Cognition’s own numbers. Fusion with Fable 5.1 on xhigh and SWE-2 on medium scored 61.7 and cost $7.90 a task. Claude Code with Fable 5.1 by itself scored 62.2 and cost $12.36. That’s half a point for 36% less money. The Astra pairing came in at 58.9 for $4.54 against Codex by itself on 61.6 for $7.47, which is 39% cheaper but you’re giving up nearly three points, so Fable is the lead I’d go with.

And until October 15, SWE-2’s share of that bill is nothing.

Turning it on is just /fusion in the CLI or Desktop, then you choose your lead, the effort and the sidekick. When you’re done, /session-stats breaks the cost down by model, which I like, because it means you can check all my maths above against your own work instead of taking my word for it.

My complaints are short. There’s no API, so I can’t plug it into my own tools or use it in any harness other than Cognition’s, and I’d really like to. Fusion doesn’t work on the free tier or on the old credit-based plans. And I wouldn’t hand SWE-2 a sprawling multi-hour terminal job and walk away, because that Terminal-Bench 4 score is there for a reason.

If you’re already paying for Devin, use the next few weeks to put your normal work through it while it’s free, and run /session-stats on a handful of sessions so October 16 isn’t a surprise. And try the build with SWE-2, clean up with a flagship thing, even if you do it manually like me.