Inception releases Mercury 3, claiming faster AI with a million-token context

Developers building tutoring, search and customer-support tools have a new option promising faster answers without giving up quality. Inception released Mercury 3 on October 19, 2026, claiming generation speeds of 1,068 tokens per second on NVIDIA GPUs and benchmark performance matching or exceeding GPT-6 Luna and Claude Haiku 5.5. Those comparisons come from Inception.

The diffusion-based language model is available through the Inception API Platform, OpenRouter, Baseten and Vercel. Inception says it is compatible with OpenAI’s API, positioning the release as a replacement developers can connect to an existing application rather than rebuild around a new interface.

A larger model, a longer working window

Inception’s architectural argument is about scale, not just speed. The company says Mercury 3 has more than 40 times as many total parameters as Mercury 1 while remaining just as fast. It presents the release as evidence that diffusion language models can grow more capable without losing their speed advantage.

The company also describes a threefold increase in intelligence across standard benchmarks from Mercury 1 mini to Mercury 3. That is Inception’s characterization of its benchmark results, not a measured threefold improvement in every task. Its competitive pitch targets speed-optimized models, naming GPT-6 Luna and Claude Haiku 5.5 rather than claiming leadership over every model.

Customer tests put speed inside a complete task

Early-access results in the launch announcement reach beyond text-generation speed. TinyFish AI cofounder Shuhao Zhang said Mercury 3 completed 50% of 50 live Mind2Web navigation tasks in the company’s production web-agent setup. He described that as significantly better than GPT-6.1-Sol, using the same token volume and number of agent steps at lower per-token cost.

Octen AI’s results distinguish model generation from the full wait for an answer. Founder and CEO Kuan Zou reported a median generation time of 0.9 seconds and a median end-to-end response time of 2.2 seconds in Sense API tests. Search itself took about 80 milliseconds; model generation was the bottleneck. Zou described Mercury 3 as offering the best speed-quality balance among the fast models Octen tested.

For RevisionDojo, cofounder and CTO Michael Tong said hints, marked answers and definitions returned in about a second, with quality matching or beating its previous models. He called the switch a drop-in replacement. Lorikeet’s Michael Gribben similarly reported lower latency and better evaluation and production performance in early testing for customer support. These are customer accounts published by Inception, not independent comparative evaluations.

Discounted entry, a deadline for predecessors

Inception is offering a 50% launch discount on Mercury 3’s listed token prices. The rates separate material sent to the model from the text it generates:

• Input: $0.10 per million tokens at launch, against a $0.20 list price.
• Output: $0.35 per million tokens at launch, against a $0.70 list price.
• Cached input: 90% below the applicable input rate.

For enterprise deployments, Inception lists dedicated capacity, automatic scaling, compliance controls and configurable data retention. It also offers to work with customers on task fit, evaluation design and performance under their deployment constraints, giving buyers a route to assess the model against their own requirements.

The release also starts a transition for existing users. Unite.AI reported that Inception’s models page closes Mercury 2 and Mercury 2.5 to new users. Customers with usage before October 19 retain access until November 15, 2026; enterprise customers remain supported under existing agreements. For those customers, evaluating Mercury 3 comes with a concrete migration deadline, not just an invitation to try another model.