SWE-1.7 Reach Near GPT 5.5 and Opus Intelligence (cognition.com)

Article summary

Cognition has launched SWE-1.7, a model that reaches frontier-level intelligence at a lower cost, advancing the cost-performance Pareto curve. The model was trained using a combination of improvements across the RL pipeline, including better infrastructure, more stable training, and higher-quality data. SWE-1.7 is available in Devin and is optimized for longer-horizon asynchronous tasks. The model's training involved preserving entropy and stabilizing training, multi-cluster training, and curating high-quality data.

Main themes

• AI model development
• RL training
• Cost-performance optimization
• Model trustworthiness
• Open-source models
• Benchmarking

What commenters say

• The company's claims about its model's capabilities are met with skepticism due to past instances of fraudulent demos.
• Some commenters believe that third-party benchmarks are necessary to verify the model's performance.
• There is a debate about the value of open-source models, with some arguing that they are inferior to proprietary models like those from OpenAI and Anthropic.
• Others argue that open-source models like GLM-5.2 and Minimax M3 are comparable to or even surpass proprietary models in certain tasks.
• The cost of using high-end models like those from OpenAI and Anthropic is a significant factor for some users, who opt for cheaper alternatives like GLM-5.2.
• Some commenters trust established companies like OpenAI and Anthropic more than smaller startups, while others prefer the flexibility and cost-effectiveness of open-source models.
• The performance difference between high-end and open-source models may not be significant enough to justify the cost difference for some users.
• The use of open-source models can be a viable option for hobby projects or smaller tasks, where the performance difference may not be as critical.