Inception Mercury 3: 1,068 Tokens per Second and How Its Benchmarks Compare

Bottom line: Inception released Mercury 3 on October 19, 2026. Three reports cite 1,068 tokens per second and a fourth says “more than 1,000.” Published quality comparisons are inconsistent: one source says the announcement included no scores, while another table shows Claude Haiku ahead in some rows.

The speed number is the one people will paste. The score number depends on which page they copied it from. OrcaRouter’s October 19 note says Inception Labs released Mercury 3 that day, and that the announcement leads with three figures: 1,068 tokens per second on NVIDIA GPUs, a 1 million token context window, and $0.10 per million input tokens while a launch discount lasts. The same note says the announcement does not contain a single benchmark score. The quality line is comparative. Mercury 3 “outperforms GPT-6 Luna and Claude Haiku 5.5.” Every quantitative claim OrcaRouter attributes to that post is speed, size, or price.

XenoSpectrum on October 20 prints a table it says Inception published. Four rows. Haiku 5.5 is ahead on two of them. Both pages are about the same Monday launch. They cannot both be a complete description of the same scoreboard unless one of them is looking at a different document than the other. The split between the secondary pages is the thing you can check tonight.

This site has already had to separate a sticker price from an average, and a score from a rank. Haiku 5.5’s short-prompt rate is $0.10. Argon’s public price has an end date. Mistral Large 4’s 38 was a score, not a trophy. Mercury 3 is the version where the writeups do not agree on whether a scoreboard exists.

Mercury 3 at 1,068, or more than 1,000, or 1,000-plus

OrcaRouter, Startup Fortune on October 19, and XenoSpectrum all print 1,068 tokens per second on NVIDIA hardware. shattered.io on October 20 prints “more than 1,000.” Startup Fortune also quotes OpenRouter, the same day, saying Mercury 3 generates “1,000+ tok/s with 1M context, 4x more than Mercury 2.5.”

Those are three different sentences. 1,068 is a count. “More than 1,000” is a floor. “1,000+” is a floor with a plus sign. If your note says 1,068, you are following OrcaRouter, Startup Fortune, and XenoSpectrum, all of them attributing the count to Inception. If your note says more than 1,000, you are following shattered.io’s spec table, which also attributes the figure to Inception’s announcement. shattered.io’s FAQ is explicit that the figures had not been independently reproduced as of that writing. Startup Fortune says the same thing in its own words: every one of those numbers comes from Inception, and the comparisons have not been replicated by an independent evaluator. XenoSpectrum says BenchLM’s October 2026 release log lists the model under October 19, with links to the announcement and a models page, and that a release log is not independent verification.

Vendor-reported, not reproduced, and not the same string on every desk.

The multipliers are easier to mix up, and they are not the same comparison.

shattered.io’s table says Mercury 3 is more than 40 times the size of Mercury 1, and that the claimed speed advantage is 4 to 8 times GPT-6 Luna and Claude Haiku 5.5. Size versus Mercury 1 is not speed versus Luna. XenoSpectrum repeats both: more than 40 times the parameters of the original Mercury, and quality “exceeding” Luna and Haiku at 4 to 8 times the generation speed. OpenRouter’s line, via Startup Fortune, is a third ratio: 4x more than Mercury 2.5, attached to the 1,000-plus throughput claim, not to the parameter claim.

Forty times a prior model, four times a later small model, and four-to-eight times two rivals’ speed. If a slide has one “4x” on it, ask which denominator. OrcaRouter adds that Inception did not publish an absolute parameter count. “More than 40x Mercury 1” without a base number is not a parameter count you can put in a capacity plan.

The price on the card is the discount, unless the card forgot

OrcaRouter separates list from launch. List, in that piece: $0.20 per million input tokens and $0.70 per million output. Launch discount: $0.10 and $0.35. Cached input: $0.02 at list, $0.01 during the launch window. Startup Fortune says the $0.10 / $0.35 launch price is a discount the firm calls 50% off its standard rate. Half of $0.20 is $0.10. Half of $0.70 is $0.35. Those two pages agree on the arithmetic.

If you copy shattered.io’s table into a spreadsheet as the ongoing rate, you may be copying the sale. OrcaRouter does not say when the launch discount ends. Until you are looking at Inception’s own pricing page, write down which of those sentences you used.

Context is the stable row. OrcaRouter: 1 million tokens, up from 260K on Mercury 2.5, maximum output 131,072. Startup Fortune: Mercury 2.5 topped out around 260,000, Mercury 3 handles a million. XenoSpectrum: from the previous generation’s 260,000 to 1 million. shattered.io: 1 million, without the prior figure in the spec table. The other three confirm the million. They do not all confirm the output cap.

One page says the scoreboard is empty. Another page prints four rows.

This is the part that changes what you do with the model this week.

