Toast 1 Beat Claude Fable 5 on Search, for a Third of the Price

TL;DR

Mixedbread, an embeddings and retrieval startup, launched Toast 1 on August 13: a small, specialized "search agent" you plug into any model as a tool, not a chatbot you talk to directly. Its job is narrow: take a query, break it into subqueries, chase down and check sources, and hand back a curated, cited context blob. On Databricks' OfficeQA Pro V2 enterprise benchmark, wiring Toast 1 into GPT-5.6 Sol got 70% answer correctness at about $1.15 to $1.20 a task, beating Claude Fable 5 running solo at 60% correctness for roughly $4 a task. On Harvey's LAB legal benchmark, adding Toast 1 cut token spend by 3.5x for an identical score. It is not open-weights, but the agent loop that runs it is, under Apache 2.0.

A Model That Only Does One Thing

Most "agentic search" today is a general-purpose model bolting on a web-search tool and hoping for the best: fire off one query, skim the top results, start writing. Toast 1 is built to be the thing that model calls instead. Give it a question and it decomposes it into subqueries, gathers evidence for each, inspects the sources it pulled rather than trusting them blindly, and returns a curated, ranked context package for the calling model to reason over.

Think of the difference between an intern who Ctrl-F's the first search result and pastes it into the reply, versus a research librarian who splits your question into parts, pulls from multiple shelves, cross-checks what disagrees, and hands you a short brief with citations. Toast 1 is trying to be the librarian, and it is priced like one that works for tips: standard queries run about $0.016 to $0.023 each, with a higher-quality "fusion" mode at $0.05 to $0.07.

The Number That Matters: 70% for $1.20

Mixedbread's own framing is "matching or outperforming Claude Opus 5 and GPT-5.6 Sol, up to 10x cheaper and 12x faster." That is a vendor's headline, so look at the actual benchmark instead. On OfficeQA Pro V2, a 90-question enterprise financial-reasoning set built by Databricks, three configurations were tested: GPT-5.6 Sol alone scored 33% correctness; Claude Fable 5 alone scored 60% at around $4 per task; GPT-5.6 Sol with Toast 1 doing the retrieval scored 70% at $1.15 to $1.20 per task.

That is the real story, and it is a better one than the marketing line: a mid-tier model with a cheap, specialized search subagent bolted on outscored the priciest frontier model running solo, for around a quarter of what that frontier model's answer cost.

Burning a Third of the Tokens on Legal Work

The second data point comes from Harvey's LAB benchmark, an open-source suite of long-horizon legal agent tasks. On 33 sampled tasks, a vanilla agent needed 80.6M tokens and 21.7 turns to reach a score of 55. Adding Mixedbread's plain search cut that to 47M tokens (down 42%) at the same score. Swapping in Toast 1 as the retrieval subagent cut it further, to 23M tokens and 11.2 turns, a 3.5x reduction from the vanilla run, again at an identical score of 55.

What You Actually Get

Toast 1 itself is API-only through Mixedbread, priced at $0.30 per million input tokens, $0.036 per million cached input tokens, and $0.72 per million output tokens. No downloadable weights. What is open is the plumbing around it: toast-harness, an Apache 2.0 agent loop on GitHub that wires up the retrieval tools (search, grep, fetch chunks, filter, rank) and tracks token spend precisely, meant to be dropped into any OpenAI-compatible coding agent. That is a deliberate split: Mixedbread sells the search brain, and open-sources the wiring that gets other people's models to use it.

Named a search-flavored product after toast, from a company called Mixedbread. Someone in marketing earned their paycheck that week.

The Caveats

Every number above is Mixedbread's own reporting, even though it is measured against genuine third-party benchmarks (Databricks' OfficeQA, Harvey's LAB, and the academic BrowseComp-Plus retrieval set) rather than an in-house eval invented for the occasion. Still, nobody outside Mixedbread has reproduced these specific runs yet. The 10x-cheaper, 12x-faster framing is also a top-line claim covering multiple benchmarks and configurations, not a single clean number, so treat it as directionally true rather than a fixed multiplier you'll see on every workload.

Key Takeaways

• Mixedbread launched Toast 1 on August 13: a specialized search agent, not a general chatbot, meant to be called as a tool by other models.
• On Databricks' OfficeQA Pro V2, GPT-5.6 Sol plus Toast 1 scored 70% correctness at $1.15 to $1.20 a task, versus Claude Fable 5 alone at 60% for about $4 a task.
• On Harvey's LAB legal benchmark, adding Toast 1 cut token usage 3.5x (80.6M to 23M) for an identical score of 55.
• Standard Toast 1 queries run $0.016 to $0.023; a higher-quality fusion mode runs $0.05 to $0.07.
• Toast 1 is API-only and closed-weight, but its orchestration harness, toast-harness, is open source under Apache 2.0.
• All figures are vendor-reported against genuine third-party benchmarks; independent replication hasn't happened yet.

Sources: Mixedbread, "Introducing Toast 1", Databricks, "Introducing OfficeQA Pro V2", Harvey, "Introducing Harvey's Legal Agent Benchmark", GitHub, mixedbread-ai/toast-harness.