Inception's Mercury 3 diffusion model claims to beat GPT-6 Luna and Claude Haiku 5.5 at a fraction of the wait

Inception has released Mercury 3, a language model built on diffusion rather than the left-to-right decoding used by conventional chat models. In its announcement, the startup calls Mercury 3 the most capable diffusion language model released so far, and says it runs four to eight times faster than GPT-6 Luna and Claude Haiku 5.5 while winning eight of the 12 benchmarks in its own comparison table. All three models were evaluated at medium settings in that table, and no outside party has yet reproduced the results.
The core architectural difference is what produces the speed. Standard models such as GPT or Claude emit one token at a time, and every token must wait on the one before it. Diffusion models, familiar from image generation, start with a rough sketch of the entire answer and refine it over several passes, committing many words at once. Inception says this approach yields 1,068 tokens per second on Nvidia GPUs, against roughly 240 for Claude Haiku 5.5 and about 130 for GPT-6 Luna in the company's own comparison. The model accepts up to 1 million tokens of context in a single pass and exposes an OpenAI-compatible API.
The benchmark table is more mixed than the headline framing. Mercury 3's widest margin is on IFBench, an instruction-following test, where it scores 76.84 points, 15 more than GPT-6 Luna. It also leads clearly on SWE-Bench Atlas, an autonomous coding benchmark, at 51.08 points while the two rivals reach 33.33 and 32.26. Claude Haiku 5.5 takes the longer coding test DeepSWE and is stronger at locating information inside long documents. GPT-6 Luna edges ahead on two tests from Artificial Analysis. Inception's own short overview is more measured than its press framing, describing Mercury 3's quality as on par with or better than the competition rather than uniformly superior.
Scaling is part of the argument. Inception says Mercury 3 has more than 40 times as many parameters as Mercury 1 while running just as fast, and that quality has roughly tripled since the small Mercury 1 mini. The company reads this as evidence that diffusion models improve predictably with scale, the same way standard language models do. The exact parameter count is not disclosed, which makes the scaling claim difficult to check independently. This is my own reading, but a 40x parameter increase at constant latency is the kind of claim that usually depends heavily on hardware, batch size and quantization settings that a benchmark table alone cannot reveal.
Inception is positioning the model for workloads where latency compounds: AI agents that take many sequential steps, search, customer support and tutoring. Early-access partners quoted in the announcement mostly reinforce the speed story. Octen AI builds search for smart glasses; its founder Kuan Zou says the search itself took about 80 milliseconds and the language model was the bottleneck, and that with Mercury 3 the median response arrived after 0.9 seconds with the full process finishing in 2.2 seconds. Education company General Learning says it now returns hints and corrections in about one second. Customer service provider Lorikeet reports faster responses alongside better results. TinyFish AI co-founder Shuhao Zhang says Mercury 3 solved 25 of 50 navigation tasks on live websites, beating GPT-6.1-Sol, though no figure is given for the OpenAI model.
Pricing is $0.20 per million input tokens and $0.70 per million output tokens, halved at launch with no stated end date for the discount. Once it lapses, output will cost more than GPT-6 Luna and Claude Haiku 5.5, which both charge $0.50 for shorter inputs; for very long prompts, Inception says, both rivals charge considerably more, and the 1 million token window is where Mercury 3's economics look strongest. The model is available through Inception's own platform as well as OpenRouter, Baseten and Vercel, and because it uses the same API shape as OpenAI's models, Inception says switching requires little effort.
Inception's track record gives the release some context. In February the company shipped Mercury 2, which it described as the first diffusion model to reason through intermediate steps before answering, and it previously raised $50 million from backers including NVentures, M12 and Snowflake Ventures, the venture arms of Nvidia, Microsoft and Snowflake. Google is pursuing the same direction: after a Gemini Diffusion demo, it released the open model DiffusionGemma in June, which Google says runs up to four times faster than standard Gemma 4 with slightly weaker results. That comparison is a useful calibration, since it shows the speed-versus-quality trade-off diffusion has historically carried. Inception's claim is essentially that Mercury 3 has closed that quality gap at the small-model tier; the unverified benchmark table is the only support for it so far.
A likely trade-off for adopters is that diffusion output can behave differently from autoregressive text in ways benchmarks do not capture, such as how reliably it follows strict output formats or handles very long chains of tool calls. Teams should treat the published numbers as a starting hypothesis and run their own evaluations on the tasks they actually ship.
Why it matters: developers building agents, voice interfaces and support tools face latency that accumulates across many model calls, and a model that decodes many tokens per pass could change what feels interactive. Inception's pricing undercuts rivals only during the launch discount and only for long inputs, so cost modeling matters. Independent benchmarks are the missing piece, and until they appear, teams with latency-sensitive products have reason to test Mercury 3 directly rather than plan around the announcement.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Inception says Mercury 3 is the most capable diffusion language model released so far. The company claims it runs four to eight times faster than GPT-6 Luna and Claude Haiku 5.5 and beats both in eight of 12 benchmarks. There are no independent tests yet. The article Inception's Mercury 3 diffusion model claims to beat GPT-6 Luna and Claude Haiku 5.5 at a fraction of the wait appeared first on The Decoder .