Inception's Mercury 3 Writes at 1,068 Tokens a Second

Every chatbot you have ever used writes like a nervous typist: one word at a time, each word chained to the one before it, unable to peek ahead. Yesterday, a small Redwood City startup shipped a model that works the opposite way. Mercury 3, from Inception Labs, drafts an entire answer at once and then edits it into shape — and the company says it does this at 1,068 tokens per second.

If that number survives contact with independent testing, a thirty-year-old assumption quietly dies: that smarter models must necessarily be slower ones. Inception's claim is not merely that Mercury 3 is fast. It is that the speed now comes free — that you no longer have to choose between a model quick enough for a voice call and one clever enough to be useful.

I have read the launch announcement, Inception's benchmark table, and the first independent hands-on test. Here is what is genuinely new, what is marketing, and why the distinction matters less than you might think.

Photo: 极客湾Geekerwan via Wikimedia Commons (CC BY 3.0). Inception reports Mercury 3 generating 1,068 tokens per second on NVIDIA GPUs; the company has not disclosed the exact GPU model or serving setup behind the figure.

The numbers that matter

Mercury 3 is Inception's fourth model, and each release has followed the same script: more intelligence, same absurd speed. Mercury 2, launched in February, was the first diffusion model with reasoning. Mercury 2.5, out on September 8, added tunable reasoning and a 260,000-token context window. Mercury 3, announced October 19 — 41 days later — is, by the company's account, the largest intelligence jump of the four, with the context window quadrupled to one million tokens.

The headline figures, all from Inception: generation above 1,000 tokens per second (1,068 on the spec sheet) on NVIDIA GPUs, four to eight times faster than the speed-optimised rivals it is being compared against. Total parameters more than forty times those of the original Mercury 1, while throughput stayed roughly flat. Pricing is aggressive: a list rate of $0.20 per million input tokens and $0.70 per million output tokens, halved during the launch window to $0.10 and $0.35, with cached input tokens at a 90% discount. New accounts get 100 million free tokens to play with.

Availability is broad for a startup launch. Mercury 3 is live through Inception's own OpenAI-compatible API — meaning developers can point existing code at it without rewriting anything — plus Baseten, OpenRouter and Vercel, with a consumer chat interface for the merely curious. Enterprise buyers are offered dedicated capacity, autoscaling, compliance controls and configurable data retention, which tells you who Inception thinks the real customer is: not hobbyists, but companies running agents at scale.

That audience choice is deliberate. In agentic systems, where a model makes dozens of small calls to finish one task, a slow model doesn't just feel sluggish — the delays compound at every step, and the agent burns time and money waiting on itself. A model that answers in a fraction of a second changes what those systems can attempt. Inception knows it: early customers are already using Mercury 3 for browser navigation, tutoring, search and support, and the company also ships two siblings, Mercury Voice for voice agents and Mercury Decide for high-volume structured decisions.

Diffusion, in plain English

Here is the idea in one paragraph. A normal language model — every GPT, Claude and Gemini release you have ever used — is autoregressive: it predicts the next word, then the next, then the next, each step waiting on the last. It is like typing with one finger. A diffusion model instead starts with a rough, noisy draft of the whole answer and refines every part of it in parallel, over a handful of passes — closer to how a sculptor works a block of marble from all sides at once. The technique is borrowed from image generation, the same coarse-to-fine process that turns static into a photograph in tools like Midjourney.

The parallel part is why it is fast: the model is never stuck waiting for word five before it can start word six. The refinement part is why Inception thinks it can be accurate: unlike an autoregressive model, which can never take back a word once emitted, a diffusion model is trained to correct its own mistakes mid-generation. Every pass is a chance to fix something.

The question that has dogged this approach since Inception emerged from stealth in 2025 is whether it scales — whether the speed advantage survives as models get bigger and tasks get harder, or whether diffusion is doomed to be the clever trick that works on small models and falls apart on serious ones. "Since we launched Mercury 1, the question we've heard most is whether diffusion can scale. Mercury 3 is over 40 times larger, significantly more capable, and still runs at over 1,000 tokens per second," said Stefano Ermon, Inception's co-founder, in the company's launch announcement. Ermon, who grew up in a small village in the Italian Alps before co-inventing the score-based diffusion methods behind modern image generators, added that the team has already started training its next model.

The benchmark claims — and the fine print

Now the part that deserves scepticism. On Inception's own benchmark table, Mercury 3 beats OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5 on 8 of 12 tests. Its best showing is SWE-Bench Atlas at 51.08, roughly 18 points clear of both rivals; on BrowseComp+ it reports 55.18 against Luna's 50.50 and Haiku 5.5's 45.78. Its worst is DeepSWE v1, where it trails Haiku 5.5, 46.31 to 51.62.

Read that paragraph again and notice what is missing: independent verification, disclosed hardware, and any explanation of what "medium settings" — the condition under which the table was measured, per one launch-day analysis — actually means. This is Inception grading its own homework, against two named rivals, on tests it selected. The 4–8x speed claim has the same problem: measured on Inception's serving stack, on undisclosed GPUs, under undisclosed conditions. Treat every comparative number in this section as a company claim, because that is what it is.

