Inception says Mercury 3 beats GPT-6 Luna and Claude Haiku 5.5 at up to 8x the speed

Language model startup Inception AI Inc. has unveiled Mercury 3, their latest flagship model. In their own tests, Mercury 3 outperformed OpenAI Group PBC's GPT-6 Luna and Anthropic PBC's Claude Haiku 5.5 on eight out of 12 benchmarks, matching them in four and surpassing them by four to eight times the speed. This victory addresses the primary concern regarding Inception's methodology: scalability.

Co-founder and CEO Stefano Ermon noted that the main hurdle lies in compute and data, rather than fundamental algorithmic limitations. Unlike GPT-6 Luna and Haiku 5.5, which generate responses token by token, Mercury 3 creates the entire response at once and refines each position in parallel over a few passes. Inception's measurements reveal Mercury 3's speed, with 1,068 tokens per second on Nvidia Corp. GPUs, compared to Haiku 5.5's 240 tokens per second and GPT-6 Luna's 130 tokens per second.

The model's parameter count remains undisclosed, but it's more than 40 times larger than Mercury 1 and boasts a context window of 1 million tokens, a significant increase from Mercury 2.5's 260,000 tokens. Ermon hinted at ongoing training for Mercury 3's successor. In coding benchmarks, Mercury 3 excelled over GPT-6 Luna (51.1 vs. 33.3 on SWE-Bench Atlas, 32.3 vs. 32.3 on Haiku 5.5) and Haiku 5.5 on DeepSWE (51.6 vs. 51.6).

However, it trailed Haiku 5.5 by a narrow margin in one long-context test. GPT-6 Luna won two other benchmarks by less than a point each. As of now, no external entity has published its independent assessment of Mercury 3. Early adopters are integrating Mercury 3 into products where every delay is noticeable. Kuan Zou, founder and CEO of Octen AI, a real-time search startup, found that a five-second delay feels broken, with Octen's median generation time at 0.9 seconds.

At tutoring app RevisionDojo, hints and marked answers are now returned in about a second. Pricing for Mercury 3 is $0.20 per million input tokens and $0.70 per million output tokens, with cached input being 90% cheaper. A launch promotion currently slashes both rates. Mercury 3, GPT-6 Luna, and Haiku 5.5 are priced at $0.10 and $0.50 per prompt, respectively.

Developers can access Mercury 3 via Inception's platform or through OpenRouter Inc., Baseten Labs Inc., and Vercel Inc. Existing OpenAI API-written software requires minimal adjustments to operate with Mercury 3. Inception AI Inc. is a venture capital-backed startup founded in 2024 by Ermon and fellow computer science professors Aditya Grover and Volodymyr Kuleshov.

Microsoft Corp.'s M12 venture arm participated in a $50 million round led by Menlo Ventures in November last year. Although Microsoft reportedly attempted to acquire Inception in May, no deal has materialized. Other interested parties include Space Exploration Technologies Corp.