Mixedbread's Toast 1 offloads the search loop from frontier models at 10x less cost

Mixedbread has launched Toast 1, a specialized retrieval agent designed to handle the search half of agentic workflows so that expensive frontier models can spend their tokens on reasoning rather than digging through documents. Given a query, Toast 1 breaks it into subqueries, gathers and inspects evidence, and returns a curated context package. It runs standalone or as a subagent, works with any retrieval backend, and is tuned to pair with Mixedbread Search. The company claims it matches or beats top frontier models on deep-search benchmarks while running roughly 7-11x cheaper and up to 12x faster, with an eight-second median latency and per-query costs starting near two cents.

The pitch rests on the economics of evidence gathering rather than raw answer quality. On Databricks’ OfficeQA Pro V2 financial benchmark, adding Toast 1 as a subagent pushed a frontier model to 70% correctness at about $1.15 per task, versus 33% without it and 60% at roughly $4 for the prior best system. On Harvey’s law-firm knowledge benchmark, answer quality held steady while token consumption dropped from 80.6M to 23M and turns were roughly halved, cutting cost more than 60%. The consistent pattern is that better retrieval frees up the frontier model’s context window instead of raising its ceiling.

Toast 1 is available now through the Mixedbread API at launch pricing of $0.30 per million input tokens and $0.72 per million output, with integrations for coding agents, a Chat Completions retrieval-tool path, and $5 in free credits to start. Mixedbread frames it as part of a broader trend of purpose-built search agents alongside efforts like Chroma’s Context-1, all aiming to bring frontier-grade retrieval to production without frontier-grade cost or latency. The practical takeaway for engineering teams: as model intelligence gets metered, splitting off retrieval into a cheap specialist is emerging as a concrete way to control agent spend.