Prescience evals: GPT-6.2-Astra cuts CUA agent task cost ~4.9× in healthcare A/B tests
RishabJainK · x · 2026-11-22
Prescience ran internal computer-use agent (CUA) evals on real patient healthcare tasks with A/B testing and found GPT-6.2-Astra significantly more efficient on video-recorded tasks: calls dropped 68→40 (1.71×), context per call fell 51k→29k tokens (1.76×), and tokens are cheaper (1.63×) — roughly 4.9× total cost savings. The author admits the model is disappointing in other ways, but the team is "AGI pilled" and builds on CUA while waiting for costs to drop, leaning on YC's $2M OpenAI credits.
More from coding & agent
- Garmin Health MCP Server lets you query VO2 max, sleep and training data via natural language — modelcontextprotocol · 2026-11-22
- Stripe MCP connector offers read-only access to Stripe data via API key — modelcontextprotocol · 2026-11-22
- Dev ports Inkling 2 Mini to vLLM, hitting ~200 tok/s decode on 2x RTX 3090 vs 50 on llama.cpp — peaster_ · 2026-11-22
- Six OAuth pitfalls from wiring Claude and ChatGPT into a remote MCP server — Dheeth · 2026-11-22
- Open-source Paperwire MCP lets AI agents send postal mail, certified letters, checks and faxes — QuixiAI · 2026-11-22
- After a year of agentic AI coding, verification—not writing code—became the bottleneck — Positive_Picture-1 · 2026-11-22