Inception Releases Mercury 3, a Diffusion Language Model It Says Beats GPT-6 Luna

Inception, an AI lab based in Redwood City, California, on October 19, 2026 released Mercury 3, a diffusion language model it calls the most capable of its kind. The company says Mercury 3 scores higher than OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5 on most of its benchmarks while generating 1,068 tokens per second.

Mercury 3 arrives six weeks after Mercury 2.5, which Inception launched on September 8, 2026 at 1,107 tokens per second by its own count (RuntimeWire). It takes aim at the cheap, fast small models that OpenAI and Anthropic sell for high volume work, but with a different design underneath.

What did Inception release?

Mercury 3 is Inception's new flagship model, announced on October 19, 2026 in a company blog post that says it is more than 40 times larger than Mercury 1 in total parameters and has a 1 million token context window, up from 260,000 on Mercury 2.5. Inception has not disclosed the parameter count itself (The Deep View).

Unlike most language models, which are autoregressive and write one token at a time, a diffusion model drafts many tokens at once and refines them in parallel over several passes. Inception says that approach lets Mercury 3 run fast on standard Nvidia GPUs. The company says its scores on a mix of standard benchmarks have roughly tripled from Mercury 1 (mini), released in February 2025, to Mercury 3, and argues that the scaling that drove autoregressive models also holds for diffusion.

Chief executive Stefano Ermon told The Deep View that Inception used swarms of Mercury 3 agents to improve the model, which he said added about 16% to its speed on top of the team's own work. That is a company claim with no published detail.

How does Mercury 3 compare with GPT-6 Luna and Claude Haiku 5.5?

In Inception's own table, published on October 19, 2026, Mercury 3 scores higher than GPT-6 Luna on 10 of 12 benchmark rows and higher than Claude Haiku 5.5 on 10 of 12. It has the top score of the three on 8 of the 12 rows.

GPT-6 Luna leads Mercury 3 on two rows, AA-Omniscience and the AA-LCR long context test. Claude Haiku 5.5 beats Mercury 3 on the MRCR v2 long context test and on DeepSWE, an agentic coding benchmark. Both rivals are their makers' current small models, so the comparison is not against outdated versions. Inception's blog does not say who ran the rival scores, and The Next Web notes that no outside group has tested Mercury 3 yet.

The speed figures in the table work out to about 8 times GPT-6 Luna and 4.5 times Claude Haiku 5.5, which matches the 4 to 8 times range Inception claims. They also suggest the larger model kept pace with its predecessor: 1,068 tokens per second for Mercury 3 against the 1,107 Inception reported for Mercury 2.5.

The scores do not line up with other vendors' tables. Anthropic's own Haiku 5.5 launch figures used a different GDPval version and scale, so numbers from the two tables should not be mixed.

How much does Mercury 3 cost?

Mercury 3 lists at $0.20 per million input tokens and $0.70 per million output tokens, and Inception is selling it at half that, $0.10 and $0.35, at launch, the company said. Cached input tokens cost 90% less than regular input.

At the launch price, Mercury 3 matches GPT-6 Luna and Claude Haiku 5.5 on input and costs 30% less on output, based on the $0.10 and $0.50 per million tokens that OpenAI and Anthropic list for their models. At the list price, the order flips: Mercury 3 would cost twice as much as both on input and 40% more on output. Anthropic's rate applies to prompts up to 100,000 tokens and rises above that.

Inception has not said how long the discount lasts (The Next Web). Ermon told The Deep View that the company is working on changes that could make the launch price permanent.

Who can use Mercury 3, and where?

Mercury 3 is available on Inception's API platform and through OpenRouter, Baseten and Vercel's AI Gateway, according to Inception. Vercel lists it under the model name inception/mercury-3 with the 50% launch discount.

The API follows OpenAI's format, so existing code can switch with few changes, and Inception offers 100 million free tokens for testing. Early partners quoted by Inception include TinyFish AI, which says Mercury 3 completed 50% of 50 live Mind2Web web navigation tasks, more than OpenAI's GPT-6.1-Sol in its own harness, and Octen AI, which reports a median generation time of 0.9 seconds. These are customer claims published by Inception.

Inception also sells Mercury Voice, tuned for voice agents, and Mercury Decide, a classification model it says beats TypeSafe AI's Jev on JevBench and LegalBench.

What we don't know yet

• How Mercury 3 performs in independent benchmarks and speed tests.
• How long the 50% launch discount lasts, and whether the price will change after it ends.
• How many parameters Mercury 3 has.
• Who ran the GPT-6 Luna and Claude Haiku 5.5 scores in Inception's table, and under what conditions.