Developer notices astra completes tasks with far fewer tokens, suspects less thinking
lucasmeijer · x · 2026-09-05
Developer lucasmeijer shares a hands-on observation: astra gets his jobs done with surprisingly few tokens. Comparing the traces, the tool calls look broadly similar, and he guesses the savings come mainly from fewer thinking tokens rather than fewer tool calls.
More from Models
- Robot control idea: local high-frequency model consuming latent predictions from a larger model — chris_j_paxton · 2026-09-05
- Astra beats Sol but not SOTA on hard wet-lab biology, citing scarce public data — nlarusstone · 2026-09-05
- Stanford's Marin 535B-A23B Open Training Run Hits 13%, Funded by Jensen Huang's Foundation — stanfordnlp · 2026-09-05
- User Cancels Fable 5.1 After One Prompt Burns 1% of Weekly Quota — BLUECOW009 · 2026-09-05
- GPT-6 Astra burns 6.73M tokens in 44 minutes for cinematic 3D reconstruction — haider1 · 2026-09-05
- Delip Rao: mathematicians are finding problems with closed-model companies — deliprao · 2026-09-05