Mercury 3 Explained: Inception's 1,068 Tokens/Sec Diffusion LLM

Mercury 3 Explained: Inception's 1,068 Tokens/Sec Diffusion LLM

The fastest frontier-class AI model in the world doesn't write left to right. On October 19, 2026, Inception Labs launched Mercury 3, a new large language model that generates text in parallel — refining a whole draft at once instead of spitting out one word at a time. The result: a claimed 1,068 tokens per second on standard NVIDIA GPUs, roughly four to eight times faster than OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5 — while matching or beating them on quality benchmarks.

This matters because speed has always been the tax you paid for intelligence in AI. Mercury 3 is the first serious claim that you might not have to pay it anymore. Whether you're a developer picking an API model, a founder pricing an AI product, or just someone who notices ChatGPT taking ten seconds to answer, here's what Mercury 3 actually is, how it works, and where the vendor's claims hold up.

Quick answer: Mercury 3 is Inception Labs' third-generation diffusion LLM, released October 19, 2026. Instead of generating text one token at a time, it drafts and refines entire responses in parallel, hitting 1,068 tokens/second in Inception's tests — 4-8x faster than GPT-6 Luna (~130 tok/s) and Claude Haiku 5.5 (~240 tok/s) — while winning 8 of 12 benchmarks in the company's own testing. List price is $0.20 per million input tokens and $0.70 output, halved during a launch promotion. It's available now through the Inception API, OpenRouter, Baseten, Vercel, and Mercury Chat.

Table of Contents

What Is Mercury 3?

Mercury 3 is the flagship model of Inception Labs, a startup founded by Stanford AI researcher Stefano Ermon along with Volodymyr Kuleshov and Aditya Grover. The company has spent two years building the commercial case for diffusion language models (dLLMs) — a fundamentally different architecture from what OpenAI, Anthropic, and Google use, as detailed in MarkTechPost's technical breakdown.

The backstory matters here. When Inception launched Mercury 1, diffusion LLMs were mostly an academic curiosity, good for simple jobs like code autocomplete. The open question the whole research community asked was whether diffusion could scale — whether it would keep getting smarter as models got bigger, the way conventional models do. Mercury 3 is Inception's answer: the company says the new model is more than 40 times larger than Mercury 1 in total parameters, with intelligence up roughly 3x across its benchmark suite, and throughput essentially unchanged. As CEO Stefano Ermon put it, the scaling laws that drove progress in autoregressive models appear to hold for diffusion too. No independent group has reproduced the results yet, but Inception has already started training the successor model.

Key specs at a glance:

  • Release date: October 19, 2026
  • Architecture: Diffusion LLM, closed weights (API access only)
  • Speed: 1,068 tokens/second on NVIDIA GPUs (vendor-measured)
  • Context window: 1 million tokens, up from 260K on Mercury 2.5
  • Max output: 131,072 tokens per response
  • Reasoning effort: four levels — instant, low, medium, high (defaults to "low")
  • Capabilities: tool calling, structured outputs, OpenAI-compatible chat API
  • Input/output: text only, no images or audio

How Does a Diffusion LLM Actually Work?

Every mainstream AI model you've used — ChatGPT, Claude, Gemini — is autoregressive. It writes the way you type: one token at a time, each new word conditioned on everything before it. That's reliable, but it's inherently slow. The model literally cannot start writing word six until it has finished word five. A long answer takes a long time no matter how powerful the hardware.

A diffusion LLM flips the process. Instead of writing left to right, it starts with a rough, noisy draft of the entire response and refines every position in parallel over a few passes. If that sounds familiar, it's because it's the same coarse-to-fine process that turns static into a photo in image generators like Midjourney — and, for that matter, in AI video tools like Kling 4.0, which uses diffusion to generate video frames. Diffusion was always great at parallel generation; the question was whether text diffusion could ever match autoregressive quality on complex reasoning. Mercury 3 is Inception's strongest evidence yet that it can.

One detail worth knowing: Mercury 3 offers four reasoning effort settings (instant, low, medium, high) and defaults to low when you don't specify one — unlike Mercury 2.5, which defaulted to medium. That means out of the box it will look a bit faster (and score a bit lower) than the same model run at higher effort. If you benchmark it yourself, pin the effort level explicitly, or your numbers won't be comparable to anyone else's.

How Fast Is Mercury 3, Really?

The headline number — 1,068 tokens per second — comes from Inception's own measurements on NVIDIA GPUs. In the same comparison, the company clocked Claude Haiku 5.5 at about 240 tokens/second and GPT-6 Luna at about 130. That gap is the whole strategic argument: for agentic workloads, where a coding or browser agent might call the model dozens of times per task, latency compounds at every step. A 4x faster model doesn't just save time — it changes what you're willing to attempt. More iterations, more alternative approaches, more verification passes.

