Mercury 3 Code Passes 2,000 Map Checks in Test That Skips Speed

Inception’s Mercury 3 produced game code that cleared 2,000 map-generation checks in a separately published eight-run API test, but the test did not assess its claimed speed advantage against rivals.

The code passed a syntax check and 2,000 randomized checks of specified map-generation conditions. Those results cover the tested conditions, not every feature of the game. Mercury 3 also answered one familiar logic puzzle correctly. On one business-math prompt, it got the core calculations right but made an error about percentage margin.

The model exceeded a requested word limit in a writing task. The eight runs offered examples of how Mercury 3 handled different prompts, but they did not provide a controlled comparison with OpenAI’s GPT-6 Luna or Anthropic’s Claude Haiku 5.5.

A rolling snapshot of live requests routed through OpenRouter listed Mercury 3’s Inception provider at 246 output tokens per second on Oct. 21. The reading covered 30 minutes of traffic. It was not a matched speed test against other models or against Inception’s GPU result.

Inception measured 1,068 tokens per second on NVIDIA GPUs when it announced Mercury 3 on Oct. 19. The company described that rate as four to eight times the speed of GPT-6 Luna and Claude Haiku 5.5. Its company-run chart placed Mercury 3 ahead in eight of 12 comparisons.

The chart also placed Mercury 3 behind Claude Haiku 5.5 on a coding test and a long-context retrieval test. The later eight-run API test did not independently reproduce Inception’s 1,068-token-per-second figure, and its results do not establish whether the claimed speed difference holds under matched conditions.

Mercury 3 uses diffusion to generate and refine multiple parts of a response in parallel. Autoregressive models generate text one token at a time. The two speed readings measure different settings: one is Inception’s NVIDIA GPU test, while the other reflects live requests over a 30-minute window.

Mercury 3 has a 1 million-token context window, up from 260,000 tokens for Mercury 2.5. Developers can access it through Inception’s API platform, OpenRouter, Baseten and Vercel. Standard pricing is $0.20 per million input tokens and $0.70 per million output tokens, with a 50% launch discount.