Intel / post

Mercury 2 posts 78% on an agentic benchmark at under $1 per million tokens

Source: x.com
Date: Published Mar 24, 2026

Context

PinchBench tests a model as the brain of an built on OpenClaw, the open-source agent project, on everyday jobs such as scheduling meetings, triaging email, researching topics, managing files and writing code. Inception says Mercury 2 completed 78% of those tasks. In its blog post, that matches or beats GPT-5 Mini at 75%, DeepSeek Chat at 72%, and Gemini 2.5 Flash and GPT-4o at 71%.

That is not the top of the board. The chart attached to the post plots success rate against execution time for 50 models, and models such as Claude Opus 4.6 and GLM-5 sit higher but further to the slower side. Inception's real claim is the combination: faster than every model at comparable accuracy, at $0.25 per million input and $0.75 per million output tokens. Its argument is that agents make dozens of calls per task, so latency and cost add up at every step.

Inception attributes the speed to how Mercury 2 generates text. It is a diffusion model that refines its output in parallel instead of producing one token at a time, and the company says the speed does not depend on specialized hardware. The result comes from Inception's own run in March 2026. PinchBench has since moved to a v2 leaderboard with 61 models and 147 tasks, where it still lists Mercury 2 as the fastest complete run.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.

Read the source x.com