But vendor numbers deserve skepticism, and here we actually have some early independent data:

  • OpenRouter's live measurement lists Mercury 3 at a P50 throughput of just 221 tokens/second on its shared infrastructure — far below 1,068. Deployment conditions matter enormously.
  • An independent tester measured roughly 850–900 visible tokens/second from the client side, over the network — close to the claim, once you include streaming and network overhead. On what you actually wait for, they found about 1.2–1.5x faster per coding answer at matched quality, and 1.9–2.9x per agent task.
  • Octen AI reported a median generation time of 0.9 seconds and median end-to-end response of 2.2 seconds for real-time search answers.
  • TinyFish AI tested Mercury 3 as a browser agent on 50 live web-navigation tasks and reported it completed 50% — significantly more than GPT-6.1-Sol at the same token volume.

So: is it a speed demon? The honest answer is that Mercury 3 is clearly faster than the big labs' fast models, roughly as accurate as them with reasoning off, and especially strong inside agent loops — a conclusion an independent hands-on test reached as well. What you measure depends heavily on where and how it's served.

Mercury 3 Benchmarks vs GPT-6 Luna vs Claude Haiku 5.5

Inception's published benchmark table (all models at medium effort) has Mercury 3 winning 8 of 12 tests against OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5 — the two models that define the "fast and cheap" tier right now. For context, GPT-6 with its Intelligent UI rollout is the model most people are comparing everything against this month. Here's the scoreboard, with the honest caveat that every number below comes from the vendor — SiliconANGLE's launch coverage and The Decoder's skeptical read both note no independent group has reproduced them yet:

Benchmark Mercury 3 GPT-6 Luna Claude Haiku 5.5
IFBench (instruction following)76.84~61.8—
SWE-Bench Atlas (agentic coding)51.0833.3332.26
GPQA Diamond (reasoning)89.23——
DeepSWE v1 (long coding tasks)46.31—51.62
MRCR v2 (long-context retrieval)75.50—78.40
AA-Omniscience46.4446.94—

The pattern is telling: Mercury 3's biggest wins are in instruction following and agentic coding — exactly the workloads where speed multiplies. Its losses are in long-context retrieval and longer, multi-step coding tasks — where Haiku 5.5 still leads. Early-access partner Bito ran its own coding benchmark and found a similar shape: Mercury 3 scored 65% overall on long agentic coding sessions at $0.48 per session, versus 76% for Haiku 5.5 at $0.69 and 87% for Opus 5.5 at $5.42 — but Mercury finished sessions in 5.8 minutes versus 17.2 for Opus. Fast and cheap, slightly less careful. That tradeoff is the whole story of this model.

Mercury 3 Pricing: What Does It Cost?

Inception priced Mercury 3 to undercut the fast-model tier. At list price it's already cheaper than Haiku 5.5's $0.10/$0.50 was at its own launch, and during the launch window it's genuinely inexpensive:

Token type List price Launch price (50% off)
Input$0.20 / 1M$0.10 / 1M
Output$0.70 / 1M$0.35 / 1M
Cached input (90% off input rate)$0.02 / 1M$0.01 / 1M

Practical notes: new accounts on the Inception API platform get 100 million free tokens — enough to properly test it before spending anything. And because agent loops repeat similar context over and over, that 90%-off cached-input tier matters more than it looks: agentic workloads are where Mercury 3 is designed to live. If you're shopping the whole fast-model market, our DeepSeek V4.1 Flash breakdown covers the open-source side of the speed race for comparison.

How to Try Mercury 3 (Step by Step)

  1. Sign up on the Inception API platform — new accounts receive 100 million free tokens, which is plenty for real testing.
  2. Grab your API key from the dashboard and note the model ID: mercury-3.
  3. Point your existing code at it. The API is OpenAI-compatible, so if you already call GPT-6 or Haiku via a chat-completions endpoint, you mostly just swap the base URL and model name.
  4. Set the reasoning effort explicitly (instant, low, medium, or high) — remember it defaults to low, so comparisons need this pinned.
  5. Or skip the API entirely: the model is live on OpenRouter as inception/mercury-3, plus Baseten and Vercel, and there's a chat interface at Mercury Chat for non-developers.

Who Is Mercury 3 For?

Mercury 3 is not trying to be the smartest model in the world — Opus-class models still win on raw reasoning. It's trying to be the smartest model at a given speed. That makes it a natural fit for:

  • AI agents and coding assistants — browser-use agents, computer-use agents, and dev tools where every extra second of latency compounds across dozens of calls. This is also why the launch is interesting alongside Anthropic's Claude Duet agent push and the rebuilt Notion AI agents: the agent era needs models that are fast and capable, and Mercury 3 is built for exactly that slot.
  • Real-time voice, tutoring, and support bots — anything where a human is waiting. Sub-second responses change what these products feel like; for a deeper look at the voice side, see our guide to free AI voice cloning tools.
  • Long-document processing — the full 1M-token context window at budget pricing makes it attractive for chewing through books, codebases, and legal documents.
  • Creators and indie builders — if you're generating video scripts or iterating on content with AI, faster models mean faster drafts. And when the draft is done, grabbing a thumbnail is instant with our free YouTube Thumbnail Downloader tool — worth a look if you make videos like our faceless YouTube video guide covers.

