LegalOn and Asana report sharp GPT-6 cost cuts

LegalOn cut estimated Codex daily costs about 65% by routing GPT-6 models, while Asana cut a browser agent’s model cost 76 times in tests.

LegalOn Technologies said it cut estimated OpenAI Codex daily costs about 65% versus unrestricted GPT-5.5 Fast mode by routing tasks across GPT-6 Luna, GPT-6.1 Sol, and GPT-6 Astra, even though its headline described the reduction as halving costs. Fast mode is off by default, parallel work was kept to preserve speed, and department and individual monthly caps govern spending, with mature businesses asked for roughly 20% greater efficiency and newer businesses allowed larger budgets. Separately, Asana used GPT-6 Astra in Codex to optimize StackAI browser-agent workflows running on GPT-6.1 Sol. A 144-run study produced a workflow averaging $0.47 and about four minutes per run, 76 times cheaper and five times faster than the original Model B setup, using a 480,000-character budget, cached history, more page text, and batched screenshot removal after 20 images. Asana said 89% of optimized input was cached at 5% of the uncached price, every optimized run correctly collected six fields for 32 catalog books, and work estimated at one to two months took about a week before shipping through its Command platform. The reports describe different companies, products, and baselines and do not contradict each other’s measurements.

• On 2026-10-08, LegalOn said estimated Codex daily costs fell about 65% versus unrestricted GPT-5.5 Fast mode, though its headline said costs were halved.
• LegalOn routes routine work to GPT-6 Luna, standard design and analysis to GPT-6.1 Sol, and complex design and agent orchestration to GPT-6 Astra, with Fast mode off by default.
• LegalOn set department and individual monthly usage caps; mature businesses were asked for about 20% greater efficiency while newer businesses kept larger AI budgets.
• On 2026-10-09, Asana said GPT-6 Astra in Codex optimized StackAI browser-agent workflows on GPT-6.1 Sol, finding cached instructions but full page-history resends.
• In a 144-run study, optimized runs averaged $0.47 and about four minutes—76 times cheaper and five times faster than the original Model B setup.
• The winning policy used a 480,000-character budget, cached browsing history, kept more page text, and removed screenshots in batches after 20 images; 89% of optimized input was cached at 5% of the uncached price.
• Every optimized Asana run correctly collected six fields for 32 catalog books; work estimated at one to two months took about a week and shipped through Asana’s Command platform.