Mercury 3 Lands at 1,068 Tokens a Second — and Without a Single Benchmark Number

Inception Labs released Mercury 3 on October 19, 2026, and the announcement leads with three numbers: 1,068 tokens per second on NVIDIA GPUs, a 1M-token context window, and $0.10 per million input tokens for as long as the launch discount lasts. What it does not contain is a single benchmark score. The quality claim is comparative — Mercury 3 "outperforms GPT-6 Luna and Claude Haiku 5.5" — and every quantitative claim in the post is about speed, size or price. That is not a criticism so much as the shape of the evidence you have to work with if you are deciding whether to move a workload onto it this week. This piece walks through what Inception actually shipped, what the unscored claims mean, and where a builder should draw the line between trying it and depending on it. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

What Inception put on the table on October 19

Mercury 3 is a diffusion language model — Inception's term is dLLM — and the company calls it the most capable diffusion-based model released to date. The architecture generates tokens in parallel rather than one at a time, which is where the throughput claim comes from. Inception says the model is more than 40 times larger than Mercury 1 in total parameters, without publishing an absolute parameter count, and that measured intelligence across its benchmark suite has risen roughly threefold from Mercury 1 (mini) to Mercury 3.

• Context — 1M tokens, up from 260K on Mercury 2.5, with a maximum output of 131,072 tokens.

• Price — $0.20 per million input and $0.70 per million output at list; $0.10 and $0.35 while the launch discount is running.

• Cached input — $0.02 per million at list, halved to $0.01 during the launch window, which is the customary 90% discount on the input rate.

• Reasoning effort — four settings: instant, low, medium and high. Mercury 3 defaults to low when the parameter is omitted, where Mercury 2.5 and Mercury 2 default to medium.

• Modalities — text in, text out. Tool calling and structured outputs are supported on the chat completions endpoint.

• API shape — OpenAI-compatible chat completions, with temperature adjustable between 0.5 and 1.0.

• Availability — Inception's own API platform, where new accounts get 100 million free tokens, and several third-party inference platforms. Enterprise deployments add dedicated capacity, autoscaling and configurable data retention.

The default effort setting is the detail most likely to be missed and most likely to matter. A model that answers at "low" effort by default will look faster and score lower than the same model at "high", and Inception's own comparison claims do not state which setting produced them. If you benchmark Mercury 3 yourself, pin the effort level explicitly or your numbers will not be comparable to anyone else's.

The quality claim is comparative, not quantitative

Read the sentence again: Mercury 3 "outperforms GPT-6 Luna and Claude Haiku 5.5". Both are speed-optimised models, not frontier flagships, and the comparison is the correct one for a model whose entire pitch is latency. But the announcement gives no margin, no benchmark names, and no per-task results, so the claim cannot be checked against a public harness without running it yourself.

The one quality result with any structure behind it is an agent test. Inception ran Mercury 3 inside its own production web-agent harness against frontier models on 50 live Mind2Web navigation tasks, and reports that it completed 50% of them, "significantly more than GPT-6.1-Sol", using the same token volume and the same number of agent steps and therefore a fraction of the per-token cost. Fifty tasks is a small sample, the harness is Inception's own, and the comparison model is named without a score. It is a real signal — task completion on live sites is harder to game than a multiple-choice suite — but it is a signal, not a measurement you can put in a procurement document.

Inception also says Mercury 3 is "on par or better than frontier speed-optimised models" rather than on par with frontier reasoning models. That distinction is doing a lot of work in the post, and it is worth taking at face value: this is a model positioned against the fast tier, not against Claude Opus 5.5 or Gemini 4 Argon.

The speed number, set against the last Mercury

1,068 tokens per second is the headline, and it is the figure most likely to be quoted out of context. Inception attributes it to "NVIDIA GPUs" without naming the hardware, the batch size, the effort setting or whether it is a single-stream or aggregate measurement. All four change the number by multiples.

