Inception released Mercury 3 on October 19, 2026, claiming 1,068 tokens per second on NVIDIA GPUs and a vendor-measured 4–8× speed advantage over GPT-6 Luna and Claude Haiku 5.5. Artificial Analysis, the standard independent arbiter the industry has grown dependent on, has no Mercury 3 page as of launch day. Every benchmark in circulation is Inception's own.

The pricing is the headline the vendor would prefer. List is $0.20 per million input and $0.70 per million output, discounted 50% during the launch window to $0.10 and $0.35. Cached inputs run 90% off list, which lands at $0.02 per million (or $0.01 promotional). A 20,000-token system prompt costs $0.0004 per call cached versus $0.004 uncached; across a fifty-step agent run, that's the difference between two cents and twenty cents of pure preamble. For a plausible support or search workload of 10M input and 40M output tokens per day, the bill works out to roughly $30/day at list, $15/day promotional, or about $900 and $450 per month respectively.

That's the floor Mercury 3 is attacking, and it arrives ten days after Anthropic reset its own speed tier when Claude Haiku 5.5 shipped at $0.10/M. The speed-tier price floor is now a contested surface rather than a resting place.

Architecturally, Mercury 3 is a diffusion LLM, Inception's fourth model since Mercury 1 shipped in June 2025, and the company says it's more than 40× larger than Mercury 1 in total parameters, with measured intelligence roughly tripling across its benchmark suite. The context window expands to 1 million tokens from 260,000 on Mercury 2.5, with maximum output of 131,072 tokens. Four reasoning_effort settings (instant, low, medium, high) ship with a default of low, down from medium on prior Mercury generations. Temperature on the OpenAI-compatible endpoint is adjustable between 0.5 and 1.0.

The customer evidence Inception foregrounds is latency, not quality. Michael Tong, co-founder and CTO of RevisionDojo, reports that "Hints, marked answers and definitions now come back in about a second" and calls it a drop-in swap. Kuan Zou, founder and CEO of Octen AI, reports that in Sense API tests "Octen search took about 80ms" on smart glasses, with 0.9-second median generation and 2.2-second end-to-end response. Michael Gribben at Lorikeet and Shuhao Zhang of TinyFish AI echo the pattern. Co-founder Stefano Ermon frames the release as a Pareto-frontier move.

The vendor harness also reports 50% completion on 50 live Mind2Web navigation tasks, and a Mercury Voice variant with time-to-first-token under 170 milliseconds at $0.40/$1.50 list ($0.20/$0.75 promotional). Mercury Decide prices input at $0.04 (promotional $0.02) with free output. Distribution runs through OpenRouter, Baseten, and Vercel; 100 million free tokens come with new accounts; Mercury 2 and Mercury 2.5 access ends November 15, 2026 for non-enterprise customers.

Artificial Analysis has historically clocked Mercury 2.5 at roughly 706 tokens per second and Mercury 2 in the 680s under its own harness. Those numbers were lower than Inception's at the time, and they shipped only after Artificial Analysis published. The 1,068-figure will get the same treatment. Until it does, this is a price story with a speed claim attached, and a quality claim parked in escrow.

Sources

https://www.inceptionlabs.ai/blog/introducing-mercury-3
https://www.morningstar.com/news/business-wire/20261019090400/inception-launches-mercury-3-setting-a-new-pareto-frontier-in-speed-and-intelligence
https://www.unite.ai/inception-releases-mercury-3-its-flagship-diffusion-llm/
https://runtimewire.com/article/inception-mercury-3-diffusion-llm-launch
https://www.orcarouter.ai/blog/inception-mercury-3