Mercury 3 Is INSANE: I Tested 13 Live Examples (+ 100 MILLION FREE Tokens)
Read the Portuguese translation →
Author: Prompt Engineer 48 · Watch on YouTube
Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.
Original language: English
Original video: Mercury 3 Is INSANE: I Tested 13 Live Examples (+ 100 MILLION FREE Tokens) · Channel: Prompt Engineer 48 · 09:50 · Watch on YouTube
Topics: AI evaluation and benchmarks, Software engineering with AI, AI developer tools, AI models
Summary
Section titled “Summary”The video tests Mercury 3, a diffusion language model, across live coding, writing, reasoning, translation, and classification examples. The presenter reports that it handles many tasks quickly, but timings vary considerably with task complexity and the visual diffusion effect slows responses. In API tests, end-to-end speeds were about 260–420 tokens per second, below the vendor’s headline claim of over 1,000 tokens per second. The presenter recommends trying the free tokens while treating marketing claims cautiously and distinguishing tested features from unverified ones.
Key points
Section titled “Key points”- Mercury 3 is described as a diffusion model that drafts and refines an answer in parallel rather than generating it strictly one token at a time.
- The presenter says new accounts receive 100 million free tokens without payment details, and that the API uses an OpenAI-style endpoint.
- The snake game worked in the preview, while the model also produced a long blog post, solved a logic puzzle correctly, and generated code with tests the presenter did not run.
- The diffusion-effect toggle was slower in the presenter’s repeated test: about 3.5 seconds on versus 1.5–2 seconds off.
- More complex interactive tools took longer and varied across runs, including a Kanban board and a Game of Life simulation.
- The model translated into Spanish, Hindi, and Japanese, though the presenter says native speakers should judge the translations.
- In a support-triage classification test, the page reported a decision time of 186 milliseconds.
- Three API runs measured approximately 258, 416, and 353 tokens per second end-to-end; the presenter attributes the gap from the headline figure partly to hidden reasoning tokens and bursty output.
- Voice mode and the claimed one-million-token context window were not tested.
Chapters
Section titled “Chapters”| Start | Chapter | Summary |
|---|---|---|
| 00:00 | Mercury 3 and the test setup | The presenter introduces Mercury 3, explains diffusion generation, and outlines the free-token offer and API access. |
| 02:06 | Live tests: writing, reasoning, and code | The video compares the visual toggle, then tests a long blog post, a logic puzzle, code refactoring, and a stock-price query. |
| 04:08 | Interactive tools, translation, and classification | The presenter tests interactive web apps, translations, plain-language writing, and Mercury 3’s fast support-triage decision mode. |
| 08:14 | API speed measurements and verdict | Streaming API runs yield lower end-to-end speeds than the headline claim, informing the presenter’s cautious final recommendation. |
Quotes
Section titled “Quotes”Mercury starts with a rough draft of the whole answer and refines all of it at once — 00:31
the toggle is a visual effect and it actually costs you a second or two. — 02:06
End-to-end, I saw roughly 260 to 420. — 08:44
Do not take the marketing number as a promise for your own app. — 09:14
Concepts and entities
Section titled “Concepts and entities”Mercury 3, Inception Labs, diffusion language model, GPT-6, Luna, Claude Haiku 5.5, RunPod
Full transcript (original language)
Section titled “Full transcript (original language)”00:00 Watch this. I type one prompt, hit enter, and a whole snake game shows up in a few seconds. Code on the left, a working preview on the right. This is Mercury 3 from Inception Labs, a diffusion language model. >> [music] >> The blog says it runs at over 1,000 tokens per second, and you get 100 million free tokens to try it. So, today I’m not reading you a press release. I am throwing real prompts at it live, and I will tell you exactly what held up and
00:31 what did not. Let’s go. First, what is a diffusion model doing differently? Normal language models write one token at a time, left to right. Mercury starts with a rough draft of the whole answer and refines all of it at once, a few passes, like a photo coming into focus. The launch post, published October 19th, says Mercury 3 beats GPT-6, Luna, and Claude Haiku 5.5, while running 4 to 8 times faster. Remember, that is the
01:01 vendor’s claim. Everything after this point is what I saw on my own screen. Now, the free part. The docs say every new account gets 100 million free tokens, and no payment details are needed. You sign in on the platform, open API keys, create a key, and you’re in. The model name is Mercury 3, and the endpoint is OpenAI style. So, most existing clients just work. After the free tokens, the blog lists 20 cents per million input tokens and 70 cents per million output, with a launch promo that
01:32 cuts that in half. Okay, enough setup chat time. Example one, the snake game. Prompt: Write a complete snake game in one HTML file. Watch the left side. The text is blurry for a moment, then it sharpens. That is the diffusion effect, 4 to 5 seconds in. The file is done, and the game is running. And yes, I played it. It starts, it moves, it scores, and it ends with game over when I hit the wall.
