Inception Labs launched Mercury 3 today — a diffusion-based language model claiming 1,068 tokens per second at $0.10 per million input tokens. That is between four and eight times faster than comparable autoregressive models at a competitive price. There is a catch: Inception shipped the launch without a single independent benchmark score, the model defaults to lower reasoning effort than its predecessor, and on coding tasks it actually trails Claude Haiku 5.5. Before you reroute your agent pipeline, here is what you need to know.
Why Diffusion Is Faster #
Traditional LLMs generate tokens left-to-right, one at a time. Every token depends on all previous ones — the process is sequential and hard to parallelize. Diffusion LLMs invert this. The model starts with a fully masked sequence and runs a denoising transformer across multiple steps, predicting all positions simultaneously using bidirectional attention. High-confidence tokens commit first; uncertain positions refine in subsequent passes.
The result: a 1,000-token output takes roughly 1,500 milliseconds on a diffusion model versus 10,000 milliseconds on a comparable autoregressive one. The speedup compounds as output length grows — exactly why Mercury 3 targets voice agents and real-time search pipelines. Google validated the architectural viability independently when it shipped DiffusionGemma 26B in June. Inception’s claim that diffusion scaling laws match autoregressive ones is a milestone, not marketing noise.
The Benchmark Problem #
Here is where it gets messy. The launch post claims Mercury 3 “outperforms GPT-6 Luna and Claude Haiku 5.5” — but every number is company-reported. Artificial Analysis had no Mercury 3 entry the day after launch. There is no published test methodology, no margin data, no per-task breakdown.
There is also a hidden variable: Mercury 3 defaults to “low” reasoning effort, while earlier versions defaulted to “medium.” That change meaningfully affects output quality. If you benchmark Mercury 3 yourself, set reasoning_effort: "medium" explicitly — otherwise your numbers will not be comparable to anyone else’s, including Inception’s own claims.
The one published third-party data point cuts against the marketing: Mercury 3 scores 46% on the DeepSWE coding benchmark. Claude Haiku 5.5 scores 52%. On SWE-Bench Atlas the picture flips — Mercury 3 hits 51% versus Haiku’s 32%. The honest read: Mercury 3 wins on speed, and on some benchmarks. It does not win on coding.
What It Is Actually Good For #
Speed that compounds over long outputs is a genuine advantage for specific workloads. Inception’s Mercury Voice variant returns the first answer token in 320 milliseconds median on customer-service prompts — the latency class voice pipelines need. The same architecture powers real-time search agents, interactive tutoring, and browser automation. On Hacker News, one developer reported building a SwiftUI app in 45 minutes using 26 million tokens at 160 tokens per second — it worked well.
The pattern: Mercury 3 is the right choice when speed matters more than peak accuracy and a human or escalation path is in the loop. It is not the right choice when a wrong answer is expensive and unmonitored.
Where to Avoid It #
- Multi-step chain-of-thought reasoning. Sequential logic does not map to parallel denoising. Complex reasoning chains degrade quality.
- EU-regulated environments. Mercury 3 is not available via EU in-region routing — a compliance gap for many European deployments.
- High-stakes unmonitored agents. Quality remains below frontier models for complex tasks. Deploy without a catch layer at your own risk.
- On-premise or self-hosted setups. Weights are not released. This is a closed, rental-only model.
How to Try It #
Mercury 3 is available now through the Inception API platform, OpenRouter, Baseten, and Vercel. The API is OpenAI-compatible — swap the base URL and model name, and existing code runs without modification. New accounts get 100 million free tokens with no payment details required.
import openai
client = openai.OpenAI(
base_url="https://openrouter.ai/api/v1",
api_key="YOUR_OPENROUTER_KEY",
)
response = client.chat.completions.create(
model="inception/mercury-3",
messages=[{"role": "user", "content": "Your prompt here"}],
extra_body={"reasoning_effort": "medium"}, # pin this
)
print(response.choices[0].message.content)
Independent benchmarks will arrive within days. Until then, the honest picture is: fast, cheap, with real production deployments — but unverified quality claims and a coding gap worth testing before you commit. For latency-bound workloads with human oversight, Mercury 3 is worth a look. For anything where accuracy is the constraint, wait for Artificial Analysis to publish their numbers first.