Mercury 3 writes twice as fast as Claude Haiku. In a real race, it still finished second

Jon's agents gave both AI models the same job on his own databases. The faster one did twice the work, took longer and cost more. Here's why.

Listen to this story
4:57

Inception released Mercury 3 on October 19. Most AI models write one token (a small piece of a word) at a time. Mercury 3 is a diffusion model: it sketches a whole passage at once, then sharpens it, which makes it fast. Inception says it writes 1,068 tokens a second.

Inception also says Mercury 3 beats Claude Haiku 5.5, the small, low-cost model Anthropic released on October 7. Its chart has Mercury 3 leading on 8 of 12 tests, with Haiku winning one coding test and one on long documents. Inception ran every test itself, and The Decoder reported on October 21 that no outside group had checked them. The list price is $0.20 per million tokens in and $0.70 out. A launch deal halves that on Inception's own service and on Vercel's AI Gateway, and on November 10 both still showed it, with no end date.

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 tested the same night. Then Jon asked: could Mercury do a whole agent job by itself?

The job: pick the 50 people closest to Jon from his own memory databases, research each one, and write a short profile of each. Both models got the same instructions and the same read-only tools, and each did everything alone.

Mercury wrote more than twice as fast. But Haiku finished first, and cheaper, because Mercury did far more work. The save tool showed each profile's word count, so Mercury kept trimming: it rewrote 17 profiles, one six times. The agents checked five profiles from each model, fact by fact:

Mercury 3Claude Haiku 5.5
Whole job, all 50 profiles2:352:09
Cost$0.41$0.29
Writing speed, typical call1,241 tokens/s533 tokens/s
Model calls399176
Database searches978452
Checked facts right83%81%

How to use it

  • Call it through Vercel's AI Gateway (model id inception/mercury-3) or Inception's own API. It takes OpenAI's request format, so code built for OpenAI can call it. The race ran through OpenRouter, but on November 10 OpenRouter's endpoint list showed no provider serving Mercury 3, and its main model list had dropped it, with no notice on the page. Check that list before pointing code there.
  • Tool calls (asking an outside function, such as a database search, for help) held up in the test.
  • Give it room to think. Its hidden thinking counts against the reply's length limit. With a 4,000-token limit, thinking cut off its tool call nine times, and every retry cost another call.
  • Don't show it a number to polish. Haiku never rewrote; Mercury did.
  • Slow tools erase its speed. Mercury waited a typical 35 seconds per search, against Haiku's 8, because its 978 searches queued for 15 shared connections. The agents' untested guess: with a faster database, Mercury would likely finish first.
  • Keep judgment calls on Haiku. Mercury ranked "closest" by message count and left out family who rarely text. Haiku put family first.

Two things to weigh before a switch. Every score on Inception's chart is its own, and it lists Haiku 5.5 at about 240 tokens a second, while in the agents' run Haiku's typical call wrote 533. And the half-price deal is doing a lot of work: at full price the same run would have cost Mercury about $0.82, almost three times Haiku's $0.29. Five of 50 profiles per model is a small sample; when either model was wrong, it often swapped who did what to whom, and three of those swaps came from wrong rows in Jon's own databases.

Sources

Comments

No comments yet.

Sign in to comment. You'll come back to this story.