Overview
Mercury 2.5 is Inception's diffusion language model, released on 9 September 2026. Where a conventional autoregressive model emits one token at a time, a diffusion LLM generates tokens in parallel — the property Inception builds its whole product line around, and which it credits for sub-300ms time to first token, 5–7× higher throughput and up to 70% lower cost than conventional models.
The launch numbers are about speed at a given quality level. Inception reports 1,107 tokens per second on widely-available NVIDIA GPUs, a 260K-token context window, and a 40% increase in intelligence over Mercury 2 — placing it, in the company's framing, alongside cost-optimized frontier models such as GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite and Claude Haiku 4.5. Inception does not publish a benchmark table for the release, so none is listed here.
The model supports tool calling and structured outputs, and is served through Inception's own API as well as Baseten and OpenRouter. It shipped beside two previews: Mercury Voice, targeting sub-170ms time to first token for voice agents, and Mercury Router, which classifies a request and sends it to the appropriate Mercury model. Weights are not published.
| Released | 2026-09-09 |
|---|---|
| License | Proprietary |
| Weights | API only |
| Context | 260K |
| Architecture | Diffusion language model (dLLM) — tokens are generated in parallel rather than one at a time; Inception calls it the largest diffusion language model ever trained |
| Modalities | Text |
| Status | Generally available — served through the Inception API, Baseten and OpenRouter, with enterprise deployments offering dedicated capacity. |
Pricing
| Input | $0.20 / 1M tokens |
|---|---|
| Output | $0.75 / 1M tokens |
Inception also published a launch promotional rate of $0.04 input and $0.15 output per million tokens, an 80% discount.
Strengths
- 1,107 tokens per second on widely-available NVIDIA GPUs, from parallel diffusion decoding rather than token-by-token generation
- 260K-token context, roughly double the 128K of Mercury 2
- Low published price: $0.20 per million input tokens and $0.75 per million output tokens
- Tool calling and structured outputs are supported, so it drops into agent and extraction pipelines
- Available through three routes — the Inception API, Baseten and OpenRouter — plus enterprise deployments with dedicated capacity
Best for
- Latency-sensitive agent loops where throughput per dollar matters more than topping a leaderboard
- High-volume structured extraction and classification, using the model's structured-output support
- Long-document work that needs the full 260K-token window at a cost-optimized price point
- Voice and real-time products, where Inception pairs Mercury 2.5 with the Mercury Voice preview
How to access
| Provider | Model ID |
|---|---|
| Inception API ↗ | mercury-2.5 |
| OpenRouter ↗ | inception/mercury-2.5 |
FAQ
What is a diffusion LLM?
A language model that generates tokens in parallel through a diffusion process instead of predicting one token at a time. Inception's case for it is hardware efficiency: it reports sub-300ms time to first token, 5–7× higher throughput and up to 70% lower cost than conventional autoregressive models.
How fast is Mercury 2.5?
Inception reports 1,107 tokens per second on widely-available NVIDIA GPUs.
What does Mercury 2.5 cost?
$0.20 per million input tokens and $0.75 per million output tokens. Inception also announced a launch promotional rate of $0.04 input and $0.15 output per million tokens — an 80% discount — and the docs list those promotional figures against mercury-2.5.
Are Mercury 2.5's weights available?
No. It is a proprietary API model, served through the Inception API, Baseten and OpenRouter.
What benchmarks has Inception published for Mercury 2.5?
None as a scored table. The launch post describes a 40% increase in intelligence over Mercury 2 and positions the model alongside GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite and Claude Haiku 4.5, but gives no benchmark figures, so we do not list any.