Cognition's SWE-2: A Cheaper, Faster Coding Agent for Devin
Cognition's SWE-2 matches near-frontier coding performance at a fraction of the cost, but only inside Devin. Here's what agencies should know.
Thank you! Your submission has been received!
For the past year, picking an AI coding agent has meant picking a trade-off: the most capable models are also the most expensive to run for hours at a time, and cost adds up fast when an agent is burning tokens on every file read, test run, and retry. Cognition — the company behind the autonomous coding agent Devin — just shipped a model aimed squarely at that trade-off. On September 10, 2026, Cognition released SWE-2, a coding model it says matches near-frontier performance at a fraction of the cost of the models it's competing with.
SWE-2 is post-trained from Moonshot AI's Kimi K3, a 2.8-trillion-parameter base model, using a single reinforcement-learning run. According to Cognition's own benchmarks, it scores 50.0% on FrontierCode 1.1 Main — within a point of Anthropic's Claude Fable 5.1 (50.9%) — while costing a claimed 64% less to run at that performance level. Cognition also reports SWE-2 lands within a few points of OpenAI's GPT-6 Astra at roughly a quarter of the cost.
Other reported benchmark results include 73.0% on DeepSWE 1.1, 92.8% on Terminal-Bench 2.1, and 27.3% on the newer, harder Terminal-Bench 4. It's also Cognition's first model with selectable reasoning-effort levels, letting you trade speed for depth depending on the task.
Compared to Cognition's previous model, SWE-1.7, the company says SWE-2 uses 58% fewer agent steps and costs 81% less per run on average to reach a better score — a meaningful efficiency jump if those numbers hold up in independent testing.
This is the part worth flagging clearly: SWE-2 has no open weights and no standalone API. It's not something you can plug into your own agent harness or self-host. It runs exclusively inside Devin — currently available in Devin Desktop and the Devin CLI, with Devin Web and Devin Fusion rolling out. If your workflow depends on an open-weight model or direct API access for custom tooling, SWE-2 isn't an option; you're adopting the Devin product, not just a model.
Cost-per-task is becoming as important a metric as raw benchmark score for teams running coding agents in production, especially agencies running agents across multiple client codebases simultaneously. A model that closes most of the performance gap with frontier models at a fraction of the cost changes the economics of how many agent-hours you can afford to run, and on what kind of work — routine refactors and test-writing become much cheaper to hand off to an agent when the cost per run drops by more than half.
It's also a signal about where the coding-agent market is heading: rather than every lab chasing the single best benchmark score, we're seeing more players optimize explicitly for the cost-performance frontier — squeezing out the same or better results with fewer steps and less compute per task. For any team evaluating coding agents as part of a client delivery pipeline, cost-efficiency claims like these are worth testing directly on your own codebase rather than taking benchmark numbers at face value, since real-world repos rarely match curated benchmark suites exactly.
Teams already using Devin, or evaluating it against alternatives like Claude Code, Cursor, or Codex-based tools, should test SWE-2 directly — the selectable reasoning-effort levels in particular are useful for tuning cost against task complexity on a per-job basis. Teams that need an open-weight model or a coding agent they can embed into their own custom harness (including, notably, teams building on OpenAI's new Agents API or an open framework) should note that SWE-2 doesn't fit that use case at all — it's a Devin-exclusive model, not a general-purpose API.
RAG assistants on your documents, AI agents that act in your tools, and LLM features inside your product.
Our cutting-edge features simplify collaboration and creativity, making your workflow intuitive and efficient. Transform your vision into reality effortlessly with Hadidiz Flow.
