Mixedbread Releases Toast 1 Specialized Search Agent
Mixedbread has launched Toast 1, a specialized search agent designed to match frontier model retrieval quality while reducing costs and latency.
Toast 1 manages the full search loop, from query decomposition to context curation, for high-efficiency retrieval.
- Mixedbread announced Toast 1, a specialized agent dedicated to managing complex search and retrieval loops.
- The model matches or exceeds the performance of frontier models like GPT-5.6 Sol and Claude Opus 5 in retrieval tasks while being up to 10 cheaper and 12 faster.
- Benchmarks show significant token efficiency gains, including a 3.5 reduction in token consumption on legal industry benchmarks.
Specialized Search for Knowledge-Intensive Tasks
Mixedbread has announced the release of Toast 1, its first specialized search agent designed for high-density knowledge work. The agent is built to handle the entire search loop, including query decomposition, evidence gathering, source inspection, and context curation. According to the company, Toast 1 is designed to operate either as a standalone retrieval agent or as a subagent that supports larger frontier models.
As general-purpose frontier models grow in size and cost, there is an increasing demand for specialized tools that can perform specific reasoning tasks more efficiently. Mixedbread positions Toast 1 as a solution to this problem, claiming it matches the search quality of models like Claude Opus 5 and GPT-5.6 Sol while operating with significantly lower overhead. The model is optimized for the Mixedbread Search infrastructure but maintains compatibility with various search backends.
Benchmark Performance: OfficeQA Pro V2
The efficiency and accuracy of Toast 1 were evaluated using several industry benchmarks, including the OfficeQA Pro V2 by Databricks. This benchmark evaluates the correctness of answers across 90 complex enterprise financial questions. The results indicated that Toast 1 establishes a new Pareto frontier for cost and quality.
When GPT-5.6 Sol was paired with Toast 1 as a subagent within the Codex environment, it achieved an answer correctness score of 70% at a cost of approximately $1.15 per task. This performance surpassed previous benchmarks reported by Databricks, where Claude Fable 5 on Databricks Genie reached 60% correctness at a significantly higher cost of roughly $4 per task. Without the Toast 1 subagent, GPT-5.6 Sol within Codex reached only 33% correctness, suggesting that the specialized search agent significantly enhances the reasoning capabilities of generalist models by providing cleaner, more relevant context.
Efficiency in Legal and Institutional Knowledge
In addition to financial analysis, Toast 1 was tested on the Harvey LAB Law Firm Knowledge benchmark. This evaluation focuses on the ability of an agent to navigate large scales of institutional legal data, where noise and context density often lead to high token consumption and false positives.
The benchmark tests showed that using Toast 1 as a dedicated subagent reduced the total tokens required for a 33-task set from 80.6 million to 23 million. This represents a 3.5 reduction in token usage while maintaining an identical task score of 55. The number of turns required to complete a task also dropped from 21.7 to 11.2. By offloading the retrieval and evidence-gathering process to Toast 1, the primary reasoning model can preserve its context window and reduce overall operational costs by over 60%.
Standalone Retrieval and Latency
Beyond its role as a subagent, Toast 1 is capable of functioning as a standalone model for deep search. Mixedbread reports that in standalone evaluations using benchmarks like BrowseComp Plus, OfficeQA Pro, and LongSeal, the model matches the quality of frontier-class sweeps. A key differentiator is the speed of response; Toast 1 typically returns answers in 8 to 10 seconds, which is faster than the response times of the larger frontier models it was compared against.
The development of Toast 1 follows Mixedbread's co-design philosophy, where the model, agent harness, and retrieval primitives—such as embedding models and the Silo infrastructure—are engineered to function together. This integrated approach aims to minimize the "agentic labor" required to find and verify information across massive document collections.
Integration and Availability
Toast 1 is available for use with existing retrieval stacks and can be integrated into coding agents or Mixedbread Stores. Its ability to decompose a single query into 16 or more tool calls across multiple rounds allows it to answer multi-faceted questions, such as comparative economic analyses, with high precision. For example, a query regarding employment rate changes in retail versus healthcare can be processed in just over five seconds, with the agent generating specific sub-queries and grep patterns to isolate the necessary data points before returning a curated context to the user or a primary model.
Enjoyed this?
Get more posts like this delivered to your inbox.
Perplexity Launches Python Search SDK for Agentic Workflows
Next →Google Announces Gemini 3.7 Flash for Coding and Agents
