Toast 1 claims MTEB parity with OpenAI. The cost math in the pitch is off by 1000x.

Mixedbread AI has announced Toast 1, an embedding model it says matches or exceeds OpenAI's text-embedding-3-large on standard benchmarks while being significantly cheaper to run ^[1]. The announcement drew 173 points and 58 comments on Hacker News ^[2], which is a reasonable proxy for how many teams treat their embedding provider as a settled decision rather than a line item.

The benchmark deltas are small and are presented as reported figures rather than independently verified ones. On MTEB, Toast 1 is said to score 68.4 overall against 67.9 for text-embedding-3-large ^[3], 58.2 on retrieval against 57.3 ^[4], 84.1 on semantic textual similarity against 83.2 ^[5], 75.8 on classification against 75.1 ^[6], and 62.3 on reranking against 61.8 ^[7]. That is a spread of 0.5 to 0.9 points across five categories ^[21][22]. The write-up concedes the improvements are marginal and that the point is to match OpenAI at lower cost ^[23].

So the whole argument sits on the cost column, and the cost column does not survive contact with a calculator. OpenAI charges $0.13 per million tokens for text-embedding-3-large ^[11]. The write-up says a system processing 100 million tokens per month faces the difference between $13,000 per month and potentially $1,000 to $3,000 per month if it self-hosts Toast 1, which ships as open weights ^[10][9]. But 100 million tokens at $0.13 per million is $13 per month ^[12]. The stated figure is 1000 times too high; $13,000 would correspond to 100 billion tokens a month ^[13]. Taken at the volume actually named, the self-hosting recommendation is between 77 and 231 times more expensive than just paying OpenAI ^[14], before anyone accounts for the engineer maintaining it.

That matters beyond one bad multiplication. Embedding inference at merchant scale is cheap enough that it usually is not the thing worth optimising. The case for swapping only opens up in the region where you are embedding tens of billions of tokens a month, or where the constraint is not price at all: latency, data residency, or the ability to run the model on infrastructure you control. Mixedbread has not finalised Toast 1's pricing and has only indicated it will be significantly cheaper ^[8], so the API-to-API comparison that most buyers would actually make cannot be run yet.

The parts of the release that do read like engineering rather than positioning are the operational ones. Toast 1 supports output dimensions of 128, 256, 512, 768, 1024 and 1536, with Matryoshka truncation so a vector can be shortened while retaining most of its semantic information ^[15]. That allows a 1536-dimension main index and a 256-dimension fast approximate index from one model ^[15], and storage is a real cost in a way that inference frequently is not. It claims strong performance across more than 50 languages without separate models ^[16], runs at roughly 1 to 2GB on a single consumer GPU with 8GB or more of VRAM, with CPU inference possible but suited to batch work ^[17], and is compatible with OpenAI-style embedding APIs and LangChain ^[18]. Mixedbread attributes the results to curated training data and transformer architecture modifications ^[20].

Watch for three things: published per-token pricing, so the comparison stops being hypothetical; independent MTEB reproduction, since half a point is inside the range where nobody should trust a vendor's own table; and whether the OpenAI-compatible interface actually holds against the reindexing bill of moving a live corpus, which the write-up does not cost out ^[19].