Glonce

Inception's Mercury 3 diffusion model claims to beat GPT-6 Luna and Claude Haiku 5.5 on speed

modelsThe Glonce editorsThe Decoder🔥63Glonce rating63Impact50Novelty55Relevance100Surprise45Players30Credibility65
Inception's Mercury 3 diffusion model claims to beat GPT-6 Luna and Claude Haiku 5.5 on speed

AI startup Inception has released Mercury 3, a diffusion-based language model. Inception says it beats GPT-6 Luna and Claude Haiku 5.5 on a range of benchmarks while running four to eight times faster, and in its announcement it calls Mercury 3 the most capable diffusion language model released so far.

The speed comes from the design. Models like GPT or Claude write one word at a time, so each new word waits for the one before it. Diffusion models, best known from image generation, first sketch a rough version of the whole answer and then refine it over several passes, generating many words at once.

The Decoder notes that the lead is narrower than the announcement suggests. In Inception's own benchmark table, Mercury 3 wins eight of 12 tests, with all three models running at medium settings. The gap is widest in instruction following: on IFBench, Mercury 3 scores 76.84 points, 15 more than GPT-6 Luna. It also leads clearly in autonomous coding on SWE-Bench Atlas, with 51.08 points against 33.33 and 32.26 for its rivals. Claude Haiku 5.5 wins the longer coding test DeepSWE and is better at finding information in long texts. GPT-6 Luna edges ahead in two tests from Artificial Analysis. Inception's short overview is more cautious than its table, describing Mercury 3's quality as on par or better.

On speed, Inception says Mercury 3 generates 1,068 tokens per second on Nvidia GPUs. In the company's comparison, Claude Haiku 5.5 manages about 240 and GPT-6 Luna about 130. No independent tests exist yet.

Inception also says Mercury 3 has more than 40 times as many parameters as Mercury 1 but runs just as fast, and that performance has tripled since the small Mercury 1 mini. It reads this as a sign that diffusion models reliably improve as they scale, just like standard language models. The exact parameter count is undisclosed.

Inception pitches the model mainly for AI agents, search, customer support and tutoring, where small delays add up; agents complete tasks in many separate steps and wait on the model at each one. Early-access partners report results in the announcement, mostly about speed. Octen AI builds search features for smart glasses. Founder Kuan Zou says "On smart glasses, an answer that takes five seconds feels broken," and that the search itself took only about 80 milliseconds, so the language model was the bottleneck. With Mercury 3 the median response arrived after 0.9 seconds, and the full process took 2.2 seconds. Education company General Learning says it now delivers hints and corrections to students in about one second. Customer service provider Lorikeet reports faster response times along with better results. At TinyFish AI, co-founder Shuhao Zhang says Mercury 3 was given 50 navigation tasks on real websites and solved 25, beating GPT-6.1-Sol; Zhang does not give a number for the OpenAI model.

Regular pricing is $0.20 per million input tokens and $0.70 per million output tokens, and Inception is cutting prices in half at launch. It has not said when the discount ends. After that, Mercury 3 would cost more for output than GPT-6 Luna and Claude Haiku 5.5, at least for shorter inputs, where both charge $0.50. For very long prompts, both rivals charge a lot more. Mercury 3 can handle up to 1 million tokens at once.

The model is available on Inception's own platform as well as OpenRouter, Baseten and Vercel. It uses the same API as OpenAI's models, so Inception says developers can switch with little effort.

For background: in February Inception released Mercury 2, which it called the first diffusion model to reason through intermediate steps before answering. Before that, the startup raised $50 million. Backers include NVentures, M12 and Snowflake Ventures, the venture arms of Nvidia, Microsoft and Snowflake. Google is working on the same approach: its Gemini Diffusion demo was followed in June by the open model DiffusionGemma, which Google says runs up to four times faster than the standard Gemma 4 but delivers slightly weaker results.

Key facts

  • Inception says Mercury 3, a diffusion-based language model, wins eight of 12 tests in its own table against GPT-6 Luna and Claude Haiku 5.5, all at medium settings, while running four to eight times faster.
  • Inception reports 1,068 tokens per second on Nvidia GPUs, against about 240 for Claude Haiku 5.5 and about 130 for GPT-6 Luna; no independent tests exist yet.
  • Mercury 3 scores 76.84 on IFBench (15 points above GPT-6 Luna) and 51.08 on SWE-Bench Atlas (rivals 33.33 and 32.26), but Claude Haiku 5.5 wins DeepSWE and GPT-6 Luna edges ahead in two Artificial Analysis tests.
  • Regular pricing is $0.20 per million input tokens and $0.70 per million output tokens, halved at launch with no stated end date; the context window is up to 1 million tokens.
  • The model runs on Inception's platform, OpenRouter, Baseten and Vercel, with an OpenAI-compatible API.

Why it matters

Most language models write one word at a time, and each word waits for the last. Mercury 3 is Inception's argument that diffusion, which refines a draft of the whole answer over several passes, can cut that wait sharply. Inception says Mercury 3 has more than 40 times the parameters of Mercury 1 yet runs just as fast, and it takes that as a sign diffusion models improve with scale like standard ones. Google is pursuing the same approach with Gemini Diffusion and the open DiffusionGemma, which Google says runs up to four times faster than Gemma 4 with slightly weaker results.

Who it affects

Inception aims Mercury 3 at apps where small delays pile up: AI agents, search, customer support and tutoring. Early-access partners include Octen AI (smart-glasses search), General Learning (student hints in about one second), Lorikeet (customer service) and TinyFish AI (web navigation). Developers on OpenAI-style APIs are the other group, since Inception says switching takes little effort.

How to use it

Mercury 3 is available on Inception's own platform and on OpenRouter, Baseten and Vercel. It uses the same API as OpenAI's models. Regular pricing is $0.20 per million input tokens and $0.70 per million output tokens, cut in half at launch; Inception has not said when the discount ends. At regular prices, output would cost more than the $0.50 that GPT-6 Luna and Claude Haiku 5.5 charge for shorter inputs, though both charge a lot more for very long prompts. The model handles up to 1 million tokens at once.

How solid is it

All figures are Inception's own, and no independent tests exist yet. Inception's table gives Mercury 3 eight wins out of 12 with all models at medium settings, while its short overview says only that quality is on par or better. Partner results are mostly about speed; at TinyFish AI, Mercury 3 solved 25 of 50 navigation tasks and beat GPT-6.1-Sol, but Zhang gives no number for the OpenAI model. The exact parameter count is undisclosed.

Risks and caveats

The lead is narrower than the announcement suggests: Claude Haiku 5.5 wins the longer coding test DeepSWE and does better at finding information in long texts, and GPT-6 Luna edges ahead in two Artificial Analysis tests. The low launch price may not last. The source does not say which rival scored 33.33 and which 32.26 on SWE-Bench Atlas, and the speed gap rests on Inception's own comparison.

“On smart glasses, an answer that takes five seconds feels broken”

— Kuan Zou, founder of Octen AI