Inception launches Mercury 3, a diffusion LLM that generates 1,068 tokens per second. Image: Inception
Delays add up when an AI agent needs multiple model calls to complete a task.
Inception introduced Mercury 3 on Oct. 19, claiming faster text generation and competitive benchmark performance against models from OpenAI and Anthropic. For developers building customer support, search and browser agents, the practical question is whether that speed translates into faster successful tasks at an acceptable cost.
Unlike conventional autoregressive models, which generate text one token at a time, diffusion models can generate and refine multiple tokens in parallel. Inception says this approach allows Mercury 3 to reach 1,068 tokens per second on Nvidia GPUs while matching or outperforming OpenAI’s GPT-6 Luna and Anthropic’s Claude Haiku 5.5 across a range of benchmarks.
According to Inception, Mercury 3 has more than 40 times as many total parameters as Mercury 1, with a reported threefold improvement across a range of standard benchmarks. The new model also supports a context window of 1 million tokens, up from 260,000 in Mercury 2.5, giving developers more room to process lengthy documents and large collections of information in a single request.
Speed could change how AI agents work
Mercury 3 is aimed at applications where delays accumulate across repeated model calls, including browser agents, live search, tutoring and customer support.
Inception said early-access partner TinyFish AI tested the model on 50 live Mind2Web navigation tasks, where it completed 50% of the tasks. The company said that result exceeded GPT-6.1-Sol’s performance using the same token volume and number of agent steps.
Education platform RevisionDojo reported that Mercury 3 returned hints, definitions and marked answers in about a second. Octen AI, which develops search technology for smart glasses, reported a median generation time of 0.9 seconds and a median end-to-end response time of 2.2 seconds.
These early-access partner reports, published by Inception, illustrate potential uses rather than independently establish a general performance advantage. For teams evaluating agents, the useful measure is the time and cost required to finish a task correctly, including tool calls, retries and corrections.
More must-read AI coverage
- SS&C Intralinks DealCentre AI vs. Datasite: Which platform is built for the future of dealmaking?
- SS&C Intralinks FundCentre AI vs. Juniper Square: Which platform better supports modern private markets fund managers?
- Why Data, Not Models, Determines AI Success
- The Rise of the AI-Native Factory: How Physical AI Is Transforming Manufacturing
The economics of faster inference
Mercury 3 costs $0.20 per million input tokens and $0.70 per million output tokens at standard rates. Its introductory prices cut those figures in half, to $0.10 and $0.35, respectively. Inception’s launch announcement does not specify when the introductory discount ends.
The company’s larger ambition is to challenge the assumption that more capable AI necessarily requires slower inference or increasingly expensive computing infrastructure. But token-generation speed alone does not establish lower total operating costs. Actual savings depend on hardware utilization, task complexity, output length and how often a model needs to retry or correct its work.
What developers should test before switching
For users, faster inference could make AI tutors feel more interactive, voice assistants less awkward and automated research tools more responsive. A million-token context window could also help applications work with extensive documentation without repeatedly splitting information into smaller requests.
Developers can access Mercury 3 through Inception’s API platform, OpenRouter, Baseten and Vercel. Its OpenAI API compatibility is intended to simplify integration into existing applications, while enterprise deployments offer dedicated capacity, autoscaling, compliance controls and configurable data retention.
Teams evaluating Mercury 3 should compare it with their current model on representative workloads. Measure end-to-end response time, answer accuracy, tool-use success and cost per completed task. Faster text generation offers less benefit when external searches, API calls or repeated corrections account for most of the delay.
The million-token context window also needs workload testing. Developers should check whether the model reliably retrieves and uses relevant details from long inputs, while tracking the additional input-token costs. Enterprise buyers should confirm which compliance controls and data-retention settings apply to their chosen deployment.
Existing Mercury customers have a migration deadline to consider. Inception says customers who used Mercury 2 or 2.5 before Oct. 19 can retain access until Nov. 15, 2026; enterprise customers remain supported under their agreements. Those teams should use the remaining period to validate compatibility and performance before moving production workloads.
Read more: Faster generation is only part of the budget equation — learn why AI agent cloud costs are difficult to predict and which controls help teams manage spending.