OrcaRouter says the announcement gives no margin, no benchmark names, and no per-task results, so the “outperforms Luna and Haiku 5.5” sentence cannot be checked against a public harness without running it yourself. It also says Inception’s comparison claims do not state which reasoning setting produced them. Mercury 3, on that page, has four effort settings: instant, low, medium, and high. The default, if you omit the parameter, is low. Mercury 2.5 and Mercury 2 default to medium. A model that answers at low by default will look faster, and can score lower, than the same model at high. If the comparison hid the setting, your rerun will not match it.

XenoSpectrum describes a different artifact. It says Inception published a benchmark comparison table, and that all three models were evaluated with reasoning set to medium. The rows it prints:

GPQA Diamond: Mercury 3 89.23, GPT-6 Luna 86.36, Claude Haiku 5.5 84.51.

IFBench: 76.84, 61.82, 54.65.

DeepSWE v1: 46.31, 37.76, 51.62.

MRCR v2: 75.50, 62.72, 78.40.

On that table Mercury 3 leads the first two rows. Haiku 5.5 leads DeepSWE v1 by 5.31 points and MRCR v2 by 2.90 points. XenoSpectrum’s own sentence is the one to keep: the model does not always post the top score, and the models differ in what they handle best. “Outperforms Haiku 5.5” is not true of every row on the table the same article prints.

Startup Fortune names three benchmarks, not four, and not the same set. It says Inception claims Mercury 3 beats Haiku 5.5 and GPT-6 Luna on IFBench, GPQA Diamond, and SWE-Bench Atlas. IFBench and GPQA Diamond are on XenoSpectrum’s table, and Mercury 3 leads both. DeepSWE v1 and MRCR v2, the two rows Haiku leads, are not in Startup Fortune’s three-name list.

So a reader of Startup Fortune can walk away thinking three wins. A reader of XenoSpectrum can walk away thinking two wins and two losses, at medium effort, on a vendor table. A reader of OrcaRouter can walk away thinking the announcement published no scores at all, and that the comparison did not even name the effort setting. A reader of shattered.io gets the outperform headline and a 4-to-8x speed claim, with a FAQ that says the figures are unreproduced.

Neither OrcaRouter nor XenoSpectrum can be declared wrong from the public pages alone. It is possible they read different URLs, or that a table went up after one of them looked, or that one of them is describing a graphic the other did not count as “the announcement.” Haiku is ahead on two printed rows. The blank-scoreboard claim and the four-row table are both in circulation as of October 20. Cite the page.

OrcaRouter does record one structured quality result, and it is not that table. Inception ran Mercury 3 inside its own production web-agent harness on 50 live Mind2Web navigation tasks, and reports 50% completion, “significantly more than GPT-6.1-Sol,” at the same token volume and the same number of agent steps. Fifty tasks. Inception’s harness. The comparison model is named without a score. OrcaRouter calls that a signal, not a leaderboard. One is a vendor table row with two other models filled in. The other is a completion rate on a private harness against a model whose score is not printed.

There is a further setting mismatch even if both pages are accurate. OrcaRouter says the default effort is low, and that the comparison claims do not state a setting. XenoSpectrum says the table used medium for all three models. Medium is not the default OrcaRouter describes. If you call the API and omit effort, OrcaRouter says you are on low, not on the table. Pin the parameter if you care which sentence you are testing.

What you can put in a config file tonight

Access, across the pages that list it: Inception’s own API, OpenRouter, Baseten, and Vercel. shattered.io, Startup Fortune, and XenoSpectrum agree on that set. OpenAI-compatible chat completions, so a base-URL swap is the migration, not a rewrite. OrcaRouter adds tool calling and structured outputs on the chat completions endpoint, text in and text out, temperature adjustable between 0.5 and 1.0. New API keys on Inception’s platform come with 100 million free tokens, according to both OrcaRouter and XenoSpectrum. That allowance is not the same thing as the launch token price. A free bucket ends. A $0.10 rate might be the sale. Write them on different lines.

Architecture, if you need the noun: diffusion language model, Inception’s term dLLM, tokens generated in parallel rather than one at a time. That is OrcaRouter’s account of where the throughput claim comes from. It is not an independent measurement of 1,068.

The table that exists, if you trust XenoSpectrum’s transcription, already falsifies the universal version of that sentence. The page that says there is no table, if you trust OrcaRouter, means you do not even have vendor rows to argue from.

If you do try it, pin effort. OrcaRouter’s warning is the operational one. Low is the default it describes. Medium is the setting on the table XenoSpectrum transcribed. A latency test at low against a quality claim at medium is how a bake-off lies to you. Log the setting next to the tokens per second. Otherwise you will reproduce neither number, and you will not know which one you missed.

What this means for you

Mercury 3’s speed is its clearest selling point. If quality matters for your use case, run your own evaluation on representative prompts and compare against your current model before switching on speed alone.