Asana used GPT-6 Astra in Codex to optimize an agent running on GPT-6.1 Sol. See how cost, speed, and quality need to be measured together.
Direct answer
On October 9, 2026, OpenAI published an Asana case where GPT-6 Astra in Codex helped optimize the flow of a browser agent running with GPT-6.1 Sol. In reported tests, the estimated cost per run dropped 76x, speed increased fivefold, and the average cost reached $0.47. The numbers are case-specific and should not be generalized without reproducing the task, quality, environment and acceptance criteria.
Cost per task is worth more than cost per token
The useful unit is the execution completed with quality. Include calls, replay, tools, failures, oversight, and infrastructure.
Optimization starts with observation
Record each agent step, time, tokens, screenshots, errors and decisions. Without a comparable trait, reducing cost becomes trial and error.
Cheaper model is not enough
An efficient model may require more attempts or produce incorrect actions. Compare success rate and total cost under the same set of tasks.
Browser agents need stable testing
Pin sites, states, permissions, and criteria. Separate external changes from the gain caused by the new flow.
Nexus Reading
Use strong models to discover bottlenecks and efficient models to repeat work, maintaining automatic assessment and review proportionate to risk.
FAQ
Is the 76-fold reduction valid for any agent?
No. It's a result of Asana's flow and testing; each company must measure its load.
Which metric to track first?
Cost per correct task, along with latency, success rate and human intervention.
Do I need to use two models?
Not necessarily, but separating optimization research and recurring execution can improve savings.
Essential guides to delve deeper into the decision
Esta análise editorial foi produzida pela Nexus a partir das fontes oficiais abaixo, consultadas em 10 de outubro de 2026. O texto é original e interpreta implicações práticas para empresas.
- OpenAI — Asana cuts model costs 76x in browser tests with GPT-6.1 Sol: official test case, method, metrics and limits.
- OpenAI Developers — Models: Official documentation to compare capabilities, prices and latency.
Date reported by the main source: October 9, 2026.