Overview
Mercury 3 is Inception's diffusion language model, released on 19 October 2026. Inception calls it the most capable diffusion-based LLM released to date and says it outperforms GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks while running 4–8x faster, at 1,068 tokens per second on NVIDIA GPUs. The launch post gives no scored benchmark table, so none is listed here.
The release is Inception's argument that text diffusion scales. Mercury 3 is more than 40x larger than Mercury 1 in total parameters yet, per Inception, just as fast, and the company says autoregressive scaling laws hold for diffusion too. The context window grows to 1M tokens, up from 260K on Mercury 2.5, with up to 131,072 output tokens.
The model supports tool calling and structured outputs, and its OpenAI-compatible API takes a reasoning effort of instant, low, medium or high. On 50 live Mind2Web browser-agent tasks Inception reports a 50% completion rate. Mercury 3 is served through the Inception API, OpenRouter, Baseten and Vercel and in the Mercury Chat app. Weights are not published.
| Released | 2026-10-19 |
|---|---|
| License | Proprietary |
| Weights | API only |
| Context | 1M |
| Architecture | Diffusion language model (dLLM) — tokens are generated in parallel rather than one at a time; Inception says it is more than 40x larger than Mercury 1 in total parameters |
| Modalities | Text |
| Status | Generally available — served through the Inception API, OpenRouter, Baseten and Vercel and in Mercury Chat, with enterprise deployments offering dedicated capacity. |
Pricing
| Input | $0.20 / 1M tokens |
|---|---|
| Output | $0.70 / 1M tokens |
Inception's docs list a 50% launch discount: $0.10 input, $0.01 cached input and $0.35 output per million tokens.
Strengths
- 1,068 tokens per second on NVIDIA GPUs, from parallel diffusion decoding
- 1M-token context window, up from 260K on Mercury 2.5, and up to 131,072 output tokens
- Low published price: $0.20 per million input tokens and $0.70 per million output tokens, with cached input 90% off
- Tool calling, structured outputs and four reasoning-effort levels on an OpenAI-compatible API
- Available on the Inception API, OpenRouter, Baseten and Vercel
Best for
- Browser and computer-use agents where each step's latency adds up
- Real-time search answers, such as Inception's Octen AI example with a 0.9-second median generation time
- Long-document work that needs the full 1M-token window at a low price
- Interactive education and chat products that need fast responses
How to access
| Provider | Model ID |
|---|---|
| Inception API ↗ | mercury-3 |
| OpenRouter ↗ | inception/mercury-3 |
Mercury — every version
The full lineage of the Mercury line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.
| Version | Released | Context | License |
|---|---|---|---|
| Mercury 3current | 2026-10-19 | 1M | Proprietary |
| Mercury 2.5 | 2026-09-09 | 260K | Proprietary |
FAQ
How fast is Mercury 3?
Inception reports 1,068 tokens per second on NVIDIA GPUs and says it runs 4–8x faster than the models it compares against.
What does Mercury 3 cost?
$0.20 per million input tokens and $0.70 per million output tokens, with cached input 90% off. At launch Inception's docs list a 50% discount: $0.10 input, $0.01 cached input and $0.35 output per million tokens. New API accounts start with 100M free tokens.
Are Mercury 3's weights available?
No. Mercury 3 is a proprietary API model, served through the Inception API, OpenRouter, Baseten and Vercel and in Mercury Chat.
What benchmarks has Inception published for Mercury 3?
No scored table. The launch post says Mercury 3 outperforms GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks without listing numbers; its one task figure is a 50% completion rate on 50 live Mind2Web browser tasks.