Mercury 3 vs Grok 4.6: One Is Superseded, the Other Is Unmeasured, and Both Are Sold on Speed

The first useful fact about comparing Mercury 3 with Grok 4.6 is that Grok 4.6 is not the current G​rok. SpaceXAI shipped it on August 12, 2026 at $2.00 per million input tokens and $6.00 per million output, and has since released Grok 4.7; Artificial Analysis now marks 4.6 as deprecated and points readers at the newer model. Mercury 3, by contrast, is two days old: Inception Labs released it on October 19, 2026 at $0.20 and $0.70 per million, halved for a launch window with no published end date, and nobody neutral has measured it at all. So the honest framing of this pairing is not "which of these two should I use". It is that both are sold primarily on a speed number, the two speed numbers are not the same kind of number, and one of the two products is on its way out of the catalogue while the other has not yet been tested by anyone who does not work for the company that made 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.

That makes this a comparison about measurement, which happens to be the most useful thing to know before you price either. Mercury 3 is a diffusion language model — it denoises a block of positions and commits them together rather than emitting one token per pass — and Inception reports 1,068 tokens per second on NVIDIA hardware, more than 40 times the total parameters of Mercury 1, a context window grown from 260,000 to 1,000,000 tokens, and a 131,072-token output ceiling. Grok 4.6 is a conventional frontier model built for long agentic runs, with a 500,000-token context, text, image and file input, and a 75% cache discount at $0.50 per million reads.

What the two speed numbers actually measure

Inception's 1,068 tokens per second is an output-throughput figure: how fast tokens leave the model once it starts producing them. Artificial Analysis measures the same quantity for Grok 4.6 and gets 65.4 tokens per second. Read those two side by side and Mercury 3 looks sixteen times faster. That is a real difference and it is not the whole difference, because throughput is only one of the three numbers a latency budget is made of.

The second is time to first token, and here the picture gets less flattering to the framing and more interesting. Artificial Analysis records a 37.19-second median time to first token for Grok 4.6 on the workload it tests, against a median of 3.75 seconds across the models on its board. Our own seven-day playground telemetry puts Grok 4.6's p50 at 8.00 seconds over the one day it has samples for. Neither of those is Mercury 3's problem or Mercury 3's advantage — Inception has published no first-token figure at all, and the only third-party latency numbers in its announcement are customer testimonials: Octen reported a median generation time of 0.9 seconds and a median end-to-end response of 2.2 seconds, RevisionDojo reported answers returning "in about a second", and Lorikeet said Mercury 3 beat other models on latency while scoring higher on its evals. Those are useful signals and they are marketing quotes, on workloads nobody outside those three companies can inspect.

The third number is the one that decides whether a fast model is usable, and it is error rate. On our own infrastructure over the last seven days, GPT-5.6 Sol's serving path recorded a 0.05% error rate. Grok 4.6, on the single day it has telemetry for — 17 October 2026 — recorded 86.7%. That is one day of samples on a model that has been live since August, so it is a warning about a serving path rather than a verdict on a model, and it is exactly the kind of number that never appears on a vendor page. For Mercury 3 there is no error rate at all, because there is no route to measure.

The spec sheet, with the provenance attached to each line

• Price — Mercury 3: $0.20 / $0.70 per million, $0.10 / $0.35 at launch, no end date published, cached input 90% off. Grok 4.6: $2.00 / $6.00 per million up to 200,000 tokens, $4.00 / $12.00 for the whole request above that, cached reads $0.50.

• Context and output — Mercury 3: 1,000,000 in, 131,072 out. Grok 4.6: 500,000 in, with no output ceiling published on our model page for it.

• Modalities — Mercury 3: text only. Grok 4.6: text, image and file in, text out.

• Throughput — Mercury 3: 1,068 tokens per second, Inception-measured. Grok 4.6: 65.4 tokens per second per Artificial Analysis; 256 tokens per second in our own playground on the day it had samples.

• Independent scores — Mercury 3: none, no Artificial Analysis page exists. Grok 4.6: 44 on the Artificial Analysis Intelligence Index, ranked 37th of 217 models, and 76.8 on its coding index, ranked 5th of 138.

• Status — Mercury 3: two days old, unmeasured. Grok 4.6: deprecated on Artificial Analysis's board, superseded by Grok 4.7.

