Watch: how diffusion writes differently
Standard models write left to right, one token at a time. Mercury 3 drafts the whole answer at once, full of blanks, then fills in every blank at the same time, over a few refinement passes. Run the illustration to see the difference.
Autoregressive
most models: GPT-6 Luna, Haiku 5.5Diffusion
Mercury 3Illustration, not a benchmark. The animation is slowed so you can follow it; the real gap is far wider. Mercury 3's reported 1,068 tokens/sec would finish this sentence in about 0.02 seconds. Timing and the refinement passes are symbolic.
Cost calculator: Mercury 3 vs its rivals
Mercury 3 is priced per token. The launch deal halves the list price, and cached input tokens cost 90% less than regular input. Inception has not published an end date for the launch pricing, and CEO Stefano Ermon told The Deep View the company is trying to make it permanent.
| Model | Pricing used | Cost per request | Est. monthly (30d) |
|---|
Notes: Haiku 5.5 keeps its $0.10 input rate only for prompts up to 100,000 tokens; above that it rises to $0.50. GPT-6 Luna numbers are the short-context tier; its long-context tier is higher. Mercury 3 launch pricing has no announced end date. All figures are estimates from list prices.
Mercury 3: the facts at a glance
| Fact | Detail | Source |
|---|---|---|
| Announced | October 19, 2026 | Inception press release |
| Reported speed | 1,068 tokens/sec on NVIDIA GPUs; claimed 4–8× faster than GPT-6 Luna and Claude Haiku 5.5 | Company claim; no independent test published |
| Context window | 1 million tokens (Mercury 2.5 had 260,000) | Inception press release |
| Scale | More than 40× the parameters of Mercury 1 | Inception press release |
| List price | $0.20 per 1M input / $0.70 per 1M output tokens; cached input 90% off ($0.02) | Inception press release |
| Launch price | $0.10 per 1M input / $0.35 per 1M output; cached input $0.01 (50% off list), no end date announced | Inception press release |
| Intelligence claim | On par with or better than GPT-6 Luna and Claude Haiku 5.5 across standard benchmarks | Company claim; independent The Deep View interview |
| Access | Inception API, OpenRouter, Baseten, Vercel; OpenAI API compatible | Inception press release |
| Open weights | Not offered at launch; API only | Launch announcement |
| Mercury 2 / 2.5 | Closed to new users; non-enterprise access ends November 15, 2026 | Superpower Daily |
Why diffusion, and why people doubted it
When Inception shipped Mercury 1 in June 2025, it was the first commercial diffusion LLM. Before that, diffusion language models were mostly an academic project, and early ones handled simple jobs like code completion. The question everyone asked: can parallel generation stay fast once the model gets smart? A 40×-bigger model answering at 1,000+ tokens per second is Inception's answer.
“Since we launched Mercury 1, the question we’ve heard most is whether diffusion can scale,” co-founder Stefano Ermon said in the release. “We’re showing that the limits to scaling diffusion models are compute and data, not fundamental algorithmic barriers.” The company says it has already started training the next model.
The practical argument is about agent loops. An agent calls the model many times to finish one task, so a small delay at each step multiplies. Inception's pitch: Mercury 3 keeps each pass fast, so agents finish sooner and can try more approaches before answering. The Deep View also reports Inception ran swarms of Mercury 3 agents for recursive self-improvement, gaining another 16% in speed, a claim worth watching, not yet replicated anywhere independent.
- Mercury 1, June 2025First commercial diffusion LLM; the proof that parallel text generation could ship.
- Mercury 2Added reasoning to diffusion models.
- Mercury 2.5Tunable reasoning; context window grew to 260,000 tokens.
- Mercury 3, October 19, 2026Biggest intelligence jump of any Mercury release; 1M-token context; still 1,000+ tokens/sec.
Where the speed is already showing up
- Tutoring. RevisionDojo (General Learning) gives every student instant feedback: hints, marked answers and definitions come back in about a second, with quality “as good as or better than the models we ran before,” says co-founder and CTO Michael Tong. It was a drop-in swap for their existing stack.
- Search agents. Octen AI reports 0.9-second median generation and 2.2-second median end-to-end responses, with search taking about 80 milliseconds.
- Web agents. TinyFish AI says Mercury 3 completed 50% of 50 live Mind2Web navigation tasks in its production web-agent setup.
All three examples come from Inception's launch materials, so treat them as company-reported until independent tests appear.
If you run Mercury 2 or 2.5 today: they are closed to new users, and non-enterprise access ends November 15, 2026. The path Inception offers is Mercury 3 or an enterprise plan with dedicated capacity, autoscaling, compliance controls, and configurable data retention.
How to try it
Mercury 3 is available through the Inception API, OpenRouter, Baseten, and Vercel. It accepts the OpenAI API format, so existing code usually needs only a base URL and key change. Links: official announcement, Unite.AI summary, The Deep View CEO interview.
Related pages on this site: GPT-6 Luna pricing calculator, Claude Haiku 5.5 pricing and migration guide.
FAQ
What is Mercury 3?
Mercury 3 is the fourth diffusion language model from Inception, announced October 19, 2026. It generates and refines many tokens in parallel instead of writing word by word. Inception reports 1,068 tokens per second on NVIDIA GPUs and a 1-million-token context window.
How much does Mercury 3 cost?
List price: $0.20 per million input tokens, $0.70 per million output tokens. At launch it is 50% off ($0.10 input, $0.35 output). Cached input tokens are 90% off the input rate ($0.02 list, $0.01 at launch). Inception has not announced when launch pricing ends.
How fast is Mercury 3?
Inception reports 1,068 tokens per second on widely available NVIDIA GPUs and claims 4–8× faster inference than GPT-6 Luna and Claude Haiku 5.5. Those are the company's own numbers; no independent benchmark has been published yet.
What is a diffusion LLM?
A standard language model writes left to right, one token at a time. A diffusion model starts with a full draft of masked tokens and refines all positions in parallel over a few passes. Parallel refinement is why diffusion models can be much faster.
Is Mercury 3 open weights?
No. Mercury 3 is API-only at launch: the Inception API, OpenRouter, Baseten, and Vercel. Inception did not announce open weights.
Where can I use Mercury 3?
Through the Inception API, OpenRouter, Baseten, and Vercel. It accepts the OpenAI API format, so existing code usually needs only a base URL and key change. Enterprise plans add dedicated capacity, autoscaling, compliance controls, and configurable data retention.
Mercury 3 vs GPT-6 Luna vs Claude Haiku 5.5?
Inception claims Mercury 3 matches or beats both on quality while running 4–8× faster. On price at launch, Mercury 3 is $0.10/$0.35 per million input/output tokens vs $0.10/$0.50 for both rivals' short tiers, about 30% less on output. Quality claims are vendor-reported.
What happens to Mercury 2 and 2.5?
Both are closed to new users, and existing non-enterprise access ends November 15, 2026. Move to Mercury 3 or take an enterprise plan.
Sources
- Inception press release, October 19, 2026, Business Wire
- Unite.AI: Inception Releases Mercury 3 (1,068 tok/s, $0.10/$0.35 launch pricing, 1M context)
- Superpower Daily (migration deadline Nov 15, 2026; Octen AI and TinyFish AI customer results)
- The Deep View: Why Inception's Mercury 3 upends AI economics (Ermon interview; +16% self-improvement claim)
- Interesting Engineering (8x speed claim; Ermon X post)
- AI Brief (pricing summary, OpenAI API format)
Ask an AI about this guide:
Agent-readable summary: /mercury-3/llms.txt