Mercury 3 vs MiniMax M3: One Number Against a Whole Record

Inception shipped Mercury 3 on October 19, 2026 with one performance figure — 1,068 tokens per second — and no benchmark score of any kind. M​iniMax shipped MiniMax M3 on June 1, 2026 with a vendor benchmark table, a downloadable checkpoint and, four months later, a published score from an independent evaluator. Put the two side by side and the interesting question is not which model is faster or cheaper. It is that one of these models can be checked by somebody other than its author today, and the other cannot. That asymmetry is the whole comparison, and it survives every price cut either lab announces. 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 two models also happen to be close enough on paper that the choice is real. Both run a one-million-token window. Both sit in the cheap tier. Both are built on architectures that are unusual for their class — diffusion for Mercury 3, sparse attention over a mixture of experts for MiniMax M3 — and both claim that architecture is what buys them their economics.

Two launch posts, two different kinds of evidence

Read Inception's announcement as a document. It makes a scaling argument: Mercury 3 is more than 40 times larger than Mercury 1 in total parameters, measured intelligence rose roughly threefold from Mercury 1 (mini) to Mercury 3, and the model is "just as fast" as its smallest ancestor. It makes a comparative quality claim: Mercury 3 is "on par or better than frontier speed-optimized models like GPT-6 Luna and Claude Haiku 5.5". And it reports exactly one task result, from Inception's own web-agent harness, on 50 live Mind2Web navigation tasks, where the model completed 50 per cent. Every one of those is vendor-reported and none names a benchmark suite, a margin, or a hardware configuration.

Read M​iniMax's material as a document and it is the opposite shape. There is a benchmark table with named suites and percentages — SWE-Bench Pro 59.0, Terminal-Bench 2.1 66.0, BrowseComp 83.5, OSWorld-Verified 68.70 — plus a model card, plus weights on Hugging Face, plus an independent listing from Artificial Analysis that scores the model at 29 on its Intelligence Index. The vendor numbers are still vendor numbers, and we have not seen them reproduced. But the model can be downloaded, run and measured by anyone who wants to, which is why the independent number exists at all.

That difference in what was published is not a marketing detail. It determines what kind of advice anyone can give you about these two models. For MiniMax M3 there is a measured throughput figure, a measured cost per index task, a measured rank. For Mercury 3 there is a claim.

The specs, with each side's source attached

• Price per million tokens — Mercury 3 $0.20 in and $0.70 out at list, halved to $0.10 and $0.35 while the launch discount runs vs MiniMax M3 $0.30 and $1.20 on requests up to 512K input tokens

• Cached input — Mercury 3 $0.02 per million, $0.01 during the launch window vs MiniMax M3 $0.06 per million

• Context window — Mercury 3 1,000,000 tokens vs MiniMax M3 1,048,576

• Maximum output — Mercury 3 131,072 tokens vs MiniMax M3 512,000 tokens on our own catalogue page; M​iniMax publishes no maximum-output figure, so that one is ours, not the vendor's

• Input modality — Mercury 3 text only vs MiniMax M3 text, image and video

• Architecture — Mercury 3 is a diffusion language model that refines blocks of tokens in parallel vs MiniMax M3 is a mixture of experts, roughly 428B total and roughly 23B active per token, with M​iniMax Sparse Attention

• Weights — Mercury 3 closed, API only vs MiniMax M3 downloadable under the M​iniMax Community License

• Throughput — Mercury 3 1,068 tokens per second, vendor-reported on unnamed NVIDIA hardware at an unspecified batch size and effort setting vs MiniMax M3 106.4 tokens per second measured by Artificial Analysis

• Independent Intelligence Index — Mercury 3 no entry as of October 20, 2026 vs MiniMax M3 29, ranked 23rd of 121 models on the revision current that day

• Reasoning control — Mercury 3 four effort settings, instant through high, defaulting to low when the parameter is omitted vs MiniMax M3 a thinking parameter with enabled, adaptive and disabled modes

• Tool calling and structured outputs — both support them on an OpenAI-compatible chat completions endpoint

Two of those lines need a caveat attached before they get quoted anywhere. The 1,068 figure is a vendor measurement without a stated harness, so it is not comparable to a number produced by an independent evaluator under a fixed methodology. And Artificial Analysis rescores its index periodically; the 29 above is the reading on October 20, 2026, and MiniMax M3 has read differently on earlier revisions. Numbers on that index move; the fact that one exists for M3 and not for Mercury 3 does not.

