Mixedbread AI Launches Toast 1 to Slash Costs of Agentic Search
The new specialized search agent optimizes the retrieval loop, offering a cheaper and faster alternative to using general-purpose frontier models for evidence gathering.
Mixedbread AI has released Toast 1, a specialized search agent designed to manage the complex retrieval loops required for agentic workflows. The tool reduces the computational overhead of iterative search by decomposing queries into subqueries, gathering evidence, and curating context.
Toast 1 operates as either a standalone model or as a sub-agent supporting larger frontier models. According to Mixedbread AI, this specialization of agentic labor results in considerably cheaper search and improved end-to-end results on realistic tasks. The company reports that Toast 1 is up to 10x cheaper and 12x faster than frontier models when performing search tasks. Standard runs are priced between $0.016 and $0.023 per query, with a median latency of 8 seconds.
Benchmarking Performance
In performance tests, Mixedbread AI claims significant efficiency gains. On the OfficeQA Pro V2 benchmark, a configuration using GPT-5.6 Sol with Toast 1 as a sub-agent achieved 70% correctness at approximately $1.15 per task. This outperformed the previous Pareto frontier set by Claude Fable 5 on Databricks Genie, which reached 60% correctness at a cost of roughly $4 per task.
Further data from the Harvey LAB Law Firm Knowledge benchmark shows that integrating Toast 1 as a sub-agent reduced token usage by 51% compared to using Mixedbread Search alone. When compared to a vanilla agent, Toast 1 used 3.5x fewer tokens while maintaining a consistent task score of 55.
The Shift to Agentic Search
The AI industry is moving toward "agentic" search, where models do not simply retrieve a single set of documents but instead iteratively search, verify, and refine their findings. While powerful, executing this loop using general-purpose frontier models is often prohibitively slow and expensive for production environments. Toast 1 joins a growing trend of specialized retrieval agents that seek to optimize the cost-performance trade-off for enterprise deployment.
Industry Implications
By decoupling evidence gathering from final reasoning, developers can utilize faster, specialized models for the iterative search process and reserve expensive frontier models for the final synthesis. This architecture lowers the barrier for deploying multi-step research agents in sectors like legal and financial services, where high accuracy is mandatory but the token costs of iterative loops can be restrictive.
Pricing and Availability
Mixedbread AI has set the pricing for Toast 1 at $0.30 per million input tokens and $0.72 per million output tokens, with cached input tokens priced at $0.036 per million. While the company's benchmarks suggest a new performance ceiling, the specific frontier models cited in their tests, such as GPT-5.6 Sol and Claude Fable 5, appear to be internal or proprietary designations and have not been widely documented in general industry releases.