Inception: Mercury 2.5 Preview

The inception/mercury-2.5-preview row was gone from the OpenRouter catalog in the 2026-09-09 fetch, seven days after the change feed recorded its arrival, and the same fetch recorded the arrival of inception/mercury-2.5. OpenRouter dropped the preview row, and at the catalog level what replaced it is a rename: the graduation case. Inception released Mercury 2.5 as a production model on 2026-09-08, OpenRouter's created field puts the production-named row at 2026-09-08T18:28:57Z the same day, and the preview's old route now redirects to that successor. The old name still leads somewhere. Whether it leads to the same weights the preview served, neither publisher says.

Inception said so plainly. Introducing Mercury 2.5, filed under Product, dated Sep 8, 2026 and signed by CEO Stefano Ermon, opens "Today, we’re releasing Mercury 2.5, our most capable production model yet." The word preview survives twice in that post and neither time is about this model: "Alongside Mercury 2.5, we’re announcing a preview of Mercury Voice and Mercury Router." Neither Inception's models page nor its platform documentation offers mercury-2.5-preview as a model id. The docs' request examples name the identifier mercury-2.5; searched in full, the two pages' markup contains the preview string only twice, both times on the models page and both inside a link out to the OpenRouter listing, and neither page uses the word preview in its text. Both fetched 2026-09-13.

OpenRouter's records are more specific, because they keep the two listings apart instead of collapsing them. Fetched 2026-09-13, openrouter.ai/inception/mercury-2.5-preview answers HTTP 307 to /inception/mercury-2.5, and …/models/inception/mercury-2.5-preview/endpoints returns HTTP 200 whose payload is the successor's — "id":"inception/mercury-2.5", one Inception endpoint, "name":"Inception | inception/mercury-2.5-20260908". The preview's own dated record is still reachable at …/models/inception/mercury-2.5-preview-20260831/endpoints and answers "id":"inception/mercury-2.5-preview" with "endpoints":[]: the record survives, the supply behind it does not. The catalog itself no longer carries this row id in any form: as observed on 2026-09-12, its 445 rows hold inception/mercury-2 and inception/mercury-2.5 and nothing under the preview slug, and a direct fetch of the live catalog on 2026-09-13 found the same. This entry keeps its binding to the withdrawn row regardless, so the name binds straight back here if it is ever a catalog row again.

Whether the two listings ran the same weights is answered by neither publisher, so it is not answered here. Their canonical slugs date them eight days apart, inception/mercury-2.5-preview-20260831 against inception/mercury-2.5-20260908, and on the fields that describe the model the two rows match: the same context length, the same output cap, the same parameter list, the same prices, and the same description, which the successor's page carries word for word down to "achieving 1,107 tokens/sec on standard GPUs". The copy did not change. It changed rows.

The preview week is worth keeping on the record, because the mechanism is the message. An autoregressive model emits one token at a time, while Inception's listing says Mercury 2.5 "produces and refines multiple tokens in parallel". That copy is the vendor's own — the OpenRouter description is Inception's text — and llm-releases, fetched 2 September 2026, files the same figures as claims: "vendor-reported figures are claims until independently verified". This page records the claim; it does not record a measurement of it. OpenRouter's page carried a different number for the same row while the row existed: P50 throughput of 359 tok/ssource, accessed 2026-09-02 across its provider, with a P50 latency of 1.37 ssource, accessed 2026-09-02, read off the page on 2 September 2026. The two figures are not the same quantity — a vendor capability claim against a traffic-derived median — and OpenRouter computes the median over a rolling 30-minute window of live traffic, so it never read the same way twice.

The durable facts are the row's. It was released 2026-08-31source, accessed 2026-09-02 as a preview, and its last-known listing, as of 2026-09-08, carried 260,000 tokens of context and a 65,536-token maximum output, text in and text out — API-only, via Inception's platform or the OpenRouter row, with no public weights. Inception's models page prices Mercury 2.5 at $0.20 per Mtoksource, accessed 2026-09-02 input and $0.75 per Mtoksource, accessed 2026-09-02 output — against Mercury 2's $0.25per million tokensopenrouter-models, last checked 2026-09-14 and $0.75per million tokensopenrouter-models, last checked 2026-09-14.

The discount outlived the window that was supposed to end it. The preview listing's banner, read on 2 September 2026, said "Limited-time 80% discount via Inception through September 8, 2026 at 07:00 UTC", which this page records as 2026-09-08 07:00 UTCsource, accessed 2026-09-02. The stated expiry passed, and the preview row was gone by the next day's fetch: OpenRouter puts the successor row's creation at 2026-09-08T18:28:57Z, eleven and a half hours after that expiry, and the successor carries the same 80% cut. Inception's announcement states both rates — "Price: $0.20 per million input and $0.75 per million output. At launch, Mercury 2.5 is 80% off at $0.04 per million input and $0.15 per million output" — the same promotional pair the withdrawn row carried when it was last seen. Fetched 2026-09-13, the successor's OpenRouter page shows an "80% off" badge with the list rate struck through, its payload carries "discount":0.8, and Inception's own models page and docs table show the same strikethroughs. No end date accompanies any of them. Searched in full, the markup of all three pages contains no "Limited-time", no "expires" and no "until"; the only written-out date on the successor's page is its release date, and the only one on Inception's models page is the Framer publish stamp in its first line. An introductory window on a preview row and a standing rate on a production row are different things to budget against, and this one quietly became the second.

The positioning was Inception's, not measured. The listing's opening claim is that Mercury 2.5 "is the fastest reasoning LLM"; it goes on to claim "a 10+ point jump in intelligence over Mercury 2" and "comparable quality to cost-optimized frontier models like GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite, and Claude Haiku 4.5"; the models page repeats the comparison, and the vendor's homepage adds "sub-300ms time to first token, 5-7x higher throughput, and up to 70% lower cost per task" — both fetched 2 September 2026. The snapshot carried no benchmark indices for this row and the vendor publishes no scores, so nothing here is independently verified. The listing named the production targets — search agents, voice pipelines, and coding subagents — and the capabilities it claimed the architecture enables: tunable reasoning levels, parallel tool calls, and schema-aligned JSON output, the last two checkable on the row, which accepted tools and response_format.

Preview was the status from 31 August to 8 September, and it carried three caveats. The documented developer path was text-only: the row's modalities were text-to-text, and the aimadetools explainer, fetched 3 September 2026, is explicit that "the currently documented developer path is a text model". The endpoint is OpenAI-compatible — the same writeup shows the model called through the OpenAI SDK against https://api.inceptionlabs.ai/v1 — and Inception's homepage says its models are "OpenAI API compatible and a drop-in replacement for traditional LLMs". And the identifier is not a contract: the writeup told developers to confirm the direct model identifier in account documentation "before shipping; preview aliases can change". Five days later it changed. The old slug still resolves — OpenRouter answers the model path under it, and Inception's own models page, republished 2026-09-09, still links to it — but what it resolves to is the successor's record, not a listing of its own. That is the distinction the caveat was about, and it is the one worth carrying forward: a preview alias is a pointer, and a pointer is not a promise about what sits behind it.