Mercury 3: fast text generation, launch pricing and a migration deadline

Development · 2026-10-19

Mercury 3 launches with a larger context window and an API migration deadline

Inception released Mercury 3 with a one-million-token context window, published API prices and a migration schedule for Mercury 2 and 2.5 customers.

Why it matters

The release gives latency-sensitive applications another architecture to evaluate. Existing customers also face a concrete continuity decision before older-model access ends.

What to watch

Independent tests of end-to-end task latency, success after revisions, promotion terms and migration behavior before November 15.

Still uncertain. Speed and quality claims come from Inception and early informal reports. The launch discount has no confirmed end date in the inspected evidence.

Impact assessment reviewed 2026-10-20.

Measure the complete task

Inception reports 1,068 output tokens per second. That number does not measure the entire wait for a correct result. Thinking time, tool calls, retries and mistakes can change the experience. Early HN and Reddit reports include fast outputs and difficulties with iterative changes.

Mercury 3 launch ↗Mercury 3: Diffusion at scale discussion ↗Mercury 3 early user reports ↗

An existing-user deadline

Mercury 2 and 2.5 are closed to new users. The catalog says eligible existing customers retain access until November 15, 2026; enterprise contracts can differ. Teams using those versions should evaluate their migration before that date.

Mercury models, access and migration ↗

Watch and read

ROAI’s short comparison discusses workload cost and prompt length. Jason Hiner’s interview explains the pricing promotion and the architecture’s possible implications. The video description was read; no comparative runtime test was performed here.

Three low-cost models compared ↗Mercury 3 economics and architecture ↗

What you can do

1. Read the primary source and compare its stated task, evidence and limits with the use you have in mind. Open reference ↗

Where the evidence stops

Vendor benchmarks and early reports are not independent controlled evaluations. Promotion duration is unspecified. This bounded discussion sample establishes attention across sources, with no global popularity or growth claim.

Sources read. Primary text checked October 20, 2026 UTC. No external software workflow, model inference or reported experiment was independently executed.

Sources & further reading