Table of Contents
Most AI chatbots write one word fragment at a time, which puts a ceiling on how fast they can respond. A start-up called Inception is betting on a different design. Its newest release, Mercury 3, is a diffusion-based language model that the company says beats well-known rivals on a range of benchmarks while running several times faster. The claims are bold, and they come with an important caveat.
What Inception Announced With Mercury 3
Inception released Mercury 3 on Monday, 19 October. It can work with a context of one million tokens, and Inception states that it beats GPT-6 Luna and Claude Haiku 5.5 across several benchmarks, at speeds four to eight times higher. The announcement lists a speed of 1,068 tokens per second on NVIDIA GPUs.
The release is Inception’s biggest test so far of whether diffusion models can grow larger without losing their speed advantage. One point matters throughout this story: every performance figure so far comes from Inception itself.
For readers new to the field, a context window is the amount of text a model can consider at once. A window of one million tokens is large enough to handle very long documents or extended conversations in a single request, though how well the model uses all that space has not been independently checked.
Bigger Model, Same Speed
How Diffusion Differs
Standard language models write text one token at a time. Diffusion models generate many tokens in parallel and then refine them. That parallel approach is the reason Inception believes it can respond so quickly.
Scaling Without Slowing Down
Inception says Mercury 3 is more than 40 times larger than Mercury 1 and still runs at over 1,000 tokens per second. “We’re showing that the limits to scaling diffusion models are compute and data, not fundamental algorithmic barriers,” said co-founder Stefano Ermon, who added that training on the next model has already begun.
That comment captures Inception’s pitch. If compute and data, rather than hard design limits, are what hold diffusion back, then larger versions should keep improving.
Mercury 3 is the company’s fourth release since Mercury 1 launched in June 2025. Its context window is larger than the 260,000 tokens that Mercury 2.5 supported.
The Web-Navigation Test
In company testing on 50 live Mind2Web web-navigation tasks, Inception said Mercury 3 completed half of them, which it described as well above GPT-6.1-Sol’s result. Again, this is a company test rather than an outside evaluation.
Pricing and Availability
List prices for Mercury 3 are $0.20 per million input tokens and $0.70 per million output tokens. A 50% launch discount brings those down to $0.10 and $0.35, and cached input tokens cost 90% less than regular input.
The model is available through the Inception API, Baseten, OpenRouter, and Vercel, and it works with the OpenAI API format. That compatibility should make it easier for developers to try without rewriting their applications. The discounted rates are low enough for teams to run sizeable experiments cheaply, while the cached-input savings help applications that reuse the same long prompt repeatedly.
Mercury 2 and 2.5 are no longer open to new users, and existing customers keep access until 15 November. Education start-up RevisionDojo said hints and marked answers now come back in about a second.
Claims Await Independent Verification
What Outside Data Shows
No independent evaluator has yet checked Inception’s benchmark comparisons. As of 20 October, the benchmarking firm Artificial Analysis had no listing for Mercury 3. It had previously measured Mercury 2.5 at about 706 tokens per second, which was below the speed Inception claimed for that model.
OpenRouter’s live traffic data showed a median throughput of 221 tokens per second on 20 October. That is faster than both rival models, but only about a fifth of Inception’s headline figure.
Why the Gap May Exist
A lab’s headline number and live traffic often differ, since real-world requests vary in length and load. The OpenRouter figure still points in the same direction as Inception’s story, because it beat both rivals, yet it is far from 1,068 tokens per second. Co-founder Volodymyr Kuleshov said on Hacker News that third-party benchmarks are still to come. Until an outside group publishes results, buyers should treat the benchmark comparisons as claims rather than facts.
What This Means for Everyday Users
Speed shapes how an AI tool feels. When marked answers or hints arrive in roughly a second, study apps, coding assistants, and support tools start to feel immediate instead of slow or sluggish. If Mercury 3 holds up, users may notice quicker replies without paying more.
For developers and teams, the sensible move is to test before committing. Run your own prompts, measure real throughput, and compare quality against the models you use today. Start with the discounted launch pricing to keep trial costs low, and remember that the discount is described as a launch offer.
Frequently Asked Questions
1. What is Mercury 3?
A diffusion-based language model from Inception with a 1-million-token context window, released on Monday, 19 October.
2. How fast is it claimed to be?
Inception lists 1,068 tokens per second on NVIDIA GPUs and says it runs four to eight times faster than GPT-6 Luna and Claude Haiku 5.5.
3. How much does it cost?
List prices are $0.20 per million input tokens and $0.70 per million output tokens, with a 50% launch discount and 90% savings on cached input.
4. Have the claims been verified?
Not independently. Artificial Analysis had no listing as of 20 October, and OpenRouter showed a median of 221 tokens per second.
Conclusion
Mercury 3 is a serious test of whether diffusion models can scale up without giving away their speed. Inception’s numbers are impressive, and pricing is low, but the evidence so far is mostly the company’s own. Independent benchmarks will show whether Mercury 3 deserves the attention it is seeking. Developers who test it themselves this month will quickly learn more than any launch announcement can say.