The context comparison is the one that flips the usual ordering. Mercury 3 offers twice Grok 4.6's window — a million tokens against half a million — and does it at a tenth of the input price. On paper that is not a close contest for long-document work. In practice, a 500,000-token window is already larger than almost any single document anyone processes, and the interesting question about a million-token window on an unmeasured model is not whether it exists but whether quality holds across it. Artificial Analysis measures long-context recall explicitly, and Grok 4.6 scores 80.3 there. Mercury 3 has no equivalent figure, which means a million-token window with an unknown recall curve is a capacity claim rather than a capability claim.

What Grok 4.6's price is actually buying

Grok 4.6 costs ten times Mercury 3 on input and roughly nine times on output at list price, and the honest answer to what that buys is: a measured position on a neutral harness, multimodal input, and a published cost per answer. That last one is the most useful. Artificial Analysis puts Grok 4.6 at $1.48 per Intelligence Index task — a figure that already includes however many tokens the model needs to finish the suite, which is the only per-answer number in this entire comparison that exists for either model.

Against that, Mercury 3's $0.70 output rate is a per-token figure with no per-answer counterpart. If a Mercury 3 task consumes the same output volume as a Grok 4.6 task, it costs about a tenth as much. If it consumes three times the volume — which is plausible for a model defaulting to low reasoning effort on work that G​rok does at higher effort — it costs about a third as much, and the strategic case weakens from "an order of magnitude" to "a meaningful discount on an unproven model". Nobody can tell you which of those is true yet. That is not a rhetorical hedge; it is the specific measurement that is missing.

Grok 4.6's own numbers, for what they are worth, are strong in the places a long agentic run needs them: 94.9 on GPQA Diamond, 42.9 on Humanity's Last Exam, 80.3 on long-context recall, and a coding index of 76.8 that puts it fifth of 138 models. Those are Artificial Analysis's runs, not SpaceXAI's, which is why they can be compared to Mercury 3 at all — and Mercury 3, having no such runs, cannot be.

The superseded-model problem, which is the real decision

Choosing Grok 4.6 today means choosing a model that its own vendor has already replaced. Grok 4.7 is on our catalogue, and the 4.6 page on Artificial Analysis explicitly redirects readers to it. For a workload you are about to build against, starting on a deprecated model is a decision to migrate later; the only question is when, and the answer is usually "during an incident".

That does not make Mercury 3 the safe choice. An unmeasured two-day-old model and a deprecated nine-week-old one are different kinds of risk, and neither is the low-risk option in this pairing. What the comparison actually establishes is narrower and more useful: if you want a measured model with a documented per-answer cost and multimodal input, Grok 4.7 is the current version of that and 4.6 is its predecessor at the same price. If you want the cheapest text-only tokens available anywhere with a million-token window, Mercury 3 is that, with the caveat that the cheapest number on its card is a promotion with no end date.

Grok 4.7 is on the OrcaRouter catalogue at SpaceXAI's own rates, behind the same OpenAI-compatible endpoint as everything else we route, with no markup added on top and automatic failover across providers when one path rate-limits — which, given the error rate our telemetry recorded on 4.6's serving path on 17 October, is not a theoretical benefit. Mercury 3 is not on the catalogue: our model API returns nothing for any Inception model, so anyone calling it goes to Inception's platform, where new accounts get 100 million free tokens, or to one of the several third-party platforms that host it.

The routing pattern that fits this pair is not a split by price. It is a split by what you are willing to have fail. Put the deprecated model's traffic on the current one, put a bounded slice of the text-only work on Mercury 3, keep a measured model as the fallback, and let failover handle the case where the new path has a bad day. The slice is a config value; the fallback is a routing rule; the decision to widen or narrow it is made by your own error rates within a week.

What to do on Monday

If the work is long text, the volume is high, and the answers are checkable, Mercury 3's rate card is worth a bounded experiment this week — the 100 million free tokens on Inception's platform cover the test, and the million-token window means you can send the whole document rather than a chunked guess at it. Watch two things: whether output token counts per task run away from your Sol or G​rok baseline, and whether quality degrades toward the end of a long context. Both are measurable in a day and neither requires anyone's benchmark.

If the work involves images, runs for hours with tools in the loop, or has to be right the first time, the answer in this pairing is not Grok 4.6. It is Grok 4.7, or one of the measured frontier models, and the honest reason is that 4.6 is deprecated and Mercury 3 is unmeasured — a choice between two models you would not put in front of a customer without a fallback behind them.

The number that would change all of this is a neutral cost per completed task for Mercury 3. One figure, on the harness that produced G​rok's $1.48, would turn this article's central uncertainty into a ratio. Until it exists, the defensible position is the boring one: the cheap tokens are real, the speed claim is plausible and single-sourced, and the only thing you can verify yourself is your own error rate on your own traffic.