What 1,068 tokens a second means at the request level

Tokens per second is a poor unit for deciding anything, because nobody buys tokens. They buy completed requests. Convert both figures into the thing the user actually experiences — how long a thousand-token answer takes to generate.

At Inception's claimed 1,068 tokens per second, a 1,000-token answer takes about 0.94 seconds of generation. At MiniMax M3's measured 106.4, the same answer takes about 9.4 seconds. Include a measured time to first token of 1.13 seconds for M3 and the gap between the two experiences is roughly tenfold, before anyone adds network latency, queueing or a provider's own overhead.

That is the case for Mercury 3, and it is a real one. A tenfold difference in generation time is the difference between a search answer that feels instant and one that feels like a page load. Inception's whole product line is built around this: the company sells Mercury Voice for time-to-first-answer-token, Mercury Decide for classification, and positions the flagship at support, tutoring, search and web agents — all workloads where a slow answer is a broken product rather than a slightly worse one.

What the tenfold number does not tell you is whether Mercury 3 is any good. The Mind2Web result — 50 per cent completion on 50 live tasks, "significantly more than GPT-6.1-Sol" — is the only structured quality signal Inception published, and 50 tasks is a sample small enough that a single difficult site moves the percentage by two points. The harness is Inception's own. The comparison model is named without a score. It is a genuine signal that the model can drive a browser, and it is not a measurement you can put in a procurement decision.

What MiniMax M3's record actually says

M3's independent picture is unglamorous and useful. An Intelligence Index of 29 against a median of 19 for models in its comparison class, 106.4 output tokens per second against a 77.4 median, a time to first token of 1.13 seconds against a 1.70-second median, and a cost of $0.51 to run the index per model. It is a model that is well above average on quality for its class, clearly above average on speed for its size, and cheap to run.

The sub-scores that circulate from third-party republishing of the same index family tell a more specific story: strong on document understanding and instruction following, weaker on autonomous multi-step agent work. The gap between its coding and agentic sub-scores is the largest internal spread in its record. If your workload is "read this long document and answer precisely", M3 is well evidenced. If it is "run for twenty minutes and deliver", the evidence is thinner and points the other way.

For Mercury 3 there is no equivalent paragraph to write, because the data does not exist. Artificial Analysis has no Mercury 3 page — its index still lists Mercury 2.5 and Mercury 2 — so one day after launch the model has no independent score, no independent speed figure and no cost-per-task number. Inception's comparison set (GPT-6 Luna, Claude Haiku 5.5) is the fast tier rather than the frontier, which is the right comparator for a latency product, but the claim carries no margin and no test names.

The price gap is smaller than the sticker suggests

The reflex is to call Mercury 3 the budget option and M3 the expensive one. Per token that is directionally true and much less dramatic than it looks: input is 1.5 times cheaper on Mercury 3, output 1.7 times cheaper. That is not a different league. It is a different shelf.

Run a concrete workload. Take a support or search assistant handling 10,000 replies a day, each with 1,500 input tokens of retrieved context and prompt scaffolding and 300 output tokens. That is 15 million input tokens and 3 million output tokens daily.

• Mercury 3 at list price — 15 × $0.20 plus 3 × $0.70, or $5.10 a day

• Mercury 3 at the launch rate — 15 × $0.10 plus 3 × $0.35, or $2.55 a day

• MiniMax M3 at its small-request rate — 15 × $0.30 plus 3 × $1.20, or $8.10 a day

So the daily difference between the two list prices is about three dollars on ten thousand conversations. If Mercury 3's token efficiency at its default low effort setting is worse than M3's — and nobody outside Inception knows, because there is no measured output-token count for Mercury 3 — that three-dollar gap can close or invert on real traffic. This is the specific reason an unscored model is hard to budget for: you cannot compute cost per completed task without a measured completion rate and a measured verbosity.

