Mixedbread launches Toast 1 search agent, claiming 70% on OfficeQA Pro V2 at about $1.15 per task
- Mixedbread says GPT-5.6 Sol using Toast 1 as a Codex sub-agent reached 70% correctness on Databricks' 90-question OfficeQA Pro V2 at about $1.15 per task, versus 60% at about $4 for Claude Fable 5 on Databricks Genie.
- Toast 1 takes over the retrieval loop by decomposing a query, gathering and inspecting sources, then returning curated context so a general-purpose model can focus on reasoning and answer generation.
- On a 33-task subset of Harvey LAB's firm-knowledge benchmark, Mixedbread reports equal task scores of 55 while Toast 1 reduced total usage to 23 million tokens, from 80.6 million for a vanilla agent, or 3.5 times fewer tokens.
- Mixedbread says Toast 1 works with any search backend but performs best with Mixedbread Search, and claims it is up to 10 times cheaper and 12 times faster than Claude Opus 5 and GPT-5.6 Sol for search.
Hacker News opinions
I like specialized LLMs for search. Humans usually need 2 to 5 rounds of queries, links, and assumption checks for anything complex, so an agent that does that quickly seems useful.
I get why instant AI answers are appealing, but crawling Reddit and Stack Overflow myself often gives me a deeper understanding. When I have time, I deliberately avoid the summary.
For ordinary troubleshooting, Gemini search already works well for me. I described a refrigerator water-dispenser problem, got the fix, found a video, and repaired it in under 15 minutes.
Google image search has become much worse for vague, exploratory searches. I used to spend hours recovering childhood memories through it, but now I hit the end of poor results in minutes.
Gemini's default search is weak, but Deep Research gives noticeably better results.
I want to know when I should use a dedicated search agent instead of a smaller general model or a conventional RAG pipeline. The distinction is not clear from this post.
The author says smaller general models trail frontier models on search. Their view is that a good index matters most, followed by a fast specialized model.
The bread-themed naming makes the product look like a joke or spoof at first glance. I have to double-check what the page actually is.
The article should explain Mixedbread Search itself. The author says it is a multimodal, multilingual product backed by the Wholembed v3 late-interaction retrieval model.
Google Search Overviews and Gemini search are still faster than alternatives for me. Gemini pulled court records about a local development in about 20 seconds, while this sounds more like AI Algolia than Google.
I want to know whether the benchmark comparisons hold the harness constant. The author says they used the same retrieval harnesses, Codex for OfficeQA Pro V2 and Harvey's vanilla benchmark for the legal tasks, with minimal tools added for Mixedbread Search and Toast 1.
I use SearXNG MCP and like the idea of a dedicated search model, but I wish Toast 1 had open weights. I would also want comparisons with Perplexity, Gemini with Search, and Parallel AI.