Inception promises GPT-6-level quality at four to eight times the speed – the numbers are unverified
The company behind the commercial diffusion models for language launched Mercury 3 on 19 October, its fourth model since Mercury 1 arrived in June 2025. The claim is ambitious: quality on par with OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5, at four to eight times the speed and half the launch price. But all the numbers behind the claims were measured by Inception itself, and no independent party has yet published a test of the model.
What was launched
In a press release via Business Wire on 19 October 2026, the Redwood City-based company describes Mercury 3 as "the most capable dLLM ever released" (AOL/Business Wire, sourceId 10983eb3). Mercury 3 is a so-called diffusion language model (dLLM) – a model class Inception itself first commercialized, according to the press release. The company claims this approach delivers high generation speed without sacrificing quality; the mechanism behind the speed advantage is thus, for now, a company statement rather than documented fact.
According to the press release, Mercury 3 is more than 40 times larger than Mercury 1, and the context window has been increased from 260,000 to one million tokens. The exact parameter count is not disclosed.
The numbers – and what they rest on
Inception's own measurements, as reported by SiliconANGLE, put the speed at 1,068 tokens per second on Nvidia GPUs, versus roughly 240 for Claude Haiku 5.5 and 130 for GPT-6 Luna in the same comparison (SiliconANGLE, sourceId 2824dfb0).
The results reported there show that Mercury 3 beat GPT-6 Luna and Haiku 5.5 on eight of twelve tests, while The Next Web writes that Inception's own benchmarks show wins over both models on a range of tests, at four to eight times the speed (The Next Web, sourceId 28f148ef). The largest lead came on coding tasks: 51.1 on SWE-Bench Atlas, versus 33.3 for GPT-6 Luna and 32.3 for Haiku 5.5. But the model also loses: on DeepSWE, another coding benchmark, it scored 46.3 against Haiku 5.5's 51.6, and it lost one long-context test.
The Next Web points to the caveat that underpins this entire story: the numbers are self-evaluated, and no outside group has yet published a test of the model (sourceId 28f148ef).
The pricing play against the fast-service tier
The price is set at $0.20 per million input tokens and $0.70 per million output tokens – halved as a launch promotion with no stated end date. Cached tokens receive a 90 percent discount, which according to the Forkast News analysis (via Yahoo Tech) is designed for high-volume, repetitive agent workflows (Yahoo Tech, sourceId b68276e1).
Forkast, which verified the prices against Inception's blog post and the OpenRouter listing, interprets this as a move that undercuts the price floor in the fast, cheap tier where GPT-6 Luna and Haiku 5.5 operate. It is worth noting the uncertainty: with no end date on the promotion, no one knows what the long-term price actually is.
What the model is already meant to be used for
Inception points to latency-sensitive applications. The education platform RevisionDojo is cited as an early partner: "Hints, marked answers and definitions now come back in about a second, at quality as good as or better than the models we ran before," says Michael Tong, CTO of General Learning, aka RevisionDojo, in the press release (sourceId 10983eb3). The quote is a customer statement published by the vendor itself.
Browser agents are another focus area. According to TechRepublic, partner TinyFish AI tested the model on 50 live Mind2Web navigation tasks, of which 50 percent were completed – a result Inception says was better than GPT-6.1-Sol at the same token volume and number of agent steps (TechRepublic, sourceId 28bf5528). This, too, is vendor-reported, not independently established.
The model that, according to the company, improved itself
Perhaps the most striking claim came from Inception CEO Stefano Ermon, who according to The Next Web told The Deep View that the company used swarms of Mercury 3 agents to improve the model itself – a form of recursive self-improvement that reportedly added around 16 percent to the model's speed, on top of the gains the team achieved manually (sourceId 28f148ef). This is a company statement that cannot be verified with the available sources.
Open questions
Several matters remain unresolved:
- No independent testing. The entire quality and speed story rests on Inception's own measurements. Until a third party replicates the numbers, the 1,068 tok/s figure and the 4–8x advantage should be read as marketing, not documentation.
- Unknown model size. The parameter count is being kept secret; the only known fact is that the model is "more than 40 times larger" than Mercury 1.
- Promotional price with no end date. The long-term price is uncertain, and with it whether the price competition against Luna and Haiku 5.5 will last.
- Acquisition rumors. Reuters reported in May that Microsoft was in talks to acquire Inception, which wanted more than one billion dollars, and that SpaceX had also made contact. No deal has been announced, and SiliconANGLE presents the rumor as unconfirmed (sourceId 2824dfb0).
Whether Mercury 3 actually beats its competitors is therefore an open question until independent testing is available. What Inception has delivered so far is a compelling pricing story built on numbers that no one but the company itself has measured.