Inception released Mercury 3, a diffusion-based language model, on 19 October 2026, and says it outperforms OpenAI’s GPT-6 Luna and Anthropic’s Claude Haiku 5.5 “while running 4–8x faster at 1,000+ tokens per second.” Co-founder Stefano Ermon announced it on X, and the company’s blog lists a price of $0.20 per million input and $0.70 per million output tokens, with a 50% launch discount.
Every benchmark and speed figure below is Inception’s own, including the speeds it gives for the rival models. Rival prices are from OpenAI’s, Anthropic’s, MiniMax’s and DeepSeek’s own pricing pages, read on 19 October 2026.
Business Pill · COST PER TASK
A short explainer on cost per task: paying for the finished job rather than the token. It teaches the general idea only and says nothing about any company in this story.
The key insight: As we read it, Inception is pricing on speed as much as on quality: by its own figures Mercury 3 runs about eight times faster than GPT-6 Luna, while its launch output price of $0.35 sits 30% below the $0.50 OpenAI and Anthropic list for their fast models, by our arithmetic.
Ermon wrote: “Today, we’re releasing Mercury 3. Mercury 3 is by far our most powerful model and the most intelligent diffusion LLM ever released. It outperforms GPT-6 Luna and Haiku 5.5 on quality while running at 1000+ tok/sec.”
Inception’s blog lists the speed as “1,068 tokens per second on NVIDIA GPUs” and the context window as “1M tokens, up from 260K for Mercury 2.5.” It says the model is on par with or better than “frontier speed-optimized models like GPT-6 Luna and Claude Haiku 5.5.”
The company says Mercury 3 is available on the Inception API Platform and through OpenRouter, Baseten and Vercel, and that it is “OpenAI API compatible, so it drops into your existing stack with no rewrites.” OpenRouter’s models list shows a 1,000,000-token context and the launch price.
In Inception’s Business Wire release, Ermon is quoted: “We’re showing that the limits to scaling diffusion models are compute and data, not fundamental algorithmic barriers. We’ve already started training our next model, and we expect to continue pushing intelligence forward without sacrificing speed.”
The release describes the difference this way: “Unlike traditional autoregressive models, which generate text sequentially, one token at a time, diffusion models can generate multiple tokens in parallel, offering a fundamentally different path to faster inference.” It calls Mercury 3 Inception’s fourth model release.
Price per million output tokens on each provider’s own page, read on 19 October 2026: Mercury 3 at its 50% launch price and its list price (Inception’s blog), GPT-6 Luna (OpenAI), Claude Haiku 5.5 (Anthropic), MiniMax M3.1 Flash (MiniMax) and DeepSeek V4.1 Flash at peak (DeepSeek). Speeds are Inception’s figures, including those it gives for GPT-6 Luna and Claude Haiku 5.5.
What Was Posted on X
1/4 Today, weโre releasing Mercury 3. Mercury 3 is by far our most powerful model and the most intelligent diffusion LLM ever released. It outperforms GPT-6 Luna and Haiku 5.5 on quality while running at 1000+ tok/sec.
Stefano Ermon, Inception’s co-founder, opened the launch thread with this post on 19 October, saying Mercury 3 “outperforms GPT-6 Luna and Haiku 5.5 on quality while running at 1000+ tok/sec.” The benchmark and speed figures in this piece are Inception’s own.
are diffusion LLMs taking off?
they say it’s over 40x larger than Mercury 1
if Mercury 1 was something like 7B, then Mercury 3 would be > 280B
The account Scaling01 quoted Ermon’s post, wrote “they say it’s over 40x larger than Mercury 1” and speculated that Mercury 3 would be above 280B if Mercury 1 was around 7B. Inception’s blog says “more than 40x larger than Mercury 1” but publishes no parameter count, so the 280B figure is the poster’s speculation, not a disclosed size.
The Price and the Launch Discount
The blog lists “$0.20 per million input and $0.70 per million output tokens.” It adds that at launch “Mercury 3 is 50% off at $0.10 per 1M input and $0.35 per 1M output” and that cached tokens are 90% off the input price. Ermon’s post calls the launch price “just $0.1 per 1M input and $0.35 per 1M output.”
OpenAI lists gpt-6-luna at $0.10 input and $0.50 output on standard, short-context pricing, and Anthropic lists Claude Haiku 5.5 at $0.10 input and $0.50 output. By our arithmetic, Mercury 3’s launch output price is 30% below those two, and its full list output price is 40% above them.
MiniMax lists M3.1 Flash at $0.60 per million output tokens, and DeepSeek lists V4.1 Flash at $1.20 at peak and $0.60 off-peak. The chart sets those against both Mercury 3 prices.
Inception’s Own Benchmark Table
Inception published a table of twelve benchmarks comparing the medium settings of Mercury 3, GPT-6 Luna and Claude Haiku 5.5. By our count, Mercury 3 has the top score on eight of them, and GPT-6 Luna and Claude Haiku 5.5 lead on two each.
