Mercury 3 vs GPT-6.1 Sol: Inception Ran This Test Itself, and That Is the Story

Mercury 3 shipped on October 19, 2026, and the only head-to-head number Inception published for it names GPT-6.1 Sol directly. Fifty live Mind2Web navigation tasks, run inside Inception's own production web-agent harness: Mercury 3 finished 50% of them, which the launch post says is "significantly more than GPT-6.1-Sol", using the same token volume and the same number of agent steps, at what it describes as a fraction of the per-token cost. That is a more specific claim than most launch posts make, and it is also a claim with no denominator. Fifty tasks is a small sample, the harness is the vendor's, GPT-6.1 Sol's completion rate is never printed, and Mercury 3 does not yet appear anywhere on the independent Artificial Analysis board that has scored GPT-6.1 Sol at 51.8. So the honest comparison is not "which model is smarter". It is what the vendor's own test can support, what the rate card says, and where GPT-6.1 Sol's measured record still wins by default. The pass-through rate is the same across the catalogue: 0% markup provider list price in, provider list price out. 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.

First, the part of the claim that is unusually checkable

Launch benchmarks are usually a table of scores on benchmarks the vendor chose, compared against models the vendor chose, with no statement about how much compute either side spent. Inception did something rarer: it fixed the token volume and the step count. "Same token volume and the same number of agent steps" is a statement about experimental control, and it matters because the most common way a fast model games an agent benchmark is to take more attempts. If Mercury 3 really matched GPT-6.1 Sol's steps and tokens and still finished more tasks, the difference is in what the model did with each step, not in how many it got.

What the claim still lacks:

• The absolute completion rate for GPT-6.1 Sol. "Significantly more" than an unstated number is not a result you can price

• The pass criterion. Mind2Web tasks differ in how strictly a navigation counts as complete, and the harness is Inception's, not a public one

• A second run. Fifty tasks is small enough that a handful of flips moves the percentage several points

• Any score on a benchmark with a published baseline, which is why the Artificial Analysis board cannot yet place Mercury 3 at all

TinyFish AI's Shuhao Zhang is quoted on the result, and the quote is a customer testimonial rather than an independent replication. That is normal for a launch and worth naming as such, because the 50% figure will circulate for months without the qualifier attached to it.

The rate card, which is where the case is actually strong

The price difference is not marginal and it does not depend on anyone's harness. Mercury 3 lists at $0.20 per million input tokens and $0.70 per million output tokens, with cached input at $0.02. Through the launch window Inception is charging half of that: $0.10 input, $0.01 cached and $0.35 output. GPT-6.1 Sol lists at $2.00 input, $0.10 cached, $2.50 cache write and $10.00 output. At list, Mercury 3's output rate is one-fourteenth of GPT-6.1 Sol's; during the launch discount it is one twenty-eighth.

Take a browser agent that runs forty steps, sends 12,000 input tokens per step and emits 400 output tokens per step. That is 480,000 input tokens and 16,000 output tokens for the run, all uncached, which is the pessimistic case for the cheaper model:

• GPT-6.1 Sol — $0.96 of input at $2.00 per million, plus $0.16 of output at $10.00 per million: about $1.12 per run

• Mercury 3 at list — $0.096 of input plus $0.011 of output at $0.70 per million: about $0.107 per run, or 10.4 times cheaper

• Mercury 3 at the launch rate — $0.048 plus $0.006: about $0.054 per run, or roughly 21 times cheaper

Now put the completion rates on top of that. If Mercury 3 finishes 50 of 100 tasks and GPT-6.1 Sol finishes some smaller number, the per-run saving is only realised on the runs that finish. A model that is ten times cheaper per step and finishes two-thirds as often is still cheaper per completed task, which is the number that belongs on a dashboard. But at 50 tasks, nobody knows the second ratio yet.

Two cost details matter more than the headline and are easy to miss. Cached input at $0.10 on GPT-6.1 Sol is half of Mercury 3's list input price and twenty times its cached price, so on a long, stable system prompt GPT-6.1 Sol's caching closes part of the gap — but only the input side, and output is where agent loops spend. And GPT-6.1 Sol reprices the entire request above 272,000 input tokens: 2× on input and cache, 1.5× on output. That cliff is exactly where long agent transcripts land, and Mercury 3's 1M-token window with no published long-prompt surcharge is a structural advantage in the same scenario.

What GPT-6.1 Sol has that Mercury 3 does not: a measured record

Everything above is a comparison against a rate card and one vendor test. The independent board is a different picture, and it only has one of these models on it.

