Mixedbread Releases Toast 1 Search Agent to Lower Costs for Frontier AI Workflows

Artificial intelligence startup Mixedbread announced the release of Toast 1, a specialized search agent built to manage complex document retrieval and context curation at a fraction of the operating cost of generalist models. First reported on Hacker News, the model is engineered to handle context-gathering tasks independently or function as a subagent alongside primary reasoning systems. According to Mixedbread, Toast 1 matches or outperforms established frontier models such as Claude Opus 5 and GPT-5.6 Sol, while operating at up to 12 times faster speeds and up to 10 times lower costs.

The product launch targets a growing operational bottleneck in enterprise AI, where top-tier foundation models expend expensive compute capacity and context windows on routine search and document retrieval tasks. Toast 1 offloads the retrieval loop entirely: upon receiving an initial prompt, the agent breaks the user query down into structured subqueries, conducts iterative source inspections, gathers evidentiary material, and returns a condensed context package to the main model. By managing the evidence pipeline, the system allows generalist models to conserve context limits and compute resources strictly for reasoning, decision-making, and final output creation.

In performance evaluations conducted on Databricks' OfficeQA Pro V2 test—a benchmark designed to measure accuracy across 90 complex financial situations—Toast 1 demonstrated marked efficiency gains. Operating as a subagent within Codex alongside GPT-5.6 Sol, the system achieved a 70 percent answer accuracy score at an average cost of $1.15 per completed task. This performance established a new state-of-the-art result for both quality and cost efficiency on the benchmark. In comparison, the previous top performer, Claude Fable 5 running on Databricks Genie, scored 60 percent accuracy at approximately $4.00 per task, while GPT-5.6 Sol without Toast 1 recorded an accuracy rate of 33 percent.

The agent showed similar efficiency improvements on Harvey LAB's Law Firm Knowledge benchmark, which evaluates an AI system's ability to navigate and extract accurate information from large legal repositories. Testing across a random sample of 33 legal tasks revealed that while overall answer accuracy remained constant across retrieval setups, token usage dropped significantly. Replacing standard filesystem search with Mixedbread Search reduced token consumption from 80.6 million to 47 million tokens. Deploying Toast 1 as a dedicated subagent further dropped token consumption to 23 million tokens while requiring half as many interaction turns, resulting in a 3.5-fold overall reduction in tokens and a cost savings exceeding 60 percent.

In addition to functioning as a subagent inside broader workflows, Toast 1 can run as an independent model designed specifically for deep search operations. Mixedbread reported that in standalone evaluations, Toast 1 performed on par with GPT-5.6 Sol while outperforming alternative models such as Kimi K3 and GLM-5.2. Standard queries using Toast 1 cost between $0.016 and $0.023 per request with a median latency of eight seconds. A higher-performing fusion configuration costs between $0.05 and $0.07 per query with an eleven-second median latency, compared to generalist frontier models that required between 20 seconds and four minutes to process identical evaluations at costs seven to 11 times higher.

Mixedbread has made Toast 1 available immediately through its Chat Completions API under initial launch pricing, alongside integration tools for developer environments such as OpenCode and package installations via npx. Although the agent was co-designed to maximize performance when paired with Mixedbread Search primitives, the company emphasized that Toast 1 is backend-agnostic and capable of indexing over existing document infrastructure without requiring backend database migrations. The release places Mixedbread alongside a growing movement of specialized search models, joining recent industry projects like SID-1 and Chroma's Context-1 that aim to separate evidence retrieval from core generative reasoning.