Independent coding test: Astra 6.2 scores better than 6.4 but burns 45% more tokens
robleclerc · x · 2026-11-18
Responding to a wave of bad reviews of Astra 6.2, robleclerc ran a quick coding test replicating recent Astra work, judged blind by two reviewers on a 10-point scale. The new Astra averaged 7.8 vs 6.4 for the current version, winning 3 of 4 tasks outright. The catch: 6.2 consumed 45% more tokens — better, but pricier to run.
More from coding & agent
- Capy launches agent swarm plans from $1 with 300K VM minutes and 3 free models — g3off_who · 2026-11-18
- n8n's self-hosted agent Nod meets NetBird: a guide to secure remote access with zero open ports — aestheticedwar1 · 2026-11-18
- Modaic launches decision-model training at $0.75/1M tokens with 2.69x vLLM throughput — code_star · 2026-11-18
- ChatGPT's $200 tier gets ultrafast: a 3D drafting-table site built in 4 minutes — gabrielchua · 2026-11-18
- Harness-wrapping interfaces are gunning to be the new app store — where you build your own apps — Rasmic · 2026-11-18
- Cognition launches Devin Bots: one bot that leads a team of agents to ship work — round · 2026-11-18