Five models refreshed with the browser harness

• Mistral Nemo
• Ling 3.0 Flash
• GPT-OSS 20B
• Llama 3.1 8B Instruct
• Qwen 3.7 Flash

The browser-enabled harness gives models tools to run and inspect their applications, interact with them, and revise their work within a 15-minute session. These fresh results replace the previous outputs for the five models.

This release includes 173 run records: 154 complete outputs with previews and 19 runs that produced no HTML. Two Qwen outputs—Platformer and Tetris—could not complete final capture and are excluded from publication.

Run costs and token counts have been reconciled with OpenRouter’s reported usage for the recorded response IDs. Each run retains its transcript and timing information so you can inspect what the model actually did.