Mixedbread's Toast 1 Claims Frontier Retrieval Quality at 10x Cheaper and 12x Faster Than Opus 5 and GPT-5.6 Sol — Priced at $0.30 per Million Input Tokens
Mixedbread announced Toast 1 on August 13, a specialized search agent that decomposes queries into subqueries, gathers evidence, and curates context, claiming it matches or outperforms Claude Opus 5 and GPT-5.6 Sol on search quality at up to 10x lower cost and 12x lower latency. Published numbers: 70% answer correctness on OfficeQA Pro V2 at ~$1.15 per task, 3.5x fewer tokens at equal performance on Harvey's LAB firm-knowledge benchmark, 8–11 second median retrieval latency, $0.016–$0.023 per standard query, and list pricing of $0.30/M input, $0.036/M cached input, $0.72/M output. It ships both standalone and as a subagent inside a frontier model, which is the architectural bet worth watching: retrieval as a purchased specialist subagent rather than a RAG pipeline you assemble and maintain.
↳ Follow the thread