Inception releases Mercury 3, a diffusion model it says writes at more than 1,000 tokens a second
Inception, Monday: Mercury 3 beats GPT-6 Luna and Claude Haiku 5.5 on the lab’s own table, at 1,068 tokens a second on Nvidia GPUs, with a 1 million token window. The launch price is $0.10 per million input tokens and $0.35 output, half the list of $0.20 and $0.70. OpenRouter’s live median on Tuesday was 235 tokens a second. The weights are not released. Stefano Ermon says the next model is already in training.
Deep dive6 minIn-depth
Maya Chen6 min read
- Inception
- Mercury
- Diffusion
- OpenRouter
- Agents

Photo: InceptionSource page
Inception released Mercury 3 on Monday and called it the most capable diffusion language model the Redwood City lab has shipped. On the company’s benchmarks it outperforms OpenAI’s GPT-6 Luna and Anthropic’s Claude Haiku 5.5, and Inception says it runs four to eight times faster, at 1,068 tokens a second on Nvidia GPUs. The context window is 1 million tokens, up from 260,000 on Mercury 2.5. The list price is $0.20 per million input tokens and $0.70 per million output tokens. At launch both are cut in half, to $0.10 and $0.35. Cached input is 90 percent off the input price. Stefano Ermon, a co-founder, said in the Business Wire release that the company has already started training the next model. He did not give a date. Inception has not said how many parameters Mercury 3 has.
The model is on the Inception API, OpenRouter, Baseten and Vercel. It speaks OpenAI’s API format, so a developer can point an existing client at it. Inception’s launch post says new accounts get 100 million free tokens, and that an enterprise deployment can take dedicated capacity, autoscaling, compliance controls and a setting for how long data is kept. The weights are not part of the release. OpenRouter’s page lists the model string as inception/mercury-3.
Many positions at once
The models behind most chat products are autoregressive. They pick one token, append it, and pick the next, each choice conditioned on everything already written. Later tokens cannot start until earlier ones exist, so the wait grows with the length of the answer. A diffusion language model, in Inception’s account and in Jean-marc Mommessin’s reading of the same design at MarkTechPost, works on many positions together. It starts from a rough draft of the whole answer and revises those positions over a small number of passes. Business Wire puts the contrast in one sentence: autoregressive models generate text sequentially, one token at a time, and diffusion models can generate multiple tokens in parallel. The lab’s argument is that fewer sequential steps is why the speed can stay high as the model gets larger. Mercury 3, it says, is more than 40 times the size of Mercury 1, about three times as capable on the lab’s standard benchmarks, and still above 1,000 tokens a second.
That is a description of the method, not a published trace of this model’s server. The launch post does not give the number of refinement steps, the Nvidia chip behind the 1,068, or a speed measured at the full million-token window. Mercury 1 shipped in June 2025. Mercury 2 added reasoning. Mercury 2.5 added a dial for how much reasoning to do, and the 260,000-token window. Mercury 3 is the fourth named release. Ermon’s line in the wire release is that the limits on scaling this design “are compute and data, not fundamental algorithmic barriers.”
Eight of twelve, on the lab’s table
Mommessin read Inception’s benchmark table, run at medium settings, and counted Mercury 3 ahead of both GPT-6 Luna and Claude Haiku 5.5 on 8 of 12 tests. The widest gaps he reports are on instruction following and agentic coding. IFBench is 76.84, against 61.82 for Luna and 54.65 for Haiku 5.5. SWE-Bench Atlas is 51.08, against 33.33 and 32.26. GPQA Diamond is 89.23, against 86.36 and 84.51. The same table has Mercury 3 behind on other tests. Haiku 5.5 leads DeepSWE v1, 51.62 to 46.31, and MRCR v2, 78.40 to 75.50. Luna is narrowly ahead on AA-Omniscience-Public, 46.94 to 46.44, and on AA-LCR v1.1, 78.30 to 77.67. Those figures are Inception’s. MarkTechPost notes that an outside group had not published its own run. The Next Web, on Tuesday, said the same: the lab’s benchmarks are its own.
The speed column on that table, as MarkTechPost prints it, puts Mercury 3 at 1,068 tokens a second, Luna at about 130 and Haiku 5.5 at about 240. MarkTechPost also lines up the published prices. At the launch rate Mercury 3 is $0.10 per million input tokens, the same short-context rate it lists for Luna and for Haiku 5.5 on prompts up to 100,000 tokens, and $0.35 per million output tokens, against $0.50 for each of those two. The list rate is a different comparison. At $0.70 per million output tokens, Mercury 3 would cost more on output than the $0.50 those pages show. Inception has not said how long the half-price launch rate lasts. The Next Web reported on Tuesday that the launch rate undercuts those two models on output tokens, and that Inception has not said when the half-price rate ends.
235 tokens a second, on the live route
OpenRouter’s page for the model, as it stood on Tuesday, lists one provider, Inception, at $0.10 and $0.35, with cache reads at $0.01 per million tokens. The page’s median throughput is 235 tokens a second. The median latency, the wait it records before output, is 1.02 seconds. Over the week shown on the page, median throughput averaged 228 tokens a second, and the slower tail was much slower: the P90 average was 88 tokens a second. End to end, the median averaged 2.73 seconds. OpenRouter dates the listing to 16 October, three days before the public post. The 1,068 and the 235 are not the same experiment. One is a vendor figure for generation speed on Nvidia GPUs. The other is the median of mixed production calls. A team pricing an agent, which waits on a model at every step, should not treat them as the same number.
The same page says the model accepts tools and a JSON schema for structured output, and that a completion can run to 131,072 tokens inside the 1 million token window. On the public app table, the largest line is an evaluation baseline, ifeval, at 11.8 billion tokens. The largest named product is Hermes Agent, Nous Research’s open-source agent, at 1.03 billion tokens. That is traffic OpenRouter is willing to show. It is not a ranking of every customer.
Four customers, and a next model with no name
Inception quotes four early users on the launch post. TinyFish said it ran Mercury 3 inside a production web-agent harness on 50 live Mind2Web navigation tasks, completed 50 percent of them, and did so with the same token volume and the same number of steps as a frontier model, “significantly more” than GPT-6.1-Sol. The post does not give Sol’s score. Kuan Zou of Octen said that in tests of the company’s Sense API, for search on smart glasses, median generation took 0.9 seconds and a median reply took 2.2 seconds end to end. Michael Tong of RevisionDojo said hints and marked answers came back in about a second, and that the swap did not require a rewrite. Michael Gribben of Lorikeet said early tests beat other models on latency and on the company’s own evals for support agents. Those are the customers’ sentences. They are not a third-party benchmark.
The Next Web, citing an interview Ermon gave The Deep View, reported that Inception used groups of Mercury 3 agents to work on the model itself, and that Ermon put the extra speed from that work at about 16 percent, on top of changes the team made by hand. The Next Web also reported that the company would not disclose a parameter count. Business Wire’s list of backers is Menlo Ventures, Mayfield, Innovation Endeavors, M12, Snowflake Ventures, Databricks Ventures, and Andrew Ng, Andrej Karpathy and Eric Schmidt. Two sibling products sit next to the flagship. Mercury Voice is built for the wait before the first token in a voice agent. Mercury Decide is a classifier. Monday’s news is the general model: a diffusion system Inception says can be sold at the speed of its smaller predecessors, at a launch price that undercuts Luna and Haiku 5.5 on output tokens until the discount ends. The figure a developer can check without the lab’s harness is the one on the OpenRouter page.


