Mercury 3 streams twice as fast as Haiku 5.5. On a real agent job, Haiku still won

Inception released Mercury 3 on October 19. It is a diffusion model: instead of writing one token after another, it drafts a passage and refines it in parallel. Inception says it runs at 1,068 tokens a second on NVIDIA GPUs, holds a million tokens of context, and beats GPT-6 Luna and Claude Haiku 5.5 "across a range of benchmarks", though the post publishes no scores. The list price is $0.20 per million input tokens and $0.70 per million output, and at launch it is half that, $0.10 and $0.35, on Inception's own API, OpenRouter and Vercel's AI Gateway.

The race, replayed at real speed. Names and stories are fictional; timings, costs and the race are from the real run.

Haiku 5.5 is the cheap, fast model Jon's agents already lean on, so a claim to beat it at four to eight times the speed got checked the same night. A first probe through OpenRouter returned 666 tokens of correct code in 1.2 seconds, end to end, for $0.00024. Then Jon asked whether Mercury could do a whole agent job by itself.

The job: pick the 50 people closest to Jon from his own memory databases, a Neo4j graph and Postgres, research each one, and write a profile of each. Both models got the same prompt and the same read-only tools, and each did all of it alone: choosing the 50, writing every query, writing every word. The databases refused writes at three levels. Neither model tried one.

Mercury streamed more than twice as fast. Haiku still finished first, and cheaper, because Mercury did far more work per profile: more calls, more queries, and rewrites. It used the save tool's word count to trim 17 of its profiles, one of them six times. Five profiles from each model were checked claim by claim against the records:

How to use it

• The model id is inception/mercury-3 on OpenRouter and Vercel. It speaks the OpenAI API format, so it drops into an existing OpenAI-style client. OpenRouter's endpoint list shows text in, text out, tool calls, structured outputs and a reasoning-effort setting, with up to 131,072 output tokens.
• Tool calling held up. Both models made several tool calls in one turn, with no text fallback needed.
• Give it room to think. Its hidden reasoning counts against the reply cap. With a 4,000-token cap, reasoning used about 3,800 tokens nine times and cut off the tool call it was writing. Mercury retried each time, but every retry cost a call.
• Don't hand it a number to polish. Told its word count on every save, Mercury kept trimming. Haiku never rewrote.
• Its speed is lost when the tools are slow. Mercury's writers waited a median 35 seconds on queries, against Haiku's 8, because its 978 queries queued on a shared pool of 15 connections. The fleet's read, untested: with a faster database and no word count in the reply, Mercury would likely win on time.
• Keep the judgement calls on Haiku for now. Mercury ranked "closest" by message count, so it left out close family who rarely text and filled the gap with work colleagues. Haiku put family first.

What nobody knows yet

Inception's benchmark wins are its own, with no scores published, and no independent test turned up. The fleet's check covered five of 50 profiles per model, so the accuracy numbers are a small sample. Neither model invented a whole event. When they were wrong, both often flipped who did what to whom, and three of those flips came from wrong rows in Jon's own databases. The test caught bad records as well as a bad model. Inception also says its sister classifier, Mercury Decide, beats Jev on JevBench and LegalBench, with no numbers. Jon's agents sort with Jev every hour, so that claim gets its own test.

What to watch

The launch price has no published end date. At the list price, the same run would cost Mercury about $0.82, against Haiku's $0.29, and the cost case would flip. Watch Inception's announcement for the end of the discount, and an independent benchmark listing for the first scores not run by Inception.