Mercury 3 Grew More Than 40x and Kept 1,000 Tokens a Second — the Scaling Is the Story

Inception Labs released Mercury 3 on October 19, 2026, and the number that should stop you is not 1,068 tokens per second. It is forty. Mercury 3 is, by Inception's own account, more than 40 times larger than Mercury 1 in total parameters, and it still runs at 1,000-plus tokens a second — the same throughput class the company has been selling since its first model. Everyone covering the launch has led with the speed figure. The speed figure was already Mercury's pitch two generations ago. What is new is that Mercury 3 is a diffusion language model that got two orders of magnitude of scale thrown at it without the usual collapse in decode speed, and that is the claim worth testing. This piece is about the scaling, what Inception says it bought, and what a builder can and cannot conclude from it a day after launch. 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.

The comparison Inception makes is to the fast, cheap tier: Mercury 3 "outperforms GPT-6 Luna and Claude Haiku 5.5", both speed-optimised models, while running faster than either. It is not positioned against GPT-6 Astra or Claude Sonnet 5.5, and reading the announcement as a frontier-model claim would be a mistake Inception is not making.

Why 40x is a harder claim for a diffusion model than for a transformer

An autoregressive model emits one token per forward pass through the network. A diffusion language model emits a whole block per pass, iteratively denoising a canvas of positions until the block resolves. That is where the throughput comes from, and it is also why scaling one is a different engineering problem from scaling the other.

In a small diffusion model, each denoising step can commit many positions at once because the remaining positions are easy to fill in. As the model grows, the quality bar per position rises, and the number of steps needed to reach that bar tends to rise with it — the parallelism that produced the speed advantage is exactly the thing that gets harder to hold onto. A 40x parameter increase that preserved 1,000 tokens per second therefore implies either a large increase in steps per block offset by a much larger increase in per-step compute efficiency, or an architectural or sampling change Inception has not described. The announcement does not say which, and that is the single most interesting unanswered question in the release.

What Inception does publish is the outcome rather than the mechanism: more than 40x the parameters of Mercury 1, and roughly a threefold increase in measured intelligence from Mercury 1 (mini) to Mercury 3 across its own suite. Both figures are vendor-reported, neither comes with an absolute parameter count, and no independent harness has run the model yet. Treat them as a directional claim — bigger and better in the same family — rather than as measurements.

What the scale-up did and did not change on the spec sheet

The generational jump shows up in the boring numbers more clearly than in the marketing ones. Against Mercury 2.5, the model it replaces:

• Context window — 1,000,000 tokens on Mercury 3 vs 260,000 on Mercury 2.5, with a 131,072-token maximum output.

• Parameters — more than 40x Mercury 1, no absolute figure published; Mercury 2.5's parameter count was never published either.

• Throughput — 1,068 tokens per second, vendor-measured on unspecified NVIDIA GPUs, vs the roughly 706 tokens per second Artificial Analysis measured for Mercury 2.5 under its own harness.

• Price — $0.20 per million input and $0.70 per million output at list, cut 50% to $0.10 and $0.35 for the launch period, which has no stated end date.

• Reasoning effort — four settings (instant, low, medium, high), defaulting to low; Mercury 2.5 and Mercury 2 default to medium.

• Modalities — text in, text out. Unchanged. No image, audio or video input, where Gemini 4 Argon and GPT-6 Astra both take multimodal input.

The context window is the change with the most practical consequence and the least press. A fourfold increase moves Mercury out of the short-document bracket and into the range where a whole repository, a long support transcript or a multi-hundred-page contract fits in one call. Diffusion decoding over a long context is not free — the canvas is larger and attention cost grows — so the fact that the context grew by 4x while throughput stayed at 1,000-plus is, if anything, a stronger statement about the engineering than the headline number.

The one quality result with a harness behind it

Inception did run an evaluation that is more useful than a benchmark table, and it is buried below the speed claims. The company put Mercury 3 inside its own production web-agent harness alongside frontier models, on 50 live Mind2Web navigation tasks — real sites, not a frozen snapshot — and reports a 50% completion rate, "significantly more than GPT-6.1-Sol", at the same token volume and the same number of agent steps. Same work, fewer dollars per token, better completion.

