One exact number, and a claim around it

Inception's launch post leads with a comparison: Mercury 3 \"outperforms GPT-6 Luna and Claude Haiku 5.5, while running 4–8x faster at 1,000+ tokens per second.\" The speed is then given precisely: Inception reports generation throughput of 1,068 tokens per second on NVIDIA GPUs. That is a specific claim, and a testable one — anyone can time the live API. But the launch materials do not disclose the GPU model, batch size, request concurrency or output length behind the figure, so the advertised result cannot be reproduced exactly from the conditions Inception published. Specific is not the same as checkable in full.

The 4–8x range is a separate assertion again. The blog does not say which metric it measures — output generation speed, end-to-end latency or time to first token — or under what serving conditions, and it should not be read as a guaranteed reduction in the response time a user feels. The two speed claims were not presented as one experiment, and they should not be combined into one.

The quality side of the headline is stated just as broadly and evidenced differently. The announcement's prose names no benchmark, no score and no margin for either rival. The scores exist, but they sit in Inception's own benchmark table rather than in the text, and no machine-readable results table is published. Read directly, the table — all three models at medium settings — gives Mercury 3 a GDPval Elo of 1,237.2 against 1,204.0 for Luna and 1,225.0 for Haiku 5.5, a GPQA Diamond score of 89.23 against 86.36 and 84.51, and rival output speeds of roughly 130 and 240 tokens per second. Mercury 3 tops eight of the table's twelve quality rows. Luna tops two — AA-Omniscience-Public accuracy and the AA-LCR v1.1 long-context test — and Haiku 5.5 tops two, MRCR v2 and DeepSWE v1. Every one of those numbers is Inception-published. No independent benchmark of Mercury 3 had appeared by the time of writing.

The models page uses a quieter word

Inception's own models page, checked on the launch date, describes the same comparison in different language: \"Comparable quality to cost-optimized frontier models like GPT-6 Luna (medium) and Claude Haiku 5.5 (medium).\" That is a framing difference, not a demonstrated contradiction: a model can lead on most rows of its own benchmark table while its vendor's summary word for the overall picture is \"comparable.\" What readers cannot determine from the launch materials is how much of the evaluation picture the printed table represents — Inception chose which benchmarks to publish, and the announcement's broader claim is not broken down beyond the table. The breadth of the reported advantage is therefore only partly determinable: broad across the table Inception printed, unverified beyond it.

The scale claim shifts its baseline in the same document. Mercury 3 has more than 40x the total parameters of Mercury 1, while \"intelligence\" has \"risen 3x\" measured from Mercury 1 (mini) to Mercury 3. Size is measured against one starting point, intelligence against a smaller one. The absolute parameter count behind the 40x is not disclosed, and the only definition offered for the quality measure is a footnote on the scaling chart: an average of eleven named benchmark scores, mixing an Elo rating with percentage results — Inception's own composite, not a metric anyone else publishes or can recompute from the outside. None of this makes the model weak. It makes the launch arithmetic impossible to check in full.

The one concrete test names a third rival

The blog's most specific result is an agent test. Inception reports that Mercury 3, running inside its own production web-agent harness, completed 50% of 50 live Mind2Web navigation tasks — 25 completed tasks. Inception says that was significantly more than GPT-6.1-Sol managed at the same token volume and number of steps, but it prints no Sol score, and \"significantly\" is the blog's word, offered with no statistical analysis. The protocol detail is thin: how the tasks were selected, what counted as success and what retry policy applied are not stated, so the margin cannot be assessed from outside. The 25 of 50 stands on its own as a vendor-reported observation; the comparison attached to it does not.

Speed still matters, and the partner accounts point at why. Inception cites Octen AI, testing Mercury 3 for search on smart glasses, reporting a median generation time of 0.9 seconds and a median end-to-end response time of 2.2 seconds. Those figures make the useful point that generation is only one component of the latency a user feels — but the announcement does not isolate what makes up the rest of the 2.2 seconds, and the gap should not be assigned to network time, startup or anything else specific.

The deadline in this launch is real

Two dates in the launch materials are firm. Mercury 3 is available today, 19 October 2026, through Inception's API and through OpenRouter, Baseten and Vercel, and it is OpenAI API compatible, so testing it against an existing workload is a configuration change rather than a rewrite — though availability through four providers does not imply identical pricing or performance on each. Inception's quick-start documentation states that every new account includes 100 million free tokens, with no payment details required; the documentation does not restrict the allocation to Mercury 3 and states no expiry. That makes an independent speed and quality check cheap to run this week.

The second date is a cost of the first. Mercury 2 and Mercury 2.5 stopped being available to new users at launch, and existing customers with usage before 19 October keep access only until 15 November 2026; enterprise customers remain supported under their agreements. Anyone in production on the older models has under four weeks to validate a replacement and migrate. The launch discount, by contrast, has no stated end date in any source checked: launch rates are half the list rates — $0.10 per million input tokens and $0.35 per million output as of 19 October 2026, with cached input at $0.01 — but a halved rate card is not a halved bill, since total spend still depends on token usage and caching. The deadline that pressures existing customers is precise. The one that would help new customers plan is missing.

Sources

  1. [1] Inception (primary), "Diffusion at Scale: Introducing Mercury 3", official blog, 19 October 2026 — 1,068 tokens per second claim, 4–8x wording, benchmark table read directly by ALFW (output speeds, GDPval Elo, GPQA Diamond and all twelve quality rows, all three models at medium settings), 40x size and 3x intelligence claims and their baselines, scaling-chart quality footnote, 1M-token context, pricing and launch discount, Mind2Web 50-task / 50% result against GPT-6.1-Sol, Octen median generation 0.9s and end-to-end 2.2s, availabilityRead source
  2. [2] Inception (primary), "Our Models", official models page, checked 19 October 2026 — "comparable quality" wording with GPT-6 Luna (medium) and Claude Haiku 5.5 (medium), Mercury 3 list and launch pricing including cached input $0.02 / $0.01, Mercury 2 and 2.5 sunset terms with the 15 November 2026 date and the enterprise exceptionRead source
  3. [3] Inception via Business Wire (primary release), "Inception Launches Mercury 3, Setting a New Pareto Frontier in Speed and Intelligence", 19 October 2026 — "outperforms" and 4–8x wording, "widely available NVIDIA GPUs", fourth release since Mercury 1 in June 2025, backers and availabilityRead source
  4. [4] Inception (primary), API quick-start documentation, checked 19 October 2026 — every new account includes 100 million free tokens, no payment details required, assigned per new user; no Mercury 3 restriction or expiry statedRead source
  5. [5] Unite.AI, "Inception Releases Mercury 3, Its Flagship Diffusion LLM", 19 October 2026 — secondary summary used only to cross-check the release details (model ID mercury-3, API compatibility, partner names); no claim in the article rests on it aloneRead source