The cached-input line matters more than the headline for agent workloads, because agents resend the same system prompt and tool definitions on every step. Mercury 3's $0.02 per million cached tokens makes a 20,000-token preamble cost four ten-thousandths of a dollar per call. M3's $0.06 costs three times that, which is still small — but on a fifty-step run the difference is cents, not fractions of a cent, and agent traffic is where cents multiply.

Licence, exit and the thing you cannot download

MiniMax M3 ships under the M​iniMax Community License. Non-commercial use is free; commercial use requires displaying "Built with MiniMax M3", a one-time notice below $20 million in annual revenue, and prior written authorisation above it, plus compliance with a prohibited-use appendix. That is a legal review rather than a formality, but it is a licence you can read before you build.

Mercury 3 has no licence to read, because there is nothing to license. It is closed. That is not automatically disqualifying — most teams calling a model over an API never touch the weights — but it changes what happens if Inception changes its price, its terms or its availability. The exit is a migration, not a redeploy.

One more asymmetry worth naming: M3 has a four-month track record. It has been on our catalogue long enough to accumulate traffic and latency percentiles, and its rate card has been stable. Mercury 3 is one day old, its launch price has no published end date, and its predecessor models are already closed to new users — Inception's own migration note says Mercury 2 and Mercury 2.5 stop accepting new signups and existing customers lose access on November 15, 2026. A vendor that retires a model line that quickly is telling you something about how it expects you to consume its products.

Where a router changes the arithmetic

This matchup is a good argument for not choosing at all in the first week. MiniMax M3 is on OrcaRouter at $0.30 in and $1.20 out per million tokens, the provider's list price with no markup on top, alongside more than 200 other models behind one OpenAI-compatible endpoint. Mercury 3 is not on our catalogue — a check of our model API returns nothing for any Inception model — 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.

The pattern that fits an unmeasured, unusually fast release is a two-path design. Put the latency-critical tier on Mercury 3 through the vendor, and keep the measured fallback on M3 or another model we serve, with automatic failover wired so that a 429 or a degraded endpoint does not take the feature down. The routing DSL exists precisely so that the split can be a rule rather than a fork in your codebase, and because provider list prices pass through unchanged, an Inception price cut or a M​iniMax tier change reaches your bill the day it is announced rather than at the next contract.

What that design does not buy you is a quality guarantee on the Mercury 3 side. Nothing does, yet.

The rule to apply this week

• If the answer must arrive in about a second and a wrong answer is recoverable — search suggestions, support triage, tutoring hints, agent steps with an escalation path — Mercury 3 is worth a week of evaluation on the strength of the latency claim alone. Budget against the list price, not the 50 per cent launch rate, since the discount has no published end date.

• If you need to know how good the model is before you commit, MiniMax M3 is the only one of these two that lets you find out. It is also the only one you can download, quantise, fine-tune or run on your own hardware.

• If your workload reads video or images, the comparison is over. M3 takes them; Mercury 3 does not.

• If you are optimising cost per completed task rather than cost per token, neither model gives you the number you need yet — M3 because its agentic weakness is measured and real, Mercury 3 because nothing about its completion behaviour has been measured at all.

The thing to watch is not Mercury 3's price. It is the first independent evaluation of Mercury 3, whenever it lands. If the throughput claim reproduces anywhere near 1,068 tokens per second under a neutral harness and the quality lands in the Haiku class Inception is claiming, this stops being a matchup and becomes a migration question for every latency-bound product on the web. If it lands at the usual vendor-to-independent discount and the quality is merely competitive, then the honest summary is the one above: a fast, cheap, unproven model against a slower, cheap, proven one — and for most teams the proven side is the one that gets the production path until the other side has a number.