Inception's Mercury 3 benchmarks lack outside checks
Inception claims 1,068 tokens per second and wins in eight of 12 tests. So far no outside lab has published its own measurements.
In short
Inception says Mercury 3 wins eight of 12 benchmarks against GPT-6 Luna and Claude Haiku 5.5 while generating 1,068 tokens per second, but nobody outside the company has verified that yet.
At a glance
- Speed: 1,068 tokens per second on Nvidia GPUs, against roughly 240 for Claude Haiku 5.5 and 130 for GPT-6 Luna.
- Scores: SWE-Bench Atlas 51.1 vs 33.3 (Luna) and 32.3 (Haiku); Haiku 5.5 takes DeepSWE 51.6 to 46.3.
- Context: up to 1 million tokens, up from 260,000 in Mercury 2.5; parameter count over 40x Mercury 1, exact figure withheld.
- Price: $0.20 per million input tokens and $0.70 per million output tokens, both halved at launch.
- Company: founded 2024, $50 million raised, Microsoft's M12 venture arm among the backers.
Inception's new diffusion language model, Mercury 3, is pitched on one number: speed. The company says it wins eight of 12 benchmarks against GPT-6 Luna and Claude Haiku 5.5 while running four to eight times faster. Every figure comes from Inception, and no third party has published results of its own.
The claimed numbers
Inception puts Mercury 3 at 1,068 tokens per second on Nvidia GPUs, against roughly 240 for Claude Haiku 5.5 and roughly 130 for GPT-6 Luna. Diffusion models do not emit text strictly left to right; they refine a whole block of text over repeated passes, which is where the throughput advantage is supposed to come from.
On the task side, the company reports 51.1 on SWE-Bench Atlas versus 33.3 for GPT-6 Luna and 32.3 for Claude Haiku 5.5, plus 76.84 against 61.84 for GPT-6 Luna on the IFBench instruction test. The context window reaches 1 million tokens, up from 260,000 in Mercury 2.5. Inception will only say the model is more than 40 times the size of Mercury 1.
Where it loses
Four of the 12 tests do not go Inception's way, and the clearest loss is DeepSWE, the longer coding evaluation, where Claude Haiku 5.5 leads 51.6 to 46.3. That gap matters for buyers weighing speed against long agentic task chains, and it is the part a self-reported benchmark sheet cannot settle.
Pricing and access
Mercury 3 lists at $0.20 per million input tokens and $0.70 per million output tokens, with cached input about 90 percent cheaper and both rates halved as a launch promotion. That does not make it the cheapest option: SiliconANGLE cites $0.10 and $0.50 for the rival models on typical prompts. The model is served through Inception's own platform plus OpenRouter, Baseten and Vercel, behind an OpenAI-compatible API.
Early customer figures also come through Inception. Smart-glasses maker Octen AI reports a 0.9-second median generation time, a tutoring product cites hints returned in about a second, and TinyFish AI completed 25 of 50 web navigation tasks. Useful signals, but still vendor-routed ones.
The open question
Chief executive Stefano Ermon frames the result as a compute-and-data story rather than an algorithmic breakthrough, which is a direct answer to skeptics who doubt diffusion can scale. Inception was founded in 2024 by Ermon, Aditya Grover and Volodymyr Kuleshov and has raised $50 million, with Microsoft's M12 among its investors; the two sources used here name different lead investors for that round. SiliconANGLE adds that Microsoft held acquisition talks in May, with Inception seeking more than a $1 billion valuation, and that SpaceX also approached the company. No deal has been announced.
FAQ
How fast is Mercury 3 compared with GPT-6 Luna?
Inception reports 1,068 tokens per second for Mercury 3 against roughly 130 for GPT-6 Luna and roughly 240 for Claude Haiku 5.5. These are vendor numbers with no independent test behind them.
How much does Mercury 3 cost per million tokens?
$0.20 for input and $0.70 for output tokens, both halved during the launch promotion, with cached input about 90 percent cheaper.
Where can developers run Mercury 3?
On Inception's own platform and through OpenRouter, Baseten and Vercel. The API is OpenAI-compatible with minor changes.
