October 20, 2026 | AI | By Mahesh

Inception, the startup behind the Mercury diffusion language models, launched Mercury 3 on October 19, 2026, a model that generates text at 1,068 tokens per second on NVIDIA GPUs with a 1 million token context window [1][2]. It costs $0.20 per million input tokens and $0.70 per million output tokens, with a 50% launch discount, and is available only through an API with closed weights [2]. Inception says it matches or beats GPT-6 Luna and Claude Haiku 5.5, and nearly all of those numbers are Inception's own.

What Mercury 3 is and what is new

Mercury 3 is a diffusion LLM. It drafts a full answer and refines every position in parallel over a few passes, where a conventional model generates one token at a time [2]. Inception says the model has more than 40 times the total parameters of Mercury 1 and that its intelligence on standard benchmarks has risen about threefold since the first Mercury mini [1]. The context window grows from 260,000 tokens in Mercury 2.5 to 1 million.

Inception argues that Mercury 3 shows text diffusion scales, and that scaling laws seen in autoregressive models also hold for diffusion [1]. The launch includes two companion products. Mercury Voice targets time to first answer token in voice apps, and Mercury Decide is a classification model for high-volume decisions [1].

"In our Sense API tests, Octen search took about 80ms; model generation was the bottleneck."

Kuan Zou, Founder and CEO, Octen AI, in Inception's announcement [1]

Under the hood: diffusion, size and serving

Inception has not disclosed the parameter count, the architecture details or the exact GPU behind the speed figure [1][2]. The 1,068 tokens per second is Inception's measurement on its own NVIDIA GPU serving stack, and it claims a speed advantage of 4 to 8 times over comparable models [1][2]. In an Octen AI test, median generation time was 0.9 seconds and median end-to-end response time 2.2 seconds [1].

The model string is mercury-3 and the API follows the OpenAI format with a reasoning_effort parameter, so it can replace an existing model in a stack with little change [2]. Early partners use it for voice and chat support, tutoring, data engineering and web navigation [1].

Benchmarks: vendor tests and one outside spot check

Inception says Mercury 3 leads in 8 of 12 tests against GPT-6 Luna and Claude Haiku 5.5, all at medium settings, and the speed figures for the rival models are also Inception's measurements [2]. Some of the larger gaps and the cases where it trails are below.

Haiku 5.5 leads on DeepSWE v1 and MRCR v2, which points to long-context retrieval and longer coding tasks as the weak spots [2]. Depth Grid covered that rival in Claude Haiku 5.5 Launch: $0.10 Input, 1M Context, Benchmarks. Inception also reports that Mercury 3 completed 50% of 50 live Mind2Web browser tasks in TinyFish's agent harness, using the same token volume and steps as a frontier model, a customer test that has not been independently verified [1][3].

Corey Noles of The Neuron ran eight spot checks through OpenRouter for a total cost of about $0.04. The model wrote a working single-file dungeon game and solved a logic puzzle correctly, but it overshot a 250-word limit by about 24 words and, in a hallucination test, added an incorrect claim about 1986 to 1988 Shuttle launches after three web searches [3]. He did not reproduce the speed figure and calls the checks practical, not a controlled benchmark.

Pricing, availability and how to access it

List price is $0.20 per million input tokens and $0.70 per million output tokens. The launch promotion halves that to $0.10 and $0.35, and cached input costs 90% less than the input price. New accounts on the Inception API platform get 100 million free tokens [1][2][3]. The launch discount is temporary, so costs roughly double when it ends [2].

Mercury 3 is available through the Inception API, OpenRouter, Baseten and Vercel AI Gateway, and in Mercury Chat for a quick test. Enterprise plans add dedicated capacity, autoscaling, compliance controls and configurable data retention [1][2]. Weights are closed, so there is no self-hosting.

Limitations, trade-offs and alternatives

The Neuron found imperfect control of exact word counts and setting rules, explanations that sound more certain than warranted and errors when tool-assisted answers pulled in web results [3]. Inception's page lists no limitations, and the speed, benchmark and agent figures are vendor-run or customer statements [1][2].

For buyers who want open weights, Depth Grid has covered Mistral Large 4, and for a closed model at a higher price, Meta Muse Wonder. Teams that want speed for voice, chat or agent loops should test Mercury 3 on their own prompts and compare cost per finished task, not tokens per second.

Sources

1. Inception. "Diffusion at Scale: Introducing Mercury 3," October 19, 2026. inceptionlabs.ai
2. MarkTechPost. "Inception Releases Mercury 3 Diffusion LLM at 1,068 Tokens per Second," October 19, 2026. marktechpost.com
3. Corey Noles, The Neuron. "Mercury 3 Review: Inception's Diffusion AI Put to the Test," October 20, 2026. theneuron.ai

Read More on Depth Grid

• Claude Haiku 5.5 Launch: $0.10 Input, 1M Context, Benchmarks
• Mistral Large 4 Preview: 1.05T Open-Weight MoE, Specs and Pricing
• Meta Muse Wonder Launch: 1M Context, $4 API Price, Benchmarks
• Nvidia in Talks to Deepen Stake in or Buy Reflection AI

Article by Mahesh | Depth Grid