SWE-2 Matches Frontier Coding at a Third of the Cost, Then Fails One Test

A new coding model gets close to the top for far less money, and has one obvious hole

On September 10, 2026, Cognition — the company behind the Devin coding agent — released a new model called SWE-2. The headline in one line: it scored 50.0% on FrontierCode 1.1 Main against 50.9% for Anthropic's Fable 5.1, a gap of less than a single point, and Cognition says running it costs 64% less to reach that same score. But the story does not end there. On Terminal-Bench 4, SWE-2 scored only 27.3%, against 55.8% for Fable 5.1 and 57.9% for GPT-6 Astra — a gap of roughly thirty points in a single column.

That contradiction is the real lesson for businesses, not the cheap number on its own. We are entering a phase where comparing models stops at "which one is smarter" and starts at a sharper question: smarter at which kind of work, at what cost, and where does it break down.

The published numbers, as they are

Cognition published a direct comparison table between SWE-2, its own base model, and two frontier-tier models. Here is the full reading:

• FrontierCode 1.1 Main: SWE-2 at 50.0%, base model Kimi K3 at 44.2%, GPT-6 Astra at 53.3%, Fable 5.1 at 50.9%.
• DeepSWE 1.1: SWE-2 at 73.0%, above Fable 5.1 at 67.4% and close to Astra at 74.1%.
• Terminal-Bench 2.1: SWE-2 at 92.8%, the highest in the table ahead of 91.4% for Fable 5.1 and 89.9% for Astra.
• Terminal-Bench 4: SWE-2 at only 27.3%, against 55.8% and 57.9%. This is the column it clearly loses.

In three of four columns SWE-2 stands level with more expensive models or ahead of them; in one column it falls off a cliff. Note also that it beat its own base, Kimi K3, in every column — that improvement is precisely what Cognition's training added.

How it was built: a large open model, then specialized training on top

SWE-2 was not built from scratch. Its base is Kimi K3, the open-weights model from Moonshot AI at 2.8 trillion parameters. What Cognition did was post-training with reinforcement learning aimed at one job: real software engineering inside actual code repositories.

The interesting engineering detail is that the model ships with three selectable effort levels — medium, high, and max. These were not trained in three separate runs or as three specialist models, but in a single training run using a cost penalty built into the reward function: success minus the cost of the attempt, multiplied by a coefficient that changes with the effort level. The result is one model that knows when to be frugal and when to dig deeper, instead of forcing you to manage three models.

The practical difference is tangible: at the medium effort level, the team says the model cuts the number of steps by 58% and cost by 81% on average compared with the previous SWE-1.7 release on the same benchmark. In project language, that means less circling around a problem before reaching a solution.

The important constraint before you get excited: no open weights, no API

This needs to be stated plainly. SWE-2 is not a model you can download onto your own servers, and you cannot call it from your application through a standalone API. It runs inside the Devin platform only: the desktop app and the command line first, expanding to Web and Fusion. Which means the cost economics Cognition is describing are tied to subscribing to its platform, not to the freedom to run the model wherever you want.

That distinction matters for anyone comparing options. An open-weights model gives you sovereignty over your data and the freedom to host inside the Kingdom; this is a closed product bought as a service. If your reason for hunting a cheaper alternative was keeping code inside your own environment, this announcement does not solve your problem. If your reason was shrinking the development-tooling bill, it might.

What this means for your business in practice

Three takeaways for an owner who pays a development bill or is weighing AI tools for a technical team:

• Specialization has become a shortcut to competing: a company far smaller than the frontier labs took an open model and trained it on one job, and now it competes at that specific job. The same logic applies to your project: a mid-sized model tuned on your data and your procedures usually beats a larger model operating with no context.
• One number in an announcement is never enough: if you looked only at FrontierCode you would conclude SWE-2 is a full replacement. The Terminal-Bench 4 column says the opposite for a particular category of tasks. Always ask: on which benchmark, what does it measure, and where is the gap.
• Cost is not the per-call price: your real cost is the number of attempts until a correct result. A cheaper model that circles endlessly can be more expensive in practice — which is exactly what the cost penalty in SWE-2's training is trying to address.

How we handle announcements like this at Origami

We do not tie client systems to a single model. We build an intermediate layer that lets the model be swapped behind the scenes, and we measure on the client's real tasks rather than on published benchmarks: actual contracts, actual invoices, actual customer conversations. That way an announcement like SWE-2 becomes an opportunity you can test in a week, not a decision that rebuilds your system.

The practical step if you have a development team: take three real tasks from your repository, run them on your current tool and on the cheaper alternative, and measure just three numbers — did the task succeed, how many attempts did it need, what did it cost. After that the decision becomes obvious, with no reliance on anyone's announcement table.

Sources

• Cognition's official SWE-2 announcement and benchmark table: cognition.com/blog/swe-2
• Moonshot AI — the Kimi K3 open-weights base model: moonshotai.github.io