Of ten agent-and-model entries with published costs, only two are not beaten on both score and price by another entry: Claude Code + Opus 5.5 (highest score)…
The leaderboard, plus cost per resolved task, tbench.ai runs each agent-and-model pair for 330 trials on 66 tasks and publishes total cost. The less obvious dominated entry is Claude Code + Sonnet 5.5. The list price per token is not the price per task.
Key takeaways
- Anthropic’s Sonnet 5.5 announcement has Sonnet 5.5 at 70.6% and Opus 5.5 at 66.4%; on tbench.ai it is Opus 5.5 at 64.8% and Sonnet 5.5 at 61.8%.
- On its published figures, Opus 5.5 scores 54.4%, GPT-6 Astra 53.3%, Sonnet 5.5 at xhigh effort 52.1%, and GPT-6 Sol 49.3%.
- Cost: token price, task price, plan price Three different prices get mixed together in most comparisons.
Cost per resolved task is cost per task divided by resolution rate, rounded to the cent.
Read the full article: https://www.trendix.tech/ai-coding-assistants-2/
Top comments (0)