Glonce

Mixedbread ships Toast 1, a search subagent 10x cheaper than frontier models

agentsThe Glonce editorsHacker News🔥70Glonce rating70Impact65Novelty55Relevance100Surprise35Players65Credibility78
Mixedbread ships Toast 1, a search subagent 10x cheaper than frontier models

Mixedbread released Toast 1, its first specialised search agent, available immediately through the Mixedbread API. The company says it provides frontier search quality, matching or outperforming Claude Opus 5 and GPT-5.6 Sol, while being up to 10 times cheaper and 12 times faster. It works best paired with Mixedbread's own search product but can run over any retrieval backend. Toast 1 can act as a standalone retrieval agent or as a subagent that a larger frontier model calls: given a query, it breaks it into subqueries, gathers evidence, inspects sources, and hands back curated context, so the frontier model spends its own compute on reasoning and producing the final answer rather than on search itself.

Mixedbread backs the launch with two third-party benchmarks. On Databricks' OfficeQA Pro V2, a set of 90 questions on complex enterprise financial situations, GPT-5.6 Sol running Toast 1 as a subagent inside Codex reached 70% answer correctness at about $1.15 per task, which Mixedbread describes as the top score and the best cost-quality combination in Databricks' evaluation. By comparison, the prior best performer, Claude Fable 5 on Databricks Genie, reached 60% correctness at about $4 per task, and GPT-5.6 Sol running inside Codex without Toast 1 reached only 33% correctness.

On Harvey LAB's Law Firm Knowledge benchmark, evaluated on a randomly selected subset of 33 tasks, Mixedbread reports that GPT-5.6 Sol's answer quality held steady across different search setups while token usage fell sharply: from 80.6 million tokens with the agent's own filesystem search, to 47 million after swapping in Mixedbread Search at an identical task score, to 23 million after adding Toast 1 as a dedicated search subagent, which also let the agent finish in half the turns of the vanilla setup. Mixedbread frames this as answer quality preserved at 3.5 times fewer tokens and over 60% lower cost.

As a standalone model, Mixedbread says Toast 1 reaches frontier-level performance on deep search benchmarks, in the same league as GPT-5.6 Sol and clearly ahead of Kimi K3 or GLM-5.2. A standard run costs about 0.023 per query with an eight-second median latency; a higher-quality fusion configuration costs about 0.07 per query with an eleven-second median latency. Mixedbread says comparable frontier-model retrieval agents in its evaluation took between 20 seconds and four minutes, making Toast 1 roughly 7 to 11 times cheaper among systems reaching similar performance.

Launch pricing through the Mixedbread API is $0.30 per million input tokens, $0.036 per million cached input tokens (cache writes are free), and $0.72 per million output tokens; Mixedbread Search invoked by Toast 1 is billed at a separate special rate. Developers can wire it in through Mixedbread's Chat Completions API as a retrieval tool, add it to coding agents with the command 'npx skills add mixedbread-ai/skills', or use it directly as a subagent through Mixedbread's OpenCode integration. A new API key comes with $5 in included credits.

Key facts

  • Toast 1 is Mixedbread's first specialised search agent, launched today with immediate API access, claimed to match or beat Claude Opus 5 and GPT-5.6 Sol on search quality at up to 10x lower cost and 12x higher speed.
  • On Databricks' 90-question OfficeQA Pro V2 benchmark, GPT-5.6 Sol with Toast 1 as a subagent inside Codex scored 70% correctness at about $1.15 per task, versus 60% at about $4 per task for the prior best (Claude Fable 5 on Databricks Genie) and 33% for GPT-5.6 Sol in Codex without Toast 1.
  • On a 33-task subset of Harvey LAB's Law Firm Knowledge benchmark, adding Toast 1 cut token usage from 80.6 million (vanilla filesystem search) to 23 million while holding answer quality constant, and halved the number of agent turns needed.
  • Launch pricing is $0.30 per million input tokens, $0.036 per million cached input tokens, and $0.72 per million output tokens; a standard query costs about 0.023 and a higher-quality fusion mode about 0.07, versus 20 seconds to four minutes and higher cost for comparable frontier retrieval agents.
  • Toast 1 is backend agnostic, so it can run over an existing retrieval index without a migration, though Mixedbread says it performs best paired with its own Mixedbread Search product; new API keys include $5 in free credits.

Why it matters

Frontier models have become capable of real knowledge work but are also the most expensive part of the stack, and search-heavy tasks burn a large share of that budget on gathering and filtering evidence rather than on reasoning. Toast 1 is Mixedbread's bet that a small model trained specifically for search, used as a subagent, can do that evidence-gathering step at a fraction of the cost of a frontier model while leaving the frontier model free to spend its context and compute on the reasoning and the final answer. Mixedbread frames the result as a new Pareto frontier on cost per task and speed per task across agentic search workloads.

Who it affects

The two benchmarks Mixedbread cites target enterprise use cases directly: Databricks' OfficeQA Pro V2 evaluates financial question answering over complex enterprise documents, and Harvey LAB's Law Firm Knowledge benchmark evaluates search over institutional legal knowledge at scale. That points Toast 1 at teams building agentic workflows in finance, legal, and other document-heavy domains, as well as any developer already orchestrating frontier models with subagents who wants a cheaper, faster search component to plug in.

How to use it

Toast 1 is live now through the Mixedbread API at launch pricing of $0.30 per million input tokens, $0.036 per million cached input tokens (cache writes are free), and $0.72 per million output tokens, with Mixedbread Search itself billed separately at what Mixedbread calls a special rate. It can be added as a retrieval tool through Mixedbread's Chat Completions API, wired into coding agents with 'npx skills add mixedbread-ai/skills', or used directly as a subagent via Mixedbread's OpenCode integration. It runs best against Mixedbread Search but does not require migrating an existing retrieval backend. A new API key comes with $5 in included credits, and Mixedbread built a public demo that searches Dwarkesh's podcast transcripts.

How solid is it

Every number in this story comes from Mixedbread's own launch post, including the headline comparisons against Claude Opus 5, GPT-5.6 Sol, Claude Fable 5, Kimi K3, and GLM-5.2. The two supporting benchmarks, Databricks' OfficeQA Pro V2 and Harvey LAB's Law Firm Knowledge, are run by outside organisations, which lends some independence to the underlying test sets, but the specific evaluation runs, comparisons, and framing described here are Mixedbread's own reporting on those benchmarks, not an independent audit of Toast 1. The legal benchmark result is drawn from a subset of just 33 tasks that Mixedbread selected for tractability, and no date is given for when either evaluation was run.

Risks and caveats

The post does not name an author, gives no explicit release date beyond 'available today', and does not describe Toast 1's architecture, model size, or training data. The per-query cost figures of 0.023 and 0.07 are stated without a currency symbol in the source, though the rest of the page prices everything in dollars. Mixedbread does not explain what distinguishes its 'highest-quality fusion configuration' from the standard Toast 1 run beyond the cost and latency difference, and it gives no separate price for Mixedbread Search when invoked by Toast 1. As with any vendor benchmark, the comparisons favour the vendor's own product and have not been independently reproduced.