Omni cuts internal token spend 80% after switching workhorse model to DeepSeek Flash
josh_wills · x · 2026-10-24
Data company Omni shared two internal model-swap experiments:
- First they tried the cheaper but weaker gpt-5.6-luna as the workhorse model. Daily spend dropped from $450 to $200, but a large share of users manually backed off to stronger models, so the savings underdelivered — "Luna wasn't smart enough."
- Two weeks later they ran a similar test with DeepSeek V4.1 Flash: back-off rates fell to under 10% of queries, total daily spend across all models dropped below $100 (less than half on Flash), an 80% reduction.
The author takes this as evidence that "intelligence is getting free."
More from Venture
- Solo founder raises $20M for Ando, says most competitors aren't real — julianweisser · 2026-10-24
- Ikigai Grants launches $50K Cohort 1 backing underfunded science with $5K checks plus mentorship — yawnxyz · 2026-10-24
- 42 decacorns created in 12 months vs ~20 per decade for 50 years — alvelda · 2026-10-24
- Damodaran: Anthropic IPO will be priced on ARR; $1.5-2T needs 30%+ margins — rohanpaul_ai · 2026-10-24
- Microsoft chief scientist Eric Horvitz: AI bubble may mirror 1847 railway boom, infrastructure wins — erichorvitz · 2026-10-24
- VC Rob Leclerc defends Big Tech mega-raises: 5-10% dilution buys far more growth — robleclerc · 2026-10-24