• Intelligence Index — Mercury 3 not listed; GPT-6.1 Sol 51.8 at maximum reasoning effort, per Artificial Analysis on October 20, 2026

• Cost per index task — Mercury 3 not measured; GPT-6.1 Sol $0.72, against a board median well above that

• Terminal-Bench 4.0 — Mercury 3 not measured; GPT-6.1 Sol 0.5606

• SciCode — Mercury 3 not measured; GPT-6.1 Sol 0.5417

• Humanity's Last Exam — Mercury 3 not measured; GPT-6.1 Sol 0.5292

• AA-LCR long-context reasoning — Mercury 3 not measured; GPT-6.1 Sol 0.83

• GDPval-AA, economically valuable knowledge work — Mercury 3 not measured; GPT-6.1 Sol Elo 1593.85

• AA-Omniscience — Mercury 3 not measured; GPT-6.1 Sol 41.5

• Context and output — Mercury 3 1,000,000 in / 131,072 out; GPT-6.1 Sol 1,050,000 in / 128,000 out

• Weights — closed on both; neither gives you a checkpoint

• Modality — text on both; GPT-6.1 Sol also takes images and files

That last line is the least glamorous and most decisive difference for some teams. GPT-6.1 Sol accepts image and file input. Mercury 3's own documentation lists supported formats as text. If your agent has to read a screenshot, a PDF or a scanned invoice, the comparison ends there regardless of price, and no amount of tokens per second changes it.

The reasoning ladders also are not the same shape. Mercury 3 offers instant, low, medium and high, with temperature adjustable from 0.5 to 1.0. GPT-6.1 Sol offers low, medium, high, xhigh and max, and no longer accepts the no-reasoning setting at all. A model with an "instant" rung is making a different product promise: it is telling you some requests should not think, and it will bill you accordingly. That is a real design difference, and it is invisible in a single intelligence number.

Where each model is the right call

Mercury 3 is the right call when the workload is latency-bound, text-only, and per-step cost dominates the budget: browser navigation, chat support, real-time tutoring, high-volume classification-adjacent work, any loop where a human or another system is waiting on each turn. The 1M context and 131,072-token output ceiling fit long agent transcripts, and the instant rung exists for the requests that should not think. The risk is that no independent evaluator has scored it, so its quality envelope is a vendor claim for now.

GPT-6.1 Sol is the right call when the work needs vision or file input, when you need a measured capability floor before you commit traffic, when you want a reasoning ladder that goes past "high", or when 272,000-token requests are the exception rather than the norm. Its cost per index task is one of the best on the board, which is the other half of the argument: cheap per token and cheap per finished job are different claims, and GPT-6.1 Sol is the one with the second number published.

The practical mistake is treating this as a winner-take-all choice. Mercury 3 is not on OrcaRouter — it is served through Inception's own API and several third-party platforms — while GPT-6.1 Sol is, at OpenAI's list price with 0% markup, so the vendor's own rate card is what you pay rather than a marked-up equivalent. That makes the split test cheap to run: keep GPT-6.1 Sol on the production path, put Mercury 3 behind the vendor's API on a slice of traffic, and compare completed tasks per dollar rather than tokens per second. If the slice holds up, move the latency-sensitive steps; if it does not, you have lost a week and a small invoice. Automatic failover matters most in exactly this window, when a day-old model is on trial and a measured one is carrying the load.

What would settle this

Three things, in order of how much they would move the answer:

• An Artificial Analysis entry for Mercury 3. One index number and one cost per task would replace every estimate in this piece. Until it exists, "no independent score" is the correct status, not "unproven" and not "frontier-class"

• A completion rate for GPT-6.1 Sol on the same 50 tasks, or a public rerun of that harness by someone who is not Inception. The controlled-token claim is good methodology; it is not yet a result anyone else can inspect

• A published answer on how diffusion output is billed. Mercury 3 generates tokens in parallel and refines them in passes, and Inception's docs do not say whether the denoising passes count toward billed output tokens. If they do not, the effective cost advantage is larger than the rate card suggests; if they do, the 1,068 tokens per second is a throughput figure rather than a billing one

Until then the defensible position is narrow and useful: Mercury 3 is the cheapest fast text model in this matchup by an order of magnitude, on a vendor test that was at least designed to control for effort, with a 1M-token window and no independent score. GPT-6.1 Sol is the more expensive, slower, measured model that also reads images and has a published cost per completed task. Buy the second for the work you cannot afford to get wrong, trial the first for the work where waiting is the actual cost.