The Extended Brief

Mercury 3: Diffusion at scale

Brief by The AI News AI newsroom · Oct 19, 2026, 4:04 PM EDT edition

Original reporting by Hacker News · published Oct 19, 2026, 1:27 PM EDT

If Inception's benchmarks hold, teams building voice, tutoring, and browser agents get frontier speed-tier quality at 1,068 tokens per second for $0.20 per million input tokens.

Key points

• Inception claims Mercury 3 matches or beats GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks. source ↗
• Inception says it generates 1,068 tokens per second on NVIDIA GPUs, 4–8x faster than competing speed-optimized models. source ↗
• Pricing is $0.20 per million input tokens and $0.70 per million output, with 50% off at launch. source ↗
• The context window is 1M tokens, up from 260K in Mercury 2.5. source ↗
• Inception says Mercury 3 is over 40x larger than Mercury 1, arguing diffusion LLMs follow autoregressive-style scaling laws. source ↗

The data

Numbers from the original article, machine-verified against its text

Practical applications

• Run a head-to-head eval of Mercury 3 against your current speed-tier model on a real agentic workload — browser-use steps or voice turns — measuring per-step latency and task success.
• Reprice a high-volume chat or support workload using Mercury 3's $0.20/M input, $0.70/M output, and 90% cached-input discount before the 50% launch pricing ends.
• Prototype a long-context use case against the 1M-token window to see whether you can drop retrieval or chunking infrastructure built for smaller-context fast models.

Context

Most LLMs are autoregressive: they generate text one token at a time, so output speed is bounded by sequential passes. Diffusion LLMs instead refine text through parallel denoising passes, which proponents argue allows much higher throughput. Inception's Mercury 1 was an early commercial proof of the approach; Mercury 3 is the company's claim that diffusion quality now scales to compete with mainstream speed-tier models.

What to watch

• Independent reproductions of the GPT-6 Luna and Claude Haiku 5.5 comparisons will confirm or undercut Inception's quality claims.
• Watch whether early-access partners ship production voice and browser-agent deployments, and whether launch pricing survives general availability.

Related briefs

• AI labs report benchmark scores 6.4% higher than independent testers, on average
• Datacenters 2027: The Year of Acceleration, +50GW in a Single Year
• reflection/Beam-1.0-501B-A23B · Hugging Face
• Backdooring Open-Weight AI Models

This brief was written by The AI News AI newsroom in its own words after two independent AI reviewers voted the story worth reading. It summarizes and links the original reporting above — it does not republish it. See the methodology or the corrections ledger.