The most expensive model in an AI stack is often doing work a cheaper one could do better. That is the wager behind Toast 1, a search agent Mixedbread released this week, which the company positions not as a rival to frontier models but as a component they ought to be delegating to.
The headline claim is that Toast 1 matches or outperforms Claude Opus 5 and GPT-5.6 Sol on search quality at as little as a tenth of the cost and up to twelve times the speed. The reasoning behind it is economic rather than technical: once inference is metered per token, spending frontier-grade compute on the mechanical parts of retrieval becomes difficult to justify on any budget.
Unbundling the Agent Loop
What Toast 1 absorbs is the search loop in its entirety, from decomposing a query into subqueries through gathering evidence, inspecting sources and curating what comes back before the calling model ever sees it. It can run standalone or slot in as a subagent that a generalist model already knows how to invoke. A trace published by the company shows it resolving a cross-sector employment-rate question through 16 tool calls across three rounds in just over five seconds.
The design intent is to protect the frontier model's context window. Retrieval output is bulky and largely disposable, while reasoning is the thing the expensive model is genuinely good at.
What the Numbers Show
Two results carry the case. On Databricks' OfficeQA Pro V2, a 90-question test of enterprise financial scenarios, GPT-5.6 Sol running in Codex with Toast 1 attached reached 70% correctness at about $1.15 per task, against 60% at roughly $4 for Claude Fable 5 on Databricks Genie and just 33% for the same Codex configuration with no search subagent. Better answers at roughly a quarter of the cost is an unusual pairing.
The legal benchmark tells a different story that may matter more. Across 33 tasks drawn from Harvey LAB's Law Firm Knowledge suite, three configurations landed on an identical score of 55 while consuming wildly different resources: 80.6 million tokens for a vanilla agent, 47 million with Mixedbread Search, and 23 million once Toast 1 was layered on top, with turns per task falling from 21.7 to 11.2. Nothing improved except efficiency, which is exactly what makes the run informative, since it isolates retrieval quality as the single variable driving cost.
Pricing the Specialist
Standalone, Toast 1 runs roughly $0.016 to $0.023 per query at a median of eight seconds, with a higher-quality fusion mode at $0.05 to $0.07 and around eleven seconds. Mixedbread reports that frontier-model retrieval agents reaching comparable quality in its own testing needed anywhere from 20 seconds to four minutes. Across deep-search evaluations spanning BrowseComp Plus, OfficeQA Pro and LongSeal, the company places it alongside GPT-5.6 Sol and ahead of Kimi K3 and GLM-5.2.
Access is immediate through the Mixedbread API, with launch pricing that includes $0.036 per million cached input tokens and free cache writes. Though co-designed with Mixedbread Search, the agent is built to run over whatever index a team already operates, removing the migration cost that tends to kill retrieval experiments before they start.
A Category, Not a Product
Toast 1 arrives beside SID-1 and Chroma's Context-1, three independent bets on the same thesis: that specialised agents will peel discrete jobs away from general-purpose models as inference budgets tighten. If that holds, the interesting question for buyers stops being which frontier model to standardise on and becomes which parts of the loop to stop paying frontier prices for.
Every figure above is vendor-reported, and the legal comparison rested on a deliberately small 33-task subset chosen to keep repeated runs tractable. Independent replication will decide how much of the claimed frontier survives real workloads.
