Mercury 3 Claims 1,068 Tokens a Second. OpenRouter Measures 241, and the Half-Price Launch Has No End Date.
Inception launched Mercury 3 on October 19, a text diffusion model it says runs at 1,068 tokens a second on NVIDIA GPUs, four to eight times faster than GPT-6 Luna and Claude Haiku 5.5 while beating them on Inception's own benchmarks, with a 1M context and launch pricing of $0.10 in and $0.35 out per million tokens. OpenRouter's live median for the only provider is about 241 tokens a second. The quality claim is a vendor chart, Artificial Analysis has no Intelligence Index score, the 50% discount has no sunset, and the 40x size claim comes with no parameter count. We read Inception's post, OpenRouter's model page, and AA's listings.
By T.W. Ghost
Inception launched Mercury 3 today, October 19, 2026, and the number in every headline is 1,068: tokens per second, on NVIDIA GPUs, by Inception's measurement. Mercury 3 is a text diffusion model rather than an autoregressive one, which is the architectural reason Inception gives for the speed, and the blog says it runs four to eight times faster than GPT-6 Luna and Claude Haiku 5.5 while "outperforming" both on Inception's benchmarks. It carries a 1M-token context, up from 260K on Mercury 2.5, and a launch price of $0.10 in and $0.35 out per million tokens, half of the $0.20 and $0.70 list.
The useful catch is not that the vendor number is wrong on its own hardware. It is that the only place you can measure the model today is the one provider serving it, Inception, through OpenRouter, and OpenRouter's live median there is about 241 tokens per second. Still fast. Not a thousand. We read Inception's blog and press release, OpenRouter's model page and API, and Artificial Analysis's listings. Here is what held up.
What You Are Calling
| Mercury 3 | |
|---|---|
| Launched | October 19, 2026 (OpenRouter lists October 16 in places) |
| Architecture | Text diffusion LLM; "more than 40x larger than Mercury 1"; no parameter count published |
| Context | 1M tokens (Mercury 2.5: 260K) |
| List price | $0.20 in / $0.70 out per 1M tokens; cached input 90% off |
| Launch price | $0.10 / $0.35, 50% off, no end date stated; OpenRouter bills $0.10 / $0.35 with cache reads at $0.01 |
| Vendor speed | 1,068 tokens/s on NVIDIA GPUs; "4 to 8x faster" than Luna and Haiku |
| OpenRouter, Oct 23 check | P50 about 241 tokens/s; P50 latency about 0.80 s; Inception the only provider |
| Where | Inception API, OpenRouter, Baseten, Vercel; 100M free tokens (vendor) |
| AA Intelligence Index | None; no model page; no changelog entry |
The speed gap is the story. Inception's 1,068 is Inception on its own GPUs. OpenRouter's 241 is the same model, same provider, measured from outside. The scout reading on October 20 saw 194 with a one-week average near 230, so the number drifts upward with load and time; cite the day-of figure and say so.
The quality claim lives in a chart. "Outperforms GPT-6 Luna and Claude Haiku 5.5" is Inception's, with scores in blog imagery, and there is no AA Intelligence Index entry to check it against. Mercury 3 has no /models/mercury-3 page on AA and no changelog line. We print no rank because none exists.
The discount has no clock. $0.10 and $0.35 is the launch rate. The blog gives no end date. Budget at the $0.20 and $0.70 list, because the promo can end without notice and nothing says it will not.
"40x larger" is a size with no size. Inception publishes no parameter count. A circulating estimate of 280B parameters is unverified and stays out.
Where It Fits
A model that answers in 0.8 seconds at a few hundred tokens a second, with a 1M context, at a dime per million input tokens, is interesting for exactly the loops where Luna and Haiku 5.5 compete: search agents, voice back ends, coding subagents, anything where latency is the product. The honest comparison is on your own loop, with OpenRouter's or Inception's measured throughput, not the 1,068. Inception also says its companion Mercury Decide outperforms Jev on decisions; we did not check that claim for this post and it stays out.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Launched Oct 19; diffusion LLM; 1M context; $0.20 / $0.70 list; 50% launch pricing; 90% cache discount | Verified (Inception blog, BusinessWire, OpenRouter) |
| 1,068 tokens/s | Vendor (Inception on NVIDIA GPUs) |
| OpenRouter P50 about 241 tokens/s, latency about 0.80 s | Verified (OpenRouter, Oct 23 check); earlier reads 194 to 230 |
| "Outperforms GPT-6 Luna and Claude Haiku 5.5" | Vendor chart; no independent score |
| AA Intelligence Index score | Unverified (none exists) |
| Promo end date | Unverified (none stated) |
| Parameter count; "40x larger than Mercury 1" | Vendor, no figure; 280B estimate unverified |
| Mercury Decide beats Jev | Vendor; not checked here |
Who Should Use It
Benchmark Mercury 3 on your latency-sensitive loops this week while the 100 million free tokens and the half-price rate last, and use the throughput you measure, not the one in the headline.
Skip it for anything where you need an independent quality number before you commit; there is none, and the vendor chart is the only evidence that it beats Luna or Haiku 5.5. Price every plan at list.
The frontier is elsewhere: Luna and Haiku 5.5 have Artificial Analysis scores (38 and 43 at max, on the current index) and Mercury 3 has a chart. Until AA scores it, the comparison Inception invites cannot be checked. Our model comparison has the current standings side by side.
Sources
- Inception: introducing Mercury 3
- OpenRouter: Mercury 3
- BusinessWire: Inception launches Mercury 3
- Artificial Analysis: Inception provider page
*Not sure whether raw speed or a verified score matters more for your agent loop? Take the free 2-minute quiz and get matched. Then read about the small model Mercury claims to beat, and the surcharge in its price list: Claude Haiku 5.5.*