Inception says Mercury 3 generates 1,068 tokens per second

Summary

• Inception launched Mercury 3, a diffusion language model it says runs 4-8 times faster than rivals.
• The company claims a 1 million-token context window, up from 260,000 in Mercury 2.5.
• Launch pricing of $0.10 per million input tokens reshapes cost math for real-time agents.

The latest

A diffusion-based language model clocking 1,068 tokens per second on Nvidia GPUs went live Monday, as Inception Labs introduced Mercury 3 across its own API, OpenRouter, Baseten and Vercel. The company says the model outperforms GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks while running four to eight times faster. Partner firms supplied the early performance figures.

Details

• The speed claim: Inception says Mercury 3 generates 1,068 tokens per second on Nvidia graphics processors, and that total throughput exceeds Mercury 1 by more than 40 times. The company did not publish the absolute throughput number behind that multiple, nor detailed benchmark tables beyond the partner tests it cites.
• Context window: The model carries a 1 million-token context window, according to the company, against 260,000 tokens in Mercury 2.5. Inception also says benchmark intelligence scores rose three times from Mercury 1 mini to Mercury 3, a comparison drawn against its own earlier releases rather than competing models.
• Pricing: List pricing is $0.20 per million input tokens and $0.70 per million output tokens, according to Inception. A 50% launch discount cuts those to $0.10 and $0.35, while cached input tokens carry a 90% discount off the input price. The company has not said how long the launch discount runs.
• Browsing test: TinyFish ran Mercury 3 against frontier models on 50 live browsing tasks drawn from Mind2Web, and Inception says the model completed 50% of them, well above GPT-6.1 Sol. Shuohao Zhang, TinyFish AI co-founder, said the test ran inside the firm's production web agent stack.
• Latency figures: In Sense API testing at Octen AI, average generation time measured 0.9 seconds and average end-to-end response time 2.2 seconds, according to the company. Octen founder and chief executive Kuan Zhu said Mercury 3 delivered the best speed-quality balance among the fast models his team tested.
• Customer accounts: Michael Tung, co-founder and chief technology officer of General Learning, which operates RevisionDojo, said the model let the company give every student instant feedback without giving up reasoning. Michael Gribben of Lorikeet's technical staff said early tests showed stronger latency and higher scores in evaluations and in production.
• Availability: Mercury 3 is available now on the Inception API platform, OpenRouter, Baseten and Vercel, and is compatible with the OpenAI API interface. Enterprise deployment comes with dedicated capacity, autoscaling, compliance controls and configurable data retention. New users get 100 million free tokens on the API platform.
• Companion products: Inception also released Mercury Voice, tuned for low response latency and already in use by Audivi AI, Altur and OpenCall, alongside Mercury Decide, a classification model the company says outperforms Jev on JevBench and LegalBench.

Background

Diffusion models generate text by refining an entire output in parallel rather than predicting one token after another, the approach that underpins Inception's speed claims against sequential language models.

Between the lines

Every performance figure here originates with Inception or with partner firms that agreed to be named, and the company has not released detailed benchmark methodology outside those partner tests. The 50% task completion rate on live browsing, if it holds in independent testing, matters more to agent developers than raw token speed, since browsing agents fail on reliability rather than latency.

What's next

Watch for independent benchmark runs on OpenRouter and Baseten, where third-party developers can measure the throughput and latency claims directly, and for Inception to set an expiry date on the 50% launch discount.

Source: Inception Labs