Inception's Mercury 3: the diffusion LLM that beats GPT-6 Luna and Claude Haiku 5.5 at 1,068 tokens per second

• Who: Inception of Redwood City, California (co-founder Stefano Ermon), the company behind the first commercial diffusion LLMs.
• What: Mercury 3, described as the most capable diffusion LLM ever released - Inception's fourth model since Mercury 1 (June 2025).
• When: Announced October 19, 2026; available the same day via the Inception API, Baseten, OpenRouter and Vercel.
• Speed: 1,068 tokens per second on NVIDIA GPUs - 4-8x faster than GPT-6 Luna and Claude Haiku 5.5, which Inception says it beats on benchmarks.
• Scale: Over 40x the size of Mercury 1, with a 1-million-token context window (up from 260,000 on Mercury 2.5).
• Price: Launch pricing $0.10 per 1M input / $0.35 per 1M output tokens (50% off list $0.20/$0.70); cached input $0.02 ($0.01 promo); 100M free tokens for new accounts.

REDWOOD CITY, Calif. — Inception, the startup behind the first commercial diffusion large language models, announced Mercury 3 on October 19, calling it the most capable diffusion LLM ever released. The headline claim is a two-for-one: benchmark performance on par with or better than frontier speed-optimized models like GPT-6 Luna and Claude Haiku 5.5, at more than 1,000 tokens per second — four to eight times faster.

Diffusion models generate many tokens in parallel instead of one at a time, which is where their speed comes from. The open question has been whether that parallel-generation advantage survives as the models get smarter. Mercury 3 is Inception's answer: over 40 times the size of Mercury 1, which launched in June 2025, while holding comparable speeds — 1,068 tokens per second on NVIDIA GPUs, per the announcement. The company says intelligence rose threefold from Mercury 1 to Mercury 3, evidence that the scaling laws that drove autoregressive models hold for diffusion too.

The evidence, such as it is

Every performance number here comes from Inception itself — worth keeping in mind. In a company-run web-agent test, Mercury 3 completed 50% of 50 live Mind2Web navigation tasks, which the company described as significantly more than GPT-6.1-Sol managed, using the same token volume and agent steps at a fraction of the per-token cost.

Early-access customers are testing it where latency is the product. Michael Tong, co-founder and CTO of tutoring company General Learning (RevisionDojo), said Mercury 3 lets every student get instant feedback “without giving up reasoning” — hints, marked answers and definitions arriving in about a second — and called the switch a drop-in replacement for the company's existing stack. Kuan Zou, founder and CEO of Octen AI, reported a median generation time of 0.9 seconds and end-to-end response of 2.2 seconds in smart-glasses tests, “the best balance of speed and quality” among the fast models Octen tested. Michael Gribben of Lorikeet, which builds customer concierges for regulated businesses, said Mercury 3 beat other models on latency in early testing while scoring higher on evaluations.

Price, access, and a 100M-token free tier

Mercury 3 is available today through the Inception API, Baseten, OpenRouter and Vercel, and it speaks the OpenAI API format, so teams can switch without rewriting code. The model identifier is mercury-3, served from Inception's chat-completions endpoint with a reasoning_effort parameter, and Inception ships Python and TypeScript SDKs. List pricing is $0.20 per million input tokens and $0.70 per million output tokens, halved at launch to $0.10 and $0.35; cached input tokens run 90% below the input rate. New platform accounts get 100 million free tokens with no payment details required.

A portfolio play — and a sunset for 2.x

Mercury 3 heads a small real-time-AI portfolio: Mercury Voice, tuned for sub-170-millisecond time-to-first-token in voice agents, and Mercury Decide, a decision model for high-volume classification that returns calibrated probabilities and can escalate uncertain cases to a human. Alongside the launch, Inception is retiring Mercury 2 and 2.5 for new users; existing customers keep access until November 15, 2026.

The bet

“Since we launched Mercury 1, the question we've heard most is whether diffusion can scale,” said co-founder Stefano Ermon. “We're showing that the limits to scaling diffusion models are compute and data, not fundamental algorithmic barriers.” Inception says it has already started training its next model. If Mercury 3's company-reported numbers hold up in developers' own evaluations — the caveat that matters — the diffusion-versus-autoregressive debate just got its strongest data point yet.