That said, the first independent test is encouraging rather than damning. The Neuron's reviewer ran Mercury 3 through eight real API calls on the day of launch — writing, maths, explanation, coding, hallucination resistance, logic — for a total reported cost of about four cents. The model was genuinely funny on creative writing but botched the constraints (274 words when asked for 250, and it wandered out of the one-room setting). It nailed a fiddly coffee-shop profit calculation, then wrote a complete single-file HTML dungeon crawler whose JavaScript passed a syntax check and survived 2,000 randomised map-generation tests without a single failure. It correctly refused to invent details about a fictitious 1987 NASA mission — then spoiled the pass by asserting, wrongly, that there were no other Shuttle launches in 1986–88. Fast, clever, occasionally sloppy: a fair summary of a first-day independent look.

The reviewer is honest about the limits of that exercise: eight spot checks are not a benchmark study, and nobody outside Inception has yet reproduced the 1,068-tokens-per-second figure. The API felt fast, which is not the same as verified.

Map: National Renewable Energy Laboratory / U.S. Department of Energy, 2025 (public domain). Models like Mercury 3 live or die on inference infrastructure — the GPU capacity, increasingly rented from neocloud GPU clouds, that turns a fast architecture into a fast product.

The three professors betting against the entire industry

It is worth understanding who is making this bet, because the credentials are almost comically strong. Inception was founded in 2024 by three professors: Ermon at Stanford, Aditya Grover at UCLA, and Volodymyr Kuleshov at Cornell — the same Stanford-centred research ecosystem that produced the diffusion methods now generating half the images on the internet. The company raised $50 million in November 2025, led by Menlo Ventures, with Microsoft's M12, NVIDIA's venture arm, Snowflake and Databricks participating and Andrew Ng, Andrej Karpathy and Eric Schmidt writing angel cheques. When Karpathy — who does not hand out endorsements casually — backs a diffusion startup, the industry pays attention.

The timing is pointed. Mercury 3 lands in a quarter where every major lab is racing to ship: Mistral announced a new flagship the day before, and Anthropic cut Haiku's price twelve days earlier. Speed and price have become the two levers every vendor is pulling at once, and Inception is trying to own both levers with an architecture nobody else is selling. Anthropic's Claude Sonnet 5.5 taking the coding crown last month shows how fast the frontier moves; Inception's answer is to change the racetrack, not just run faster on it.

We downloaded both images used in this article from Wikimedia Commons ourselves and read each file page to confirm the licence before embedding — the H100 photo is CC BY 3.0, the data-centre map is U.S. public domain. Small thing, but it is the kind of check that separates reporting from reshuffling press releases.

Where I land

Here is my position, stated plainly: the benchmark table is marketing, but the architecture bet is real, and the marketing matters less than the price tag suggests. Forget the 8-of-12 framing — what actually counts is that a diffusion model is now within shouting distance of the best small frontier models on coding and agentic tasks while charging ten cents per million input tokens at launch. That is not a research demo. That is a product a startup can build a voice agent on tomorrow morning, which is exactly what early customers like RevisionDojo are doing. "Mercury 3 allowed us to give every student instant feedback without giving up reasoning," said Michael Tong, the co-founder and CTO of General Learning, quoted in the announcement — and for a tutoring product, a one-second response is the whole ballgame.

The shoddy part is the fine print Inception hopes you will not read: undisclosed GPUs, "medium settings", a comparison table with no independent referee. The colossally interesting part is everything else — a 40x scale-up with no speed loss, a million-token context window, and an API you can switch to without rewriting code. If the scaling claim holds through independent evaluation, the autoregressive monopoly on serious language AI is over, and the next era of competition will be fought on efficiency, not size. If it doesn't, Mercury 3 is still the fastest cheap model on the market, which is a perfectly good business.

Either way, the typist just got company. The sculptor has entered the studio.

Your questions, answered

What is a diffusion language model?

A language model that generates text by starting with a noisy draft of the whole response and refining all of it in parallel, instead of predicting one word at a time. The approach is adapted from image-generation diffusion models. Inception's Mercury family is the first commercial line built this way.

How fast is Mercury 3, really?

Inception reports 1,068 tokens per second on NVIDIA GPUs — four to eight times faster than the speed-optimised rivals it compares itself against. Those are company-measured figures on undisclosed hardware; no independent party has reproduced the exact number yet, though hands-on testers describe the API as feeling fast.

How much does Mercury 3 cost?

List price is $0.20 per million input tokens and $0.70 per million output tokens, with a 50%-off launch promotion ($0.10/$0.35) and cached input tokens at a 90% discount. New API accounts get 100 million free tokens.

Is Mercury 3 better than GPT-6 Luna and Claude Haiku 5.5?

On Inception's own benchmark table it leads both on 8 of 12 tests, with its strongest result on the SWE-Bench Atlas coding benchmark. Those are company claims measured at "medium settings" — independent verification is still pending, and the first hands-on review found rough edges on precision tasks.

Sources: Inception's launch announcement via BusinessWire (October 19, 2026); The Neuron's independent hands-on review (October 20, 2026); Interesting Engineering; Unite.AI; MarkTechPost; Startup Fortune; TechCrunch (November 2025 funding coverage); Menlo Ventures founder profile. Benchmark comparisons are Inception's own unless stated otherwise; quote attributions are given inline.

By Daniel Rao — Daniel Rao covers frontier AI labs, model releases and the infrastructure behind them for TechPulse USA.