Inception Labs released Mercury 3 on October 19, 2026, and this Mercury 3 diffusion LLM is built for one thing: speed. The company says it generates over 1,000 tokens per second and handles a 1M-token context window, at $0.20 per million input tokens.
That makes it a different kind of model from the chatbots most people use. It is aimed at developers who need fast, cheap answers in voice apps, agents, and high-volume tasks.
Here are the quick facts, all from Inception’s own launch post.
- Released October 19, 2026, via the Inception API, OpenRouter, Baseten, and Vercel
- Speed: 1,068 tokens per second on NVIDIA GPUs
- Context window: 1M tokens, up from 260K in Mercury 2.5
- Price: $0.20 input / $0.70 output per million tokens, half off at launch
- OpenAI-compatible API, with 100M free tokens to test it
If you only build with chat apps, you can skim the rest. If you pay for API calls, the price and speed are the interesting part.
What makes a diffusion language model different?
Most LLMs, like GPT or Claude, write one token at a time, left to right. Each word has to wait for the one before it. That is why long answers stream slowly.
A diffusion language model works more like image generators do. It starts with a rough draft of the whole answer and refines many tokens in parallel. Inception laid out the idea in its original Mercury paper, and The Decoder covered the first release.
Parallel refinement is where the speed comes from. The trade-off is that the approach is newer and less proven than standard models.
How fast and cheap is Mercury 3, really?
Inception claims 4 to 8 times the speed of comparable speed-tuned frontier models. Its partner Octen AI reports a median generation time of 0.9 seconds and a median end-to-end response of 2.2 seconds on its search API.
Pricing is where it gets tempting. Here is the full rate card.
| Item | Standard price | Launch price |
|---|---|---|
| Input tokens (per 1M) | $0.20 | $0.10 |
| Output tokens (per 1M) | $0.70 | $0.35 |
| Cached input tokens | 90% cheaper than input | 90% cheaper than input |
The launch discount is 50%. Inception has not said how long it lasts, so check before you budget around it. For comparison, our look at DeepSeek V4.1 Flash pricing shows how crowded the cheap-model tier already is.
Is Mercury 3 actually smart enough?
Inception says Mercury 3 matches or beats GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks. It also says the model is about 3 times smarter than Mercury 1 on standard tests, with over 40 times the parameters.
Treat those as vendor claims. The benchmark chart in the post is an image, and no independent tests are out yet.
One partner result shows the limits. On Mind2Web, a set of 50 live web navigation tasks run by TinyFish, Mercury 3 completed 50% of them. The partner says that beats GPT-6.1 Sol, but half the tasks still failed.
Which Mercury 3 model should you use?
Inception ships three products, each aimed at a different job.
| Model | Built for | Best fit |
|---|---|---|
| Mercury 3 | General reasoning and coding | Agents, chat, long documents |
| Mercury Voice | Low time to first answer token | Voice assistants and call bots |
| Mercury Decide | High-volume classification | Tagging, routing, moderation |
Voice is the obvious win. If you read about Grok Voice 3.0’s price and speed, you know delay is what makes a voice bot feel robotic.
How do you try the Mercury 3 diffusion LLM?
The API is OpenAI-compatible, so most existing code only needs a new base URL, API key, and model name. You also get 100M free tokens on the Inception platform.
Run your own test before switching anything important. Inception itself says to evaluate on your workload and serving constraints, and that is good advice for any new model.
What should you check before you commit?
Speed numbers depend on the hardware and the load. A figure measured on NVIDIA GPUs in a lab may look different when thousands of users hit the same endpoint at once.
Start with a small slice of real traffic. Compare answer quality, cost per request, and response time against what you use today. Mercury 3 supports prompt caching, so repeated long prompts could cost far less than the headline rate suggests.
Also read the data terms. Inception lists configurable data retention for enterprise plans, which matters if you send customer text through the API.
Frequently Asked Questions
What is Mercury 3?
It is a diffusion-based large language model from Inception Labs, released October 19, 2026. It is built for very fast output and long context.
How many tokens per second does Mercury 3 generate?
Inception reports 1,068 tokens per second on NVIDIA GPUs. That figure comes from the company, not an outside test.
How much does the Mercury 3 API cost?
It costs $0.20 per million input tokens and $0.70 per million output tokens. A launch discount cuts that to $0.10 and $0.35.
Where can I use Mercury 3?
You can use it through the Inception API Platform, OpenRouter, Baseten, Vercel, and the Mercury Chat app.
Is it worth switching?
If your app waits on slow answers, such as voice, autocomplete, or classification at scale, Mercury 3 is worth a weekend test, especially at launch prices. For hard reasoning or tasks where a wrong answer costs real money, stay with your current model until independent benchmarks show up. Fast is only useful when it is also right.
Leave a Reply