What independent measurement exists is for the previous generation. Artificial Analysis measures Mercury 2.5 at roughly 706 tokens per second and Mercury 2 in the 680s, both under its own harness. Those are real, reproducible third-party figures, and they are lower than the vendor numbers that accompanied those launches. That gap is the normal vendor-to-independent discount in this category, not evidence of bad faith — but it is the discount you should mentally apply to 1,068 until someone reproduces it.

As of October 20, 2026, Artificial Analysis has no Mercury 3 page: its model index still lists Mercury 2.5 and Mercury 2, and the Mercury 3 URL returns nothing. So the speed claim has no independent check yet, one day after launch. That will change; the useful thing to watch is not whether the number lands near 1,068 but whether the gap to Mercury 2.5 stays as wide under a neutral harness as Inception's framing implies.

What the launch discount is worth, and what it is not

The 50% cut is stated as temporary — "at launch" — with no end date. That makes it a promotional rate rather than a rate card, and any cost model built on $0.10/$0.35 has an expiry date you cannot see. The list price is the number to plan against.

A concrete comparison makes the scale clear. Take a support or search workload of 10 million input tokens and 40 million output tokens per day. At list price, Mercury 3 costs $2 for the input and $28 for the output, or $30 a day. At the launch rate the same traffic costs $1 and $14, or $15 a day. Over a month that is roughly $900 against $450 — the promotion is worth about as much as the model's entire list price advantage over a mid-tier autoregressive model. Useful, but not a reason to architect around it.

The cached-input line is the one that changes real economics for agent workloads, because agents resend the same system prompt and tool definitions on every step. At $0.02 per million cached tokens, a 20,000-token preamble costs $0.0004 per call instead of $0.004. On a fifty-step agent run that is two cents of preamble rather than twenty, which is the difference between a viable per-task price and a marginal one.

Inception is not selling a chatbot

The launch is framed around a portfolio rather than a single model, and that framing tells you who the product is for. Mercury Voice is tuned for time-to-first-answer-token in voice agents; Mercury Decide is a classification and scoring model on a separate endpoint, priced at two cents per million input and nothing per output. Inception's customer quotes are all latency-sensitive production systems: a web-agent harness at TinyFish, tutoring feedback at RevisionDojo, search generation at Octen, support at Lorikeet.

Those are workloads where a two-second response is the product requirement and a five-second one is a bug. They are also workloads where being wrong is recoverable — a bad hint, a bad search answer, an escalated support turn — which is exactly the risk profile a model without published benchmark scores can safely be tried on.

Where this leaves a builder this week

Mercury 3 is not on the OrcaRouter catalogue. Our model API currently returns nothing for any Inception model, and the Mercury 3 model page does not exist on our site, so anyone who wants to call it goes to Inception's own platform or to one of the third-party hosts Inception lists. We would rather say that plainly than imply a route we do not have.

What our routing layer is good for in this situation is the part around the new model rather than the model itself. The pattern that works with an unproven, unusually fast release is a two-model path: send the latency-critical tier to Mercury 3 through Inception's key, and keep the fallback and the hard cases on models we do carry — more than 200 of them behind one OpenAI-compatible endpoint, at provider list price with 0% markup, with automatic failover if a provider degrades. You get the speed where it pays and a second opinion where it matters, without a second integration for the fallback.

Three things are worth watching over the next few weeks. First, an independent Mercury 3 measurement, which will settle both the throughput number and the quality position in one go. Second, whether Inception publishes the effort setting and hardware behind its own figures, since a benchmark without those is not reproducible. Third, whether the launch discount gets an end date — the list price is 33% below Mercury 2.5's output rate and roughly a fifth of Claude Haiku 5.5's, and a company that has just scaled a model 40 times has reasons to want the promotional number to be the one people remember.

Until then the honest summary is this: Mercury 3 is a real release with a defensible architectural story, a genuinely aggressive price, and a quality claim that nobody outside Inception has yet tested. For a latency-bound workload with a human or an escalation path in the loop, that is enough to justify a week of evaluation. For anything where a wrong answer is expensive and unmonitored, it is not yet enough to justify a migration.