Inception's Mercury 3: A Diffusion LLM at 1,000+ Tokens per Second. What to Test Before You Swap It In

Inception Labs says its new diffusion language model, Mercury 3, matches or beats GPT-6 Luna and Claude Haiku 5.5 on a range of benchmarks while running 4 to 8 times faster. If that holds for your workload, it matters most for agents and voice, where every model call adds waiting time. All the numbers below come from Inception, and I have not run the model.

This article summarizes the October 19 launch post, Diffusion at Scale: Introducing Mercury 3. I found it through the TLDR AI newsletter of October 20 and then read the primary post. I did not check independent benchmarks, the API docs or the pricing page.

What Inception claims

From the post:

• Quality. "On par or better than" frontier speed-optimized models such as GPT-6 Luna and Claude Haiku 5.5.
• Speed. 1,068 tokens per second on NVIDIA GPUs, which the post describes as 4 to 8 times faster than those models.
• Context. 1M tokens, up from 260K for Mercury 2.5.
• Price. $0.20 per million input tokens and $0.70 per million output tokens. At launch it is 50% off, at $0.10 and $0.35. Cached tokens are 90% off the input price.
• Scale. Inception says Mercury 3 is more than 40 times larger than Mercury 1, and that scaling laws seen in autoregressive models also hold for diffusion. That is a research claim from the vendor. The post does not publish the data behind it in the text I read.

The post does not name the specific benchmarks in the summary I read, so I cannot tell you which tasks "a range of benchmarks" covers.

Why diffusion matters for latency

A typical LLM generates one token after another. A diffusion language model, as Inception frames it, refines output in parallel passes, which is where the speed claim comes from. I am describing the vendor's framing, not explaining the model's internals, which the post does not detail.

For agents the effect is simple arithmetic. An agent that waits on the model at every step compounds small delays. Inception makes the same point about browser-use agents.

Customer evidence in the post

The post includes quotes from early-access partners. They are testimonials, not independent evaluations:

• TinyFish AI says that on 50 live Mind2Web navigation tasks in its production web-agent harness, Mercury 3 completed 50% of them, "significantly more" than GPT-6.1-Sol, with the same token volume and agent steps.
• RevisionDojo says hints and marked answers come back in about a second, and that Mercury 3 was a drop-in swap.
• Octen AI says that in its tests Mercury 3 had a median generation time of 0.9 seconds and a median end-to-end response of 2.2 seconds, the best balance among the fast models it tested.
• Lorikeet says that in early testing Mercury 3 beat other models on latency while scoring higher on evals and in production.

These are four companies describing their own tests, each quoted in the vendor's launch post. Fifty tasks is a small sample for a success rate.

Availability

Per the post, Mercury 3 is on the Inception API platform and through OpenRouter, Baseten and Vercel, and it is OpenAI API compatible. Inception offers 100 million free tokens to try the API, which I did not verify. It also lists Mercury Voice and Mercury Decide, a classification model, as part of a real-time product line.

What I would test

These are my own untested suggestions.

1. Your own eval, not theirs. Run your task set against your current fast model and Mercury 3. Compare accuracy first, then latency.

2. Measure end-to-end, not tokens per second. Time to first token and total response time for your prompt sizes matter more than peak throughput. Octen's 0.9 second generation versus 2.2 second end-to-end gap is a reminder.

3. Test long context. A 1M token window is a claim about capacity. Check quality at the lengths you use.

4. Check tool calling and structured output. Agents depend on these. The post's summary does not cover them.

5. Price your real mix. Launch pricing is half off. Compute cost at the regular $0.20 and $0.70 rates, with and without caching.

6. Check data retention and compliance if you are in a regulated area. The post says enterprise plans have configurable retention. I did not read the terms.

What I would not conclude

• That diffusion LLMs now beat autoregressive models in general. The claim is limited to speed-optimized models on Inception's chosen benchmarks.
• That a 50-task partner test predicts your success rate.
• That the speed holds on your hardware or through every provider. The 1,068 tokens per second figure is stated for NVIDIA GPUs.

Frequently asked questions

What is Mercury 3?

A diffusion-based large language model from Inception Labs, launched October 19, 2026, according to the company's post.

How fast is it?

Inception reports 1,068 tokens per second on NVIDIA GPUs and 4 to 8 times the speed of GPT-6 Luna and Claude Haiku 5.5.

How much does it cost?

$0.20 per million input and $0.70 per million output tokens, with a 50% launch discount and 90% off cached input, per the post.

What is the context window?

1M tokens, up from 260K for Mercury 2.5.

Where can I use it?

The Inception API platform, OpenRouter, Baseten and Vercel, per the post. It is described as OpenAI API compatible.

Are the benchmark results independent?

Not in the post I read. The comparisons and partner quotes come from Inception.

Is it good for agents?

It may suit latency-sensitive agents, which is the use Inception highlights. Test tool calling and accuracy on your own tasks before relying on it.

Sources

• Inception Labs, Diffusion at Scale: Introducing Mercury 3 (October 19, 2026): all model, price and partner claims.
• TLDR AI newsletter, October 20, 2026: discovery pointer only.