Toast 1

Mixedbread launched Toast 1 on 13 August 2026 as a specialised search agent rather than a general chat model. Its design centres on a single job: take a question, decompose it into subqueries, gather evidence from indexed sources, inspect that evidence, and hand back a curated package of relevant material for a larger Announced launch pricing from the company is $0.30 per million input tokens, $0.036 per million cached input tokens, and $0.72 per million output tokens, with reported per-query cost of roughly $0.016–$0.023 at an eight-second median latency and roughly $0.05–$0.07 at an eleven-second median in its highest-quality conf

Mixedbread announced Toast 1 as its first specialised search agent, built to run the full retrieval loop: query decomposition, evidence gathering, source inspection, and context curation. The model is positioned for high-density knowledge work and can act either as a standalone retrieval agent or as a subagent supporti On the OfficeQA Pro V2 benchmark from Databricks — ninety complex enterprise financial questions — Toast 1 is reported to establish a new Pareto frontier for cost and quality. When GPT-5.6 Sol was paired with Toast 1 as a subagent inside the Codex environment, it reached a 70% answer-correctness score at roughly $1.15

Mixedbread's first-party changelog documents several Toast 1 capability additions in September 2026. On 3 September, hosted search tools were added to the Chat Completions and Responses APIs so Toast 1 can run the full search harness server-side against user Stores with the same tool schemas, defaults, and truncation a On 10 September, MCP servers can be declared in the Responses API so Toast 1 can search via external MCP tools, following the OpenAI remote MCP tool shape including call approval and recording every call on the response. On 14 September, structured outputs via OpenAI's Responses and Chat Completions APIs allow Toast 1

Mixedbread's quality evals page evaluates retrieval and agent benchmarks against the strongest competing models using public datasets, standard splits, and evaluation scripts. Retrieval numbers come from the Wholembed V3 model served through the Stores API, while agent numbers come from Toast 1 or third-party agents ru On BrowseComp-Plus — a deep-research evaluation with multi-hop questions over 100k web documents designed to isolate retrieval quality from model capability — the Mixedbread (GET_DOCUMENT) configuration leads with 90.48% accuracy at 11.53 calls, followed by Reason-ModernColBERT (GET_DOCUMENT) at 87.59% with 13.27 calls