02:06 Example two, there is a diffusion effect toggle in the corner. It changes how the text appears. So, I wanted to know if it changes the speed. Same prompt. Explain how a transformer works in 300 words, once with the effect off, once on. I ran each three times. With the effect off, it finished in about 1 and 1/2 to 2 seconds. With it on, 3 and 1/2 every single time. So, the toggle is a visual effect and it actually costs you a second or two. Pretty to watch, but turn it off if you want the fastest answer.
02:36 Example three, a long one. I asked for a thousand word blog post with headings and a conclusion. About 6 and 1/2 seconds later, the whole post was sitting there. Long outputs are exactly where a normal model makes you wait. So, this one felt different. Example four, a logic puzzle with a known answer. Alice is twice as old as Bob was when Alice was as old as Bob is now, and their ages add up to 63. It set up the equations, solved them in under 3 seconds, and said, “Alice is 36 and Bob
03:07 is 27.” That is correct. It even checked itself. 9 years ago, Alice was 27 and Bob was 18, and 36 is twice 18. Good. Example five, real code work. I passed it a messy function with a loop and an index, and asked it to refactor it and add pedantic tests. 4 seconds. It gave back a clean version, plus a full test file, empty list, negatives, zero, duplicates, order preserved, even a generator input. I have not run those tests myself, so I’m calling this
03:38 promising, not proven. Example six, a trap. I asked for the exact closing price of Nvidia stock yesterday. A plain model would guess. This one searched 12 sources, gave a figure in about 5 seconds, and pointed out that one site disagreed by 8 cents. I did not verify that number, so do not trade on my video, but flagging a disagreement is the right behavior. Example seven, Conway’s Game of Life
04:08 with a canvas, play, pause, and step buttons. This one took 10 seconds on my first run, and about 23 on the second. Same prompt, different load, so treat any single number with care.
04:40 Example eight, the stress test, a Kanban board with drag and drop, add card, delete card, saved in local storage. 44 seconds on the first run, 28 on the second. The model spends a lot of that time thinking before it writes. When it finished, there was a three-column board with seated cards. So, the speed is not constant. Harder prompts take longer.
05:15 Example nine, translation, Spanish, Hindi, and Japanese in under 3 seconds. All three came out fluent to me on the first read, but I would want native speakers to judge them. Example 10, a futuristic landing page for a made-up product, Mercury Cloud Hero, three feature cards, pricing about 13 seconds, and look at the preview, a headline, a call to action, even a little terminal card with a deploy log. I did not ask for the terminal, it just added one.
06:03 >> Example 11, a regex playground with live match highlighting. This one took about 27 seconds. When it finished, there is a pattern box, preset buttons for email, URL, date, and more, a sample text, and flag toggles. A real little tool from one sentence.
06:38 Example 12 is sorting visualizer bubble, merge, and quick sort with a speed slider. About 21 seconds, and now it is sorting live. Comparisons, swaps, elapsed time, all updating. The colored bars are the algorithm running.
07:11 And one quick plain text test, explain special relativity to a 10-year-old in five short paragraphs. Under 4 seconds, not code, not a tool, just clear writing. Now something different. Inside the chat side, there is a decide mode. The idea is instead of writing text, the model makes fast classification decisions. The preset is support triage with a customer who was charged twice and is angry. I hit decide. It answered four questions
07:42 at once, want a refund, yes, team billing agency, a little over two and a half out of three, churn risk, yes. And the page says it decided in 186 milliseconds. The page estimates about 1 cent per thousand calls and claims 80% lower input price than GPT-6 Luna. Those cost figures are the pages, not mine, but the speed, that part I watched happen. There is also a voice mode and a search playground. I did not test voice for this video and I will not pretend I did. The 1 million token context window is
08:14 also untested, so I will not claim anything about it. Now, the part I can verify properly, I took the API key and called Mercury 3 myself with streaming. Do not worry, I’m not showing the key and I will rotate it after this video. Run one, a 250 word explainer at medium reasoning. First token after 2.8 seconds, total 3.1 seconds. 806 tokens, that is about 258 tokens per second end-to-end. Run two, 600 words at low reasoning effort. 1,968
08:44 tokens in 4.7 seconds, about 416 per second. Run three, same prompt at medium, 3 seconds total, about 353 per second. So, where is the thousand tokens per second? Two things. About a third to a half of those completion tokens were hidden reasoning tokens and the text arrives in bursts. If you only time the stretch after the first token, the numbers look huge. End-to-end, I saw roughly 260 to 420. That is still fast, but it is not what the headline says. These were single runs from my network, so your numbers will vary. The takeaway
09:14 is simple, quick answers, tools, classification, coding help, it feels instant. Do not take the marketing number as a promise for your own app. So, here’s my verdict. What I liked, the speed on long outputs, the correct puzzle, solid code and tests, and that it flagged its own disagreement on the stock price. What I would watch, the headline number, the hidden reasoning tokens, and everything I did not test. If you want to try it, the free tokens are the whole point. Make an account, grab a key, and run your own prompts. Do not take my word or the blog’s.
09:50 If you are looking for cheap and reliable GPUs, I use RunPod, and the link is in the description. Tell me in the comments, what is the first thing you would throw at Mercury 3? I will see you in the next one.