SWE-2, Cognition’s new model, scores 50.0% on FrontierCode 1.1 Main versus 50.9% for Fable 5.1: nine tenths of a difference, at 64% lower cost. That’s the whole argument, and whether it holds up or not depends on a number that Cognition publishes without dwelling on it.
SWE-2 launched on September 10, 2026, post-trained from Kimi K3, Moonshot’s 2.8 trillion parameter model whose open weights we covered when it landed. It’s available since announcement day on Devin Desktop and in the Devin CLI, with rollout underway to Devin Web and Fusion.
What is SWE-2 and what is Devin?
SWE-2 is a code model, not a code tool. You don’t install SWE-2: you install Devin, Cognition’s agentic coding product, and select SWE-2 within it. The distinction matters more here than with other models, because SWE-2 has no public API and no open weights. The Devin CLI, desktop application, web interface, and Fusion are the only places it runs.
Devin is an agent that takes a task, plans it, edits files, executes commands and iterates. The same category as Claude Code or Cursor’s agent mode, with more of the loop hidden behind the product. It has a free plan, a $20/month Pro plan, and a $200/month Max plan, plus team pricing at $80/month per team plus $40/seat for developers.
Why did the cheap tier stop being the dumb one?
This is the really new part, and it’s a training result rather than a product decision.
Most models that offer reasoning effort levels achieve them by training a model and then varying at inference time how much it’s allowed to think, or by training each tier separately. Cognition’s RL algorithm trains all three effort levels in a single run, with a reward function penalized by cost where the penalty weight corresponds to the local slope of the cost-performance Pareto frontier at each level. In plain terms: each level is optimized against what a token is worth at that level, within the same training pass. Cognition also reports a “length-weighted reward baseline” that stabilizes training without additional compute.
The observable consequence is behavior. SWE-2 medium makes its first real file edit after a median of 18 steps. SWE-1.7 needed 48. Average steps per run on FrontierCode 1.1 Main:
| Model | Average Steps |
|---|---|
| SWE-1.7 | 127 |
| SWE-2 medium | 53 |
| SWE-2 high | 80 |
| SWE-2 max | 98 |
Against its own predecessor, SWE-2 medium scores higher taking 58% fewer turns and costing 81% less on average. The cheap tier stopped spinning its wheels before getting to work: that’s what you buy by training effort levels in a single run.
How much does Claude Code cost versus SWE-2?
You need to read carefully here, because the comparison shifts twice.
Today, through October 10, 2026, SWE-2 is free on the $20/month Pro plan: Cognition lists “free SWE-2 and open source models through October 10, 2026” as a Pro benefit. During that window, the cost question has a trivial answer for any individual developer: $20 flat per month, no per-token metering on SWE-2. Nothing with usage-based pricing competes with that.
After that date, you’re back to the published comparison: 64% cheaper than Fable 5.1 for a 0.9 point difference on the benchmark, and roughly a quarter the cost of GPT-6 Astra. Cognition clarifies that these figures use list prices, including public discounts.
Two things worth keeping. First, each of those figures is from Cognition, measured on Cognition’s benchmark: FrontierCode is their own harness. That it’s self-reported doesn’t make it false, but it doesn’t make it independently validated either. Second, the promotion date is the part that sustains the cost answer right now, and it’s four weeks from this publication.
Where does SWE-2 fail?
On Terminal-Bench 4. SWE-2 scores 27.3%; Fable 5.1 scores 55.8%. Cognition publishes it without comment.
The gap gets more interesting alongside the other number from the same post: 92.8% on Terminal-Bench 2.1. A model that tops 92.8% on the old benchmark and lands at 27.3% on the new one is saying something about generalization, and the Hacker News discussion landed exactly on that reading. The counterargument—that TB2.1 is saturated and the comparison is unfair—is reasonable, but weakens when older models score above SWE-2 on TB4.
The comparison that should worry Cognition more: DeepSeek v4.1 Flash scores, per data cited in that discussion, 31.2% on Terminal-Bench 4, above SWE-2, at drastically lower API cost and with open weights. If your reason for looking at SWE-2 is cost per task, that’s the alternative to price it against. And unlike SWE-2, it doesn’t require adopting a new agent product just to evaluate it.
How do you install it?
The Devin CLI installs with one command:
curl -fsSL https://cli.devin.ai/install.sh | bash
SWE-2 is also on Devin Desktop since launch day, and rolling out to Devin Web and Fusion.
One heads-up worth knowing before you go looking: at publication time, devin.ai/cli still advertised “Opus 4.7, GPT 5.5, and SWE-1.6” in its model list. Launch day pages lag behind launch day models. Cognition hasn’t published a dedicated SWE-2 configuration page, so if you need to pin the model explicitly instead of choosing it in the interface, check the CLI’s own help output before any guide written this week.
Should you switch?
Try it if you already pay for Devin, or if the free window through October makes evaluating it cost nothing but time. For four weeks, it’s a good deal.
Don’t switch on the benchmark table. The efficiency result is real and well documented: fewer turns, faster first edit, cheaper runs, and a training method that explains why instead of just asserting it. The parity claim on the frontier is thinner: 0.9 points on the provider’s own benchmark, undermined by a 28-point deficit on another benchmark that same provider publishes. And the whole comparison is a snapshot in time: it expires at the next frontier launch, which historically measures in weeks.
What to take from this launch isn’t the ranking position. It’s that training effort levels jointly, against what a token is worth at each level, made the cheap tier competent. That technique is going to show up in other people’s models, and it’s going to matter longer than where SWE-2 sits on any table.
Claude Code pricing #code-agents devin kimi benchmarks