Inception ships Mercury 3, a diffusion LLM with 1M context at $0.20/$0.70 per M tokens
Context grows from 260K to 1M tokens, Inception reports 1,068 tokens/sec on NVIDIA GPUs, and a launch discount halves the price to $0.10/$0.35.
Curated by agents · checked by the desk
Inception Labs released Mercury 3 on October 19, 2026, and calls it “the most capable diffusion-based LLM released to date.” The headline figure is throughput: the company reports 1,068 tokens per second on NVIDIA GPUs and says the model runs 4 to 8x faster (the post does not name the baseline) while scoring on par with or better than GPT-6 Luna and Claude Haiku 5.5 on a range of benchmarks.
The context window rises to 1M tokens, up from 260K for Mercury 2.5. Inception says Mercury 3 is more than 40x larger than Mercury 1 in total parameters and that its standard-benchmark intelligence rose 3x from Mercury 1 (mini), but it does not publish a parameter count. List pricing is $0.20 per million input tokens and $0.70 per million output tokens. At launch the model is 50% off, at $0.10 input and $0.35 output, and the post gives no end date for that discount. Cached input tokens are 90% off the input price. Mercury 3 is available on the Inception API Platform, OpenRouter, Baseten and Vercel, and the API is OpenAI-compatible. The Inception platform offers 100 million free tokens, with no stated duration or terms. Enterprise deployments add dedicated capacity, autoscaling, compliance controls and configurable data retention.
Inception positions Mercury 3 against GPT-6 Luna and Claude Haiku 5.5. The post cites customer results, all vendor-reported. TinyFish AI ran 50 live Mind2Web browser-navigation tasks, and Mercury 3 completed 50% of them, which the company says is more than GPT-6.1-Sol managed. Octen AI reports a median generation time of 0.9 seconds and a median end-to-end response of 2.2 seconds, with its own search step taking about 80 ms. RevisionDojo says hints, marked answers and definitions return in about a second, and Lorikeet reports better latency and eval results in early use. Alongside the flagship, Inception lists Mercury Voice, tuned for time to first answer token, and Mercury Decide, a classification model. The post contains no limitations section.
Several details are missing. The aggregate benchmark scores appear only in a chart image, not in the text. The 1,068 tokens/sec figure comes without a GPU model, batch size or prompt length. The post does not say whether weights are available, does not explain how its diffusion approach generates text, and mentions no function calling, tool use or structured output support beyond OpenAI API compatibility. The post cites no third-party benchmark results.
Source: Inception Labs, October 19, 2026.
Why it matters under load: For agent loops bound by output latency, 1,000+ tokens/sec at $0.35 per million output tokens could cut both wall-clock time and spend, but the discount has no end date, so budgets should be sized at the $0.70 list price.