Mercury 3 Pros and Cons — Honestly

Pros:

  • Genuinely the fastest frontier-class model available — no one else is near 1,000 tokens/second at this quality level.
  • Launch pricing is aggressively cheap ($0.10/$0.35 per million tokens), and 100M free tokens make testing risk-free.
  • 1M-token context window — quadruple Mercury 2.5's.
  • OpenAI-compatible API with tool calling and structured outputs — drop-in for existing code.
  • Early evidence (Bito, TinyFish, Octen) suggests the speed advantage survives real-world use, especially in agent loops.

Cons:

  • All quality claims are vendor-reported. No independent group has published full benchmark reproductions yet.
  • The 1,068 tok/s figure is a best-case server-side number; OpenRouter's shared tier measures 221 tok/s. Your mileage will vary.
  • Trails Haiku 5.5 on long-context retrieval and long multi-step coding — speed doesn't fix every weakness.
  • Text only. No images, no audio, no video understanding.
  • Closed weights — you can't download or self-host it. (If open weights matter, Google's DiffusionGemma is the diffusion model to watch instead.)
  • Launch pricing has no published end date. Budget for the list price, not the promo.

What Mercury 3 Means for the AI Race

Step back from the benchmark table and the interesting claim isn't really "Mercury 3 beats Luna." It's Inception's bet that diffusion scales like everything else — that the whole industry's decade of scaling wisdom transfers to a parallel architecture. If true, this is the first crack in the autoregressive monopoly, and it won't be the last: Google is already exploring diffusion text models, and Inception says Mercury 3's successor is in training now.

For the rest of us, the practical effect is simpler: the price-performance frontier just moved. Fast, cheap, capable models are how AI gets embedded into everything — every agent loop, every support chat, every coding assistant. Mercury 3 doesn't have to dethrone the giants to matter. It just has to make "fast enough and smart enough" a viable default — and at 1,000 tokens a second, it might.

Frequently Asked Questions

What is Mercury 3?

Mercury 3 is Inception Labs' third-generation diffusion LLM, launched October 19, 2026. It's a large language model that generates text in parallel rather than one token at a time, which makes it dramatically faster than conventional models while matching their quality on most benchmarks.

How does a diffusion LLM work?

Instead of writing left to right like ChatGPT or Claude, a diffusion LLM starts with a rough draft of the entire response and refines every part of it in parallel over a few passes — the same coarse-to-fine idea behind AI image generators. Fewer sequential steps means much faster output on the same hardware.

How fast is Mercury 3 compared to GPT-6 Luna and Claude Haiku 5.5?

Inception measures 1,068 tokens/second for Mercury 3, versus about 240 for Claude Haiku 5.5 and 130 for GPT-6 Luna — a 4–8x gap. Independent tests confirm a large speed advantage, though smaller in real-world conditions: roughly 850–900 tokens/second client-side, and 1.9–2.9x faster per agent task.

How much does Mercury 3 cost?

List price is $0.20 per million input tokens and $0.70 per million output tokens, with cached input at 90% off. During the launch window it's 50% off: $0.10 input and $0.35 output per million tokens. New API accounts get 100 million free tokens.

Is Mercury 3 open source?

No. Mercury 3 is a proprietary model available only through APIs — Inception's own platform, OpenRouter, Baseten, Vercel, and Mercury Chat. The parameter count hasn't been disclosed. If you need open weights, Google's DiffusionGemma is the open alternative to explore.

Who should use Mercury 3?

Developers building AI agents, coding assistants, real-time chat and voice apps, and anyone processing very long documents. If raw reasoning quality matters more than speed, larger models like Opus 5.5 still win — Mercury 3 is for when speed and cost matter as much as smarts.

Our Verdict

Mercury 3 is the most interesting model launch of the month — not because it wins every benchmark (it doesn't), but because it changes the shape of the tradeoff everyone else is competing on. Take the vendor benchmarks with the appropriate grain of salt and wait for independent reproductions, but the core claim — diffusion scales, and it's fast — already has enough third-party support to take seriously. The 100 million free tokens mean there's no reason not to test it yourself. That's the real verdict: try it in an agent loop and watch the clock. You'll feel the difference.

Methodology: This article was researched on October 23, 2026, from Inception Labs' official launch post, the company's Business Wire announcement, and independent reporting from SiliconANGLE, The Decoder, MarkTechPost, TipRanks, and an independent hands-on benchmark published on Medium. Benchmark figures are vendor-reported unless noted; speed claims were cross-checked against OpenRouter's live throughput data and third-party measurements. No sponsorship, affiliate deal, or payment influenced this coverage. Last updated: October 23, 2026. — The AI Tools · About us · Contact

Comments

Post a Comment