In the table, Mercury 3 scores 51.08 on SWE-Bench Atlas against 33.33 for GPT-6 Luna and 32.26 for Claude Haiku 5.5, and 89.23 on GPQA Diamond against 86.36 and 84.51. GPT-6 Luna leads on AA-Omniscience accuracy (46.94 against 46.44) and AA-LCR v1.1 (78.30 against 77.67).
Claude Haiku 5.5 leads on MRCR v2 (78.40 against 75.50) and DeepSWE v1 (51.62 against 46.31). The same table gives output speeds of 1,068 tokens per second for Mercury 3, about 130 for GPT-6 Luna and about 240 for Claude Haiku 5.5; by our arithmetic that is about 8.2 and 4.5 times faster.
The 40x Claim
Ermon wrote that “Mercury 3 demonstrates that text diffusion scales. At over 40x the size of Mercury 1, it’s vastly more intelligent while remaining just as fast.” The blog says Inception scaled Mercury “more than 40x in total parameters” and that intelligence across standard benchmarks “has risen 3x from Mercury 1 (mini) to Mercury 3.”
Inception does not give a parameter count for Mercury 3 or for Mercury 1. A post by the account Scaling01 speculated that if Mercury 1 was around 7B, Mercury 3 would be above 280B; that is the poster’s speculation, and Inception has not published a size.
Who Is Using It
Ermon wrote that early adopters include TinyFish, Lorikeet, RevisionDojo, Bauplan, Octen and Croux. In Inception’s post, TinyFish co-founder Shuhao Zhang says Mercury 3 completed 50% of 50 live Mind2Web navigation tasks, “significantly more than GPT-6.1-Sol.”
Octen founder Kuan Zou is quoted as saying Mercury 3 delivered “a median generation time of 0.9 seconds and a median end-to-end response time of 2.2 seconds” in the company’s Sense API tests. These are customer statements published by Inception.
The Structural Read
As we read it, the comparison is drawn in the fast tier, not against frontier models: Inception benchmarks Mercury 3 against GPT-6 Luna and Claude Haiku 5.5, the models OpenAI and Anthropic price at $0.50 per million output tokens.
As we read it, the launch discount carries much of the pricestory: at the $0.70 list price Mercury 3’s output costs 40% more than those two models, by our arithmetic, while at $0.35 it costs 30% less, and Inception does not say how long the discount runs.
As we read it, the customer quotes Inception chose stress latency in agent loops: TinyFish describes multi-step web navigation and Octen a smart-glasses search with a 0.9-second median generation time, workloads where waiting compounds with every step.
Stefano Ermon, Inception co-founder, in the Business Wire release, 19 October 2026
“Weโre showing that the limits to scaling diffusion models are compute and data, not fundamental algorithmic barriers.”
Three Implications
A SPEED CLAIM BUILT ON ITS OWN FIGURES The 1,068 tokens per second for Mercury 3 and the roughly 130 and 240 for GPT-6 Luna and Claude Haiku 5.5 all come from Inception’s own table.
A PRICE THAT DOUBLES AFTER LAUNCH Inception lists $0.70 per million output tokens and offers $0.35 at launch; OpenAI and Anthropic list $0.50 for their fast models.
A SIZE CLAIM WITHOUT A SIZE Inception says Mercury 3 is more than 40 times the size of Mercury 1 but publishes no parameter count; the 280B figure in a Scaling01 post is that poster’s speculation.
The framework opens: “AI has crossed from the training era to the inference era. This is not a technical nuance. It is a fundamental economic restructuring that changes who makes money, how they make it, and where value concentrates in the AI economy”.
As we read it, Mercury 3 is a bet on the inference side of that shift: Inception competes on tokens per second and price per token rather than on frontier capability, a bid for a share of where, in the framework’s words, value concentrates. The framework was published in March 2026, before Mercury 3, and does not mention Inception or diffusion models.
What Is Not Established
The benchmark scores, the speeds for all three models and the customer results are Inception’s own figures and quotations. A Google News search on 19 October turned up no third-party benchmark of Mercury 3, and Inception’s materials carry no comment from OpenAI or Anthropic.
Inception does not say how long the 50% launch discount lasts. The chart uses list prices read on each provider’s page on 19 October 2026, and the percentages in this piece are our arithmetic.
Business Engineer Framework
The State of The Inference Economy
A Business Engineer analysis of how the shift from training to inference changes who makes money in AI, and how.
Inception released Mercury 3 on 19 October 2026 and says the diffusion model outperforms GPT-6 Luna and Claude Haiku 5.5 on quality at 1,068 tokens per second, more than 40 times the size of Mercury 1. It lists $0.70 per million output tokens, cut to $0.35 at launch, against $0.50 for OpenAI’s and Anthropic’s fast models.
A note on sourcing. This piece is built on Inception’s own blog post, benchmark table and Business Wire release and on co-founder Stefano Ermon’s X thread, all of 19 October 2026. Benchmarks, speeds (including those for rival models) and customer results are Inception’s figures. Rival prices were read on OpenAI’s, Anthropic’s, MiniMax’s and DeepSeek’s own pages on 19 October. Percentages, speed multiples and the benchmark count are our arithmetic. Nothing here is a forecast, and nothing here is financial or investment advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
