AI News Feed
Market watch
Apple Watch Ultra 4 Action Button Can Trigger Workouts, Navigation and ShortcutsDeveloper Says Block Blast Code Can Adjust Difficulty Around Ad ValueAnthropic Circulates Draft AI Safety Bill Allowing Court-Ordered Model BlocksOura adds Cardio Effort and automatic VO2 Max estimates to its smart ringApple's first foldable iPhone Duo to start at $1,999 as pre-orders open FridayUkraine Brigade Claims First Aerial-Drone Evacuation of Wounded Soldier
Apple Watch Ultra 4 Action Button Can Trigger Workouts, Navigation and ShortcutsDeveloper Says Block Blast Code Can Adjust Difficulty Around Ad ValueAnthropic Circulates Draft AI Safety Bill Allowing Court-Ordered Model BlocksOura adds Cardio Effort and automatic VO2 Max estimates to its smart ringApple's first foldable iPhone Duo to start at $1,999 as pre-orders open FridayUkraine Brigade Claims First Aerial-Drone Evacuation of Wounded Soldier
Companies

Inception Releases Mercury 3, Claiming Speed Edge Over GPT-6 Luna and Claude Haiku 5.5

Inception AI says its Mercury 3 diffusion model beats GPT-6 Luna and Claude Haiku 5.5 on 8 of 12 tests at 4-8x speed.

AI News Feed · 2026-10-21 08:41

The company reported Mercury 3 at 1,068 tokens per second on Nvidia GPUs in its comparison, against about 240 for Haiku 5.5 and about 130 for GPT-6 Luna. Unlike conventional models that generate one token after another, Mercury drafts a full response and refines all positions in parallel over several passes. Co-founder and CEO Stefano Ermon said the question Inception has heard most since Mercury 1 is whether diffusion can scale; he attributed limits to compute and data rather than fundamental algorithmic barriers.

Inception did not disclose Mercury 3's parameter count, saying only that it is more than 40 times the size of Mercury 1. The context window is 1 million tokens, up from 260,000 in Mercury 2.5. Ermon said training on Mercury 3's successor is already under way.

On SWE-Bench Atlas, a coding benchmark, Mercury 3 scored 51.1, compared with 33.3 for GPT-6 Luna and 32.3 for Haiku 5.5. It trailed Haiku 5.5 on DeepSWE, 46.3 to 51.6, and on one long-context test, according to the company. GPT-6 Luna won two other tests by less than one point each. No outside group has published independent results for Mercury 3.

Early customers are using the model in latency-sensitive products. Kuan Zou, founder and CEO of real-time search startup Octen AI, said a five-second answer on smart glasses 'feels broken'; Octen measured a median generation time of 0.9 seconds. Michael Tong, co-founder and CTO of tutoring app RevisionDojo, said hints and marked answers now return in about a second.

Mercury 3 lists at 20 cents per million input tokens and 70 cents per million output tokens, with cached input 90% cheaper. A launch promotion halves both rates. GPT-6 Luna and Haiku 5.5 each list at 10 cents per million input tokens and 50 cents per million output tokens for typical prompts. Developers can access Mercury 3 through Inception's platform or OpenRouter, Baseten and Vercel; software built for OpenAI's API requires minor changes.

Inception was founded in 2024 by Ermon and fellow computer science professors Aditya Grover and Volodymyr Kuleshov. Menlo Ventures led a $50 million round last November, with Microsoft's M12 among investors. Reuters reported in May that Microsoft was in talks to acquire Inception, which sought more than $1 billion, and that SpaceX had also approached the company. No deal has been announced.