Inception Mercury 3: The Diffusion LLM Claiming 1,000+ Tokens per Second, Explained

TL;DR: Inception’s Mercury 3, announced on October 19, 2026, is a diffusion-based large language model. It writes text by refining many tokens in parallel instead of predicting them one at a time. Inception says it matches or beats GPT-6 Luna and Claude Haiku 5.5 on a range of benchmarks while running 4–8x faster, at 1,068 tokens per second on Nvidia GPUs. It has a 1 million-token context window and costs $0.20 per million input tokens and $0.70 per million output tokens, half that during the launch discount. You can use it through Inception’s own API, OpenRouter, Baseten and Vercel. The weights are not open.

Inception Mercury 3 is the biggest test so far of an idea that has hovered around AI research for a few years: that the diffusion technique behind image generators can also produce fast, useful text. Most chatbots you know, from ChatGPT to Claude, are autoregressive. They generate an answer one token at a time, left to right. Mercury 3 doesn’t work that way, and Inception argues that is exactly why it’s so quick. Here’s what the company announced, what the numbers mean, and what to keep in mind before you switch any workload to it.

What Inception announced

In its launch post, Inception calls Mercury 3 “the most capable diffusion-based LLM released to date.” The company also published a press release claiming a “new Pareto frontier in speed and intelligence.” In plain terms, Inception is saying no other model it knows of is both this fast and this capable.

Mercury 3 at a glance

How a diffusion LLM is different

A standard LLM writes like a person typing. It picks the most likely next token, appends it, and repeats. Every new token waits for the one before it, so long answers take longer, and reasoning models that “think” before answering add even more waiting.

A diffusion model starts from a rough, noisy draft of the whole output and cleans it up over several passes. Inception co-founder and chief scientist Volodymyr Kuleshov put it this way in a Hacker News post on launch day: diffusion “generates tokens in parallel through iterative denoising rather than one at a time, so as reasoning models think longer per answer, the speed advantage over autoregressive models of the same size grows.”

If you’re new to how models count text, tokens are the small chunks of words that models read and write, and the unit that API pricing is based on. The speed figure of 1,068 tokens per second works out to several hundred words of output in about a second.

“Diffusion scales”: the bigger claim

The quiet question around diffusion text models has been whether they get smarter as they get bigger, the way autoregressive models do. Inception says yes. It says Mercury 3 is more than 40x larger than Mercury 1 in total parameters and that “intelligence across a range of standard benchmarks has risen 3x from Mercury 1 (mini) to Mercury 3,” while speed stayed about the same. “The scaling laws that drove progress in autoregressive models hold for diffusion too,” the company writes.

That’s a significant claim if it holds up. Independent benchmark results weren’t available at launch, so treat the comparisons with GPT-6 Luna and Claude Haiku 5.5 as Inception’s own numbers for now. For context on the models it’s measuring against, see our breakdown of GPT-6 Sol vs Luna.

What early customers say

Inception quoted several early-access partners. These are customer testimonials published by the company, not independent tests:

• Browser agents: TinyFish AI co-founder Shuhao Zhang says that on 50 live Mind2Web navigation tasks, Mercury 3 completed 50% of them, “significantly more than GPT-6.1-Sol,” using the same token volume and number of agent steps.
• Tutoring: RevisionDojo CTO Michael Tong says hints, marked answers and definitions “now come back in about a second,” and that switching models was “a drop-in swap.”
• Search on smart glasses: Octen AI CEO Kuan Zou reports a median generation time of 0.9 seconds and a median end-to-end response time of 2.2 seconds in its tests.
• Customer support: Lorikeet’s Michael Gribben says that in early testing Mercury 3 “beat other models on latency while also performing higher on evals and in production.”

The pattern is clear. Inception is pitching Mercury 3 for jobs where waiting is the main problem: voice assistants, agents that call a model dozens of times per task, and live chat. That’s the same latency problem behind the push toward autonomous AI agents, where small delays add up at every step.

Pricing: what it costs in practice

At the list price of $0.20 per million input tokens and $0.70 per million output tokens, Mercury 3 sits in the budget tier. With the 50% launch discount, it’s $0.10 and $0.35. Cached input tokens, which are repeated prompt text the system has already processed, are 90% cheaper than regular input. Inception is also offering 100 million free API tokens to new developers.

A rough example: an app that sends 10 million input tokens and receives 2 million output tokens in a month would pay about $3.40 at list price ($2.00 + $1.40), or about $1.70 at the launch rate. Your real bill depends on how much the model outputs. A lower per-token price doesn’t always mean a lower total if a model needs more tokens to finish a task.

How to try Mercury 3

1. Chat: Inception offers a web chat to test it with no setup.
2. API: The Inception API is OpenAI-compatible, so existing code that calls an OpenAI-style endpoint can switch by changing the base URL and model name.
3. Gateways: It’s listed on OpenRouter as inception/mercury-3, which supports tool calling and JSON-schema structured outputs. It’s also available on Baseten, and Vercel added it to its AI Gateway on launch day.
4. Coding tools: Inception’s documentation lists integrations, including Cursor.

Caveats before you switch

• Benchmarks are self-reported. Wait for independent leaderboards before replacing a model you’ve already tuned prompts for.
• Not open-weight. Kuleshov confirmed on Hacker News that the weights aren’t being released, so you can’t self-host it.
• Speed-tier comparison. Inception compares Mercury 3 with fast, cheaper models (Luna, Haiku), not with flagship reasoning models. It’s built for latency-sensitive work, not to replace the largest models.
• Launch pricing is temporary. Budget on the $0.20/$0.70 list price.

The bigger picture

Mercury 3 joins a wider push to make AI faster and cheaper to run rather than just bigger, alongside custom inference chips and software like Amazon’s TorchNeuron for Trainium. Inception also sells Mercury Voice, tuned for voice agents, and Mercury Decide, a classification model. The company noted that Mercury 2 was named to TIME’s Best Inventions of 2026 earlier in October. If independent tests confirm Inception’s numbers, Mercury 3 would be the strongest evidence yet that diffusion is a real alternative to the one-token-at-a-time design behind nearly every chatbot. Until then, it’s an impressive and inexpensive model that’s worth testing for jobs where speed matters most.

About Teck Hustlers Editorial Team

The Teck Hustlers Editorial Team covers technology, business, finance and gaming news for readers who want the story behind the headline. Every article is researched, written and edited by our in-house team, checked against company announcements, filings and reputable reporting, and updated when new facts emerge. We focus on what each development means for founders, professionals, investors and everyday users. Have a tip or a correction? Reach us through our contact page.