Wait Which Model?
← Back to directory
Inception

Mercury 3

Released Oct 19, 2026 · knowledge cutoff unpublished

Status
Unknown
Location
United States
Type
LLM
Modality
Text
Context window
1M
Max output
131K
Speed
—
Price ($/MTok in / out)
$0.20 / $0.70
Cost per task
—
Open weights
No
News1 story
Strengths & weaknesses

Strengths

  • Diffusion decoding refines whole blocks of the answer in parallel instead of token by token, so long outputs land in a burst — the gain is largest on generation-heavy calls
  • Reasoning effort is a per-call dial from instant to high, so one endpoint covers both latency-critical calls and harder ones
  • Tool calling and JSON-schema structured outputs on an OpenAI-compatible API, so it drops into existing agent tooling without an adapter

Weaknesses

  • Defaults to low reasoning effort where Mercury 2.5 defaulted to medium — out-of-the-box answers are shallower than the launch comparisons, which ran at medium, unless the setting is pinned
  • Holds up on one-shot coding questions but drifts over long multi-turn agent sessions, and because every turn still has to read the whole conversation first, the speed edge shrinks as context grows
  • Every quality claim at launch is Inception's own — no independent evaluator had measured it by the day after release

Announced 2026-10-19 by Inception and its founders on X ('Today, we're releasing Mercury 3'); the blog post's listing date reads 2026-10-15 and OpenRouter's listing 2026-10-16, so the public announcement date is used. Available the same day on the Inception API Platform, OpenRouter, Baseten and Vercel; list price is $0.20 input / $0.70 output per million tokens ($0.02 cached input), with a 50% launch promotion ($0.10 / $0.35, end date unstated). Text-only input. Inception claims about 1,068 output tokens per second on NVIDIA GPUs (its own figure, not an Artificial Analysis measurement) and a 1M-token context, up from Mercury 2.5's 260K. Inception's benchmark table (medium effort for all models) reportedly leads GPT-6 Luna and Claude Haiku 5.5 on 8 of 12 tests, including GPQA Diamond, IFBench and SWE-Bench Atlas, but the table could not be read from the primary page, so every benchmark cell is null. Mercury 2 and 2.5 (not tracked here) are closed to new users, and existing customers keep access until 2026-11-15. No Artificial Analysis coverage yet; parameter count and knowledge cutoff are unpublished.

Benchmarks0 of 10 reported

MMLU-Proretired

—

GPQA Diamond

—

SWE-bench Verified

—

SWE-bench Pro

—

Terminal-Bench 2.1

—

AIME (latest)retired

—

Humanity's Last Exam

—

LMArena Elo

—

GDPval-AA v2

—

ARC-AGI-2

—
For developers

API model strings

  • inceptionmercury-3

Licence

Proprietary — weights not released

Retirement

No retirement announced

Lineage

No recorded predecessor