Fifty tasks is a small sample and the harness belongs to the vendor, so the margin is not a number you can plan against. But agentic web navigation is the workload least amenable to benchmark contamination and most sensitive to the two things Mercury sells: latency and cost per step. A model that finishes half of fifty live tasks is doing something real. It is also, at 50%, failing half of them, which is the correct frame for this class of model in production — an agent that completes half its tasks with a human escalation path is deployable; one presented as reliable is not.

Where the launch discount sits in the actual budget

The list price is $0.20 in and $0.70 out per million tokens. The launch price is half that. Neither is a rounding error against the tier Mercury is aimed at: Claude Haiku 5.5 lists at $1.00 and $5.00, and GPT-6 Luna at $0.10 and $0.50. So Mercury 3 at launch is priced at the very bottom of the market it is competing in, and at list it is still roughly a seventh of Haiku 5.5's output rate.

Run the arithmetic on a support-deflection workload of 20 million input and 10 million output tokens a day, which is a mid-size deployment rather than a pilot. At list, that is $4 of input and $7 of output — $11 a day, about $330 a month. At the launch rate it is $5.50 a day. The discount is worth roughly $165 a month at that volume, and it disappears on a date nobody has published. Build the case on $330.

Cached input is the line that actually moves agent economics. At $0.02 per million list, a 20,000-token preamble of system prompt and tool definitions costs four hundredths of a cent per call instead of four tenths. Across a forty-step agent run that is a cent and a half of preamble instead of sixteen cents — not transformative on its own, but it is the difference between an agent loop whose cost is dominated by the model's work and one whose cost is dominated by re-reading its own instructions.

What diffusion at this size still has not solved

Three things are worth stating plainly, because the launch framing does not.

First, there is no independent measurement. As of October 20, 2026, Artificial Analysis — the neutral harness this blog uses for cross-model comparison — has no Mercury 3 page at all. Its newest Mercury entries are 2.5 and 2, measured at about 706 and 684 tokens per second. Every Mercury 3 figure in circulation, including every figure in this article, traces back to Inception.

Second, text-only is now a real limit rather than a stylistic one. The competitors Mercury 3 is priced against in the cheap tier are increasingly multimodal, and the long-context window it just gained is most useful for documents that arrive as PDFs and screenshots. A 1M-token window you can only fill with extracted text is worth less than the spec sheet suggests.

Third, the default effort setting is low. A model that answers at low effort by default will look faster and score lower than the same model at high, and Inception's comparative claims do not state which setting produced them. Anyone benchmarking Mercury 3 against anything should pin the effort parameter explicitly, or the comparison is not reproducible — including against Inception's own numbers.

How to hold this model without betting a production path on it

Mercury 3 is not on the OrcaRouter catalogue. Our model API returns nothing for any Inception model — the 209 models we route do not include it — so calling Mercury 3 means going to Inception's own platform, where new accounts get 100 million free tokens, or to one of the inference platforms Inception lists. Saying that plainly is more useful than implying a route we do not have.

What a router is genuinely good for here is the shape of the deployment rather than the model. The pattern that fits an unproven but unusually fast release is a two-path one: send the latency-critical work to Mercury 3 on 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 when a provider degrades. The pass-through pricing matters for the same reason the launch discount does: when Inception ends the 50% cut, a routed model's rate changes the day the vendor changes it, with no renegotiation.

The honest summary after one day is that Mercury 3 is a real scaling result wearing a speed headline. The 40x claim is unverifiable, the throughput claim has no neutral reproduction, and the quality claim is comparative against the fast tier rather than quantitative. But the direction is unambiguous: diffusion language models were a latency curiosity at 260K context and a 2.5-generation model, and at 1M context and 40x the parameters they are now in the price bracket where the fast tier actually lives. The number to watch is not whether 1,068 reproduces. It is whether the next independent index puts Mercury 3 within a few points of GPT-6 Luna and Claude Haiku 5.5 — because if it does, the cheap tier stops being a tier and becomes a decision.