Core Announcement: SWE-2 Released
Cognition has released SWE-2, a coding-agent model built on Kimi K3, which has 2.8T parameters, and optimized through reinforcement learning (RL) post-training.
Facts disclosed in the source:
- Model: SWE-2
- Base model: Kimi K3 (2.8T parameters)
- Optimization method: RL post-training
- Benchmark: FrontierCode 1.1 Main
- Weight availability, API access, and rollout timing: Not specified
Performance: Near Fable 5.1
SWE-2 scored 50.0% on FrontierCode 1.1 Main, a benchmark maintained by Cognition. Fable 5.1 scored 50.9%, leaving a difference of 0.9 percentage points.
SWE-2 also scored 2.5 percentage points higher than GPT-5.6 Sol, which posted 47.5%. Within the comparison provided in the source, SWE-2 shows a competitive result.
Benchmark Context
FrontierCode is Cognition’s upgraded version of SWE-bench. Benchmarks of this kind are intended to assess how well models handle software-engineering tasks, such as interpreting issue descriptions, changing code, and passing tests. Benchmark results are useful reference points, but production performance can still vary with the codebase, toolchain, task mix, and human-review process.
Cost Comparison: Lower Price Than Fable 5.1
According to the source, SWE-2 costs 64% less than Fable 5.1. Given the relatively narrow score difference, that pricing gap may make SWE-2 worth evaluating for users focused on inference costs in coding-agent workflows.
| Model | FrontierCode 1.1 Main | Price vs. Fable 5.1 | Notes |
|---|---|---|---|
| SWE-2 | 50.0% | 64% lower | Kimi K3 with RL post-training |
| Fable 5.1 | 50.9% | Baseline | — |
| GPT-5.6 Sol | 47.5% | — | — |
Note: Scores and pricing information in the table come from the source. Information not disclosed is marked with an em dash.
Use Cases and Deployment Considerations
- Teams that may want to evaluate SWE-2: Development organizations with substantial code-fixing, code-review assistance, or engineering automation workloads that are also sensitive to model costs.
- Questions to resolve before production use: Teams should verify access options, pricing terms, data-handling policies, and performance on their own repositories and task distributions.
- What to measure in an evaluation: Beyond benchmark scores, teams should consider task completion rates, tool-use reliability, patch maintainability, and the time required for human review.
Final Thoughts
SWE-2’s result suggests that post-training strategy can be an important part of a coding agent’s competitiveness alongside the choice of base model. For teams evaluating such systems, performance, cost, and fit with real engineering workflows should be assessed together.