I've been bullish on diffusion language models for more than a year. This week, Inception Labs gave me another reason to double down.
On October 19, the company launched Mercury 3, its most capable model yet. It claims quality comparable to, and on several benchmarks better than, GPT-6 Luna and Claude Haiku 5.5, while generating text at a blistering 1,068 tokens per second on NVIDIA GPUs.
That's a combination worth taking seriously. Diffusion models have offered startling speed for a while, but they've generally had ground to make up on intelligence. With Mercury 3, that gap looks a whole lot smaller.
And Inception has been moving quickly. Mercury first arrived in February 2025. Mercury 2 followed in February 2026, bringing reasoning and 1,000-plus-token-per-second generation. Mercury 2.5 launched September 8. Just 41 days later, we've got Mercury 3.
This team is cooking with gas.
Why diffusion matters
Most familiar chatbots use an autoregressive approach: they produce a response one token at a time. Each new token depends on what came before it. Diffusion takes a different route, generating and refining multiple pieces of a response in parallel.
Think of the difference between a writer typing a sentence from left to right and an editor revising several parts of a draft simultaneously. Both approaches still require real computation, and real-world speed depends on hardware and serving conditions. But diffusion changes the underlying trade-off.
The bigger question has always been whether diffusion could keep that advantage as models got larger and more capable. Speed is wonderful. Speed married to genuinely useful intelligence is where things get interesting.
Inception says Mercury 3 has more than 40 times the total parameters of Mercury 1, with a roughly threefold improvement on its aggregate intelligence measure across standard benchmarks, while retaining similar generation throughput. Its launch post makes an ambitious claim: "The scaling laws that drove progress in autoregressive models hold for diffusion too."
That's the part of the announcement that really grabbed me. If further independent research confirms this trajectory, diffusion has a much stronger claim to becoming a lasting alternative in mainstream language AI.
Mercury 3: The numbers
Here are the specifications Inception has published:
- Generation speed: 1,068 tokens per second on NVIDIA GPUs, measured by Inception.
- Context window: 1 million tokens, up from 260,000 with Mercury 2.5.
- Standard API pricing: $0.20 per million input tokens and $0.70 per million output tokens.
- Launch discount: $0.10 per million input tokens and $0.35 per million output tokens, with cached input at $0.01 per million tokens during the promotion.
- Developer features: reasoning controls, tool calling, structured outputs, and OpenAI-compatible API integration.
Inception's published comparison chart shows Mercury 3 ahead of both GPT-6 Luna and Claude Haiku 5.5 on tasks including SWE-Bench Atlas, BrowseComp+, DeepSearchQA, and GPQA Diamond. For example, it reports 55.18 on BrowseComp+ versus 50.50 for Luna and 45.78 for Haiku 5.5. The model doesn't win everything: Haiku leads its DeepSWE v1 and MRCR v2 comparisons, and Luna edges Mercury on one long-context evaluation.
Those are Inception's benchmark results, not an independent leaderboard verdict. Still, the overall picture is impressive. And the more practical claim is easy to understand: a fast model that can perform many of the everyday tasks handled by today's smaller frontier models, often at a remarkably low price.
I put Mercury 3 through eight API runs. Here's what happened.
I tested Mercury 3 through OpenRouter, using writing, business arithmetic, explanation, coding, hallucination resistance, and logic questions. I also ran variations of an instruction-following exercise. The total reported API cost for all eight runs was $0.04160. Yes, about four cents.
A few of the results stood out.
Creative writing: entertaining, but sloppy with constraints. I asked for a 250-word science-fiction story about a maintenance machine becoming an accidental ruler, with only two speaking characters, a one-room setting, forbidden words, and a final twist. Mercury gave me a genuinely funny story about a machine worshipped by roaches. But the story ran around 274 words and drifted beyond the one-room setting. The creative spark was there; the constraint checking wasn't perfect.
Business math: excellent. I handed it a coffee-shop problem involving prices, costs, a supplier discount, and a monthly membership fee. Mercury correctly calculated $900 in daily gross profit, $756 in monthly gross savings, and $576 in net monthly benefit. Better yet, it challenged the assumptions, calculated a break-even point, and explained the decision plainly. One minor nit: it confused a higher profit per premium cup with a higher percentage margin. The central recommendation was sound.
