Cognition SWE-2 and Devin Pricing: Benchmarks, Costs and What the $48B Valuation Means for AI Coding Tools
Cognition released SWE-2, a coding model post-trained from Kimi K3, two days after raising $2 billion at a $48 billion valuation. Benchmarks against GPT-6 Astra and Fable 5.1, what is included in Devin's $20 Pro plan, and how to evaluate it.
Cognition, the company behind the Devin coding agent, had a busy September 2026. On September 8 it announced a $2 billion funding round at a $48 billion valuation, and two days later it released SWE-2, its most capable in-house coding model. The two announcements are connected but answer different questions: the funding is about whether AI coding can support several large companies, and SWE-2 is about whether Cognition can deliver near-frontier coding quality at a lower cost. This article covers both, plus what developers actually get on Devin’s plans.
SWE-2: what it is
According to Cognition’s launch post, SWE-2 is post-trained from Moonshot AI’s Kimi K3, a 2.8-trillion-parameter open model that had already undergone extensive reinforcement learning for agentic coding. Cognition applied its own RL on top, which it says added five to six points on many benchmarks. Key training changes include:
- Cost penalties in RL, applied per reasoning-effort level in a single training run, so the model learns to be efficient at each level.
- Three times as many RL environments as before, with instruction-following overlays.
- Faster rollout serving using an online draft model and low-precision kernels.
Weights are not published. SWE-2 is available in Devin Desktop and CLI, with rollout to Devin Web and Fusion.
Benchmarks (self-reported)
Cognition’s published comparison:
| Benchmark | SWE-2 | Kimi K3 | GPT-6 Astra | Fable 5.1 | GPT-5.6 Sol | SWE-1.7 |
|---|---|---|---|---|---|---|
| FrontierCode 1.1 Main | 50.0% | 44.2% | 53.3% | 50.9% | 47.5% | 42.0% |
| DeepSWE 1.1 | 73.0% | 68.5% | 74.1% | 67.4% | 72.7% | 37.7% |
| Terminal-Bench 2.1 | 92.8% | 88.3% | 89.9% | 91.4% | 88.8% | 81.5% |
| Terminal-Bench 4.0 | 27.3% | 21.5% | 57.9% | 55.8% | 37.3% | 7.6% |
Cognition’s headline claim is cost: on FrontierCode 1.1 Main it says SWE-2 comes within one point of Fable 5.1 while being 64% cheaper, and within a few points of GPT-6 Astra at a quarter of the cost. AI-TLDR’s model summary notes that FrontierCode is Cognition’s own benchmark, which grades whether a project maintainer would merge the pull request, using tests, rubrics and other verifiers.
Two points deserve attention. First, all figures are published by Cognition, including on its own benchmark. Second, SWE-2 trails badly on Terminal-Bench 4.0, the hardest test in the table, at 27.3% against 57.9% for GPT-6 Astra. For the most difficult long-horizon terminal tasks, frontier models still lead clearly.
Efficiency compared with SWE-1.7
The biggest practical change may be behaviour. Cognition says SWE-2 at medium effort scores higher than SWE-1.7 while taking 58% fewer turns and costing 81% less on average on FrontierCode. It makes its first real code edit after a median of 18 steps, compared with 48 for SWE-1.7, which users had criticised for over-exploring simple tasks. Cognition also describes better end-to-end tests, more willingness to find alternative routes when blocked, and a habit of verifying a user’s hypothesis instead of simply agreeing. These are internal observations, so they are worth checking on your own repositories.
Why post-training on open models matters
Cognition’s approach, also used for SWE-1.7 on Kimi K2.7, avoids the cost of pre-training a model from scratch. The company starts from the strongest open base it can find and spends its compute on reinforcement learning in software environments that resemble Devin’s real work. The benefit is speed and lower cost; the dependency is that each generation relies on the quality and licence terms of someone else’s base model. For buyers, the practical question is less who trained the base and more whether the resulting model performs reliably on their own codebase.
What Devin’s plans include
According to Devin’s pricing page, checked on November 12, 2026:
| Plan | Price | Notes |
|---|---|---|
| Free | $0 | Light quota, limited model availability |
| Pro | $20/month | Higher quotas, frontier models from several providers, free use of SWE-2, cloud agents |
| Max | $200/month | Significantly higher quotas |
| Teams | $80/month plus $40 per full developer seat | Up to 100 users, SSO, cost controls |
| Enterprise | Custom | VPC deployment, SCIM, dedicated support |
Usage allowances refresh daily and weekly, and extra usage can be bought at API pricing. “Free use of SWE-2” on the Pro plan is a clear signal of strategy: Cognition wants its own cheaper model to carry everyday work, reducing what it pays to third-party model providers.
The $48 billion valuation
TechCrunch reports that the $2 billion round was led by Andreessen Horowitz, Accel, Founders Fund, General Catalyst and Avenir, four months after a raise at $26 billion. Cognition said its annualised run-rate revenue grew from $492 million to $900 million over that period, without explaining the calculation. Its enterprise customers include Mercedes-Benz, NASA, Goldman Sachs and Citi.
TechCrunch also notes the cost side. Citing The Information, it reports that Cognition leases an Nvidia cluster costing hundreds of millions of dollars a year and could burn around $800 million in 2026. Training its own models on open-source bases is partly a way to reduce dependence on expensive models from OpenAI and Anthropic. The publication frames the round as a sign investors believe AI coding is not a winner-take-all market, especially after Cursor’s sale to SpaceX earlier in the year.
How to evaluate SWE-2 for your team
- Run it on real tickets. Pick 20 to 30 closed issues from your repositories and compare SWE-2’s pull requests with those from your current tool, judged by your own reviewers.
- Test effort levels. Medium effort is faster and cheaper; check whether high or max is needed for your complex tasks.
- Track cost per merged change, not just benchmark scores.
- Keep frontier models for the hardest work. The Terminal-Bench 4.0 gap suggests routing the most complex tasks elsewhere.
- Review security settings. Cloud agents need scoped repository access and branch protection, as with any coding agent.
For background on the wider market, see our coverage of AI-native IDEs and running Cursor cloud agents in Cloudflare sandboxes.
Our analysis
SWE-2 shows the strategy many AI coding companies are converging on: build on a strong open base model, add targeted reinforcement learning, and compete on cost per task rather than absolute benchmark leadership. On Cognition’s numbers, it gets close to frontier models on most coding tests at a fraction of the price, while still falling well short on the hardest terminal tasks. For developers, the $20 Pro plan with free SWE-2 use makes it cheap to test; for companies, the right approach is to measure merged changes and costs on their own code before shifting workloads.