Explaining AI: strong, with familiar oversimplifications. I asked it to explain how language models reason to a child, a nontechnical adult, and a software engineer. The responses were generally accessible and thoughtful, with some overly confident claims about what models can or cannot understand. On a stricter follow-up, it hit exactly 50 and 150 words as requested. That version was substantially more expensive than a looser version, an interesting reminder that precision and additional reasoning can change the cost of a request.
Coding: the most impressive demonstration. I asked for a complete offline, single-file HTML dungeon crawler, including a randomized 25-by-25 map, multiple connected rooms, melee and ranged enemies, line-of-sight rules, potions, gold, and a restart button. Mercury produced the actual game code. The JavaScript passed a syntax check, and I ran 2,000 randomized map-generation checks with zero failures for the room-count, connectivity, or initial-placement conditions tested. That isn't the same as proving every combat interaction works perfectly during a full playthrough, but it's a strong first-pass result for a fairly ambitious prompt.
Hallucination resistance: a qualified pass. I asked about a deliberately fictitious NASA mission in 1987 involving astronaut Daniel Mercer and a nonexistent Helios-7 probe. Mercury spotted the fabrication and refused to invent answers. However, it used three web searches, so this was a tool-assisted result, and it introduced an unrelated error by claiming there were no other Shuttle launches in 1986–1988. That's a useful caution: even a model that correctly rejects a false premise can add incorrect supporting details.
Logic: clean and cheap. Faced with the classic three-mislabeled-fruit-box puzzle, Mercury chose the correct box, explained both possible draws, and stayed within the word limit. Cost: $0.00014. It's a well-known puzzle, so I wouldn't use it alone as evidence of groundbreaking reasoning, but it handled the task flawlessly.
These were practical editorial spot checks rather than a controlled benchmark study. I didn't run the same tasks repeatedly across temperatures, measure latency under matched hardware, or independently reproduce Inception's 1,068-token-per-second figure. The API experience felt fast, but company-reported peak throughput and what any given user sees through a third-party provider can differ considerably.
The agent story may matter even more than the chatbot story
I suspect Mercury 3's most important audience will include developers building systems that make many small, repeated model calls. Browser agents, search tools, customer support, tutoring, code review, and other agentic workflows can become painfully slow when every step waits for another answer.
Inception is leaning directly into that. The company says Mercury 3 completed 50% of 50 live Mind2Web browser-navigation tasks in a production agent harness, outperforming GPT-6.1 Sol in that specific evaluation. Customers including TinyFish, RevisionDojo, Octen, and Lorikeet describe using it for navigation, learning, search, and support. Those are company-published tests and partner accounts, but they show where the architecture could have real impact.
Mercury 3 also arrives alongside a growing diffusion-powered lineup. Mercury Voice is a separate model optimized for low-latency voice-agent workloads, and Mercury Decide is a separate model for structured, high-volume decisions. Mercury 3 itself supports tool-based agent workflows. The broader family is becoming more versatile, though I would not assume every capability belongs natively to the same endpoint.
Why I've been watching Inception
I met Inception cofounder and CEO Stefano Ermon in Las Vegas last November. Hearing him explain what the team was building, why diffusion deserved a serious shot at language, and watching what their technology could already do made an impression on me.
I've been a believer ever since. That doesn't mean giving Inception a free pass on benchmarks or treating every claim as settled. I want to see independent comparisons, repeatable agent results, and more evidence of reliability on long, complicated tasks. But this release looks like a meaningful step toward the future Ermon was describing.
A year ago, asking whether diffusion language models could become serious competitors to traditional language models was an interesting research question. Mercury 3 makes that question feel much less hypothetical.
Should you try Mercury 3?
Absolutely. Start with Mercury Chat if you want to experience it without building anything. Developers can use the Inception API platform with 100 million free tokens for new accounts, follow the API quickstart, or try the model on OpenRouter and Vercel AI Gateway. Inception also lists Baseten as an availability partner.
My early verdict: Mercury 3 is a genuinely exciting model at a startling price. It still has rough edges, particularly around precision and the occasional overconfident explanation. But it's now delivering useful writing, reasoning, and substantial coding output in territory once dominated by autoregressive models.
I've spent more than a year arguing that diffusion deserves a bigger role in the AI conversation. After Mercury 3, I expect a lot more people will be joining it.
