Mercury 2.5 Preview

Mercury 2.5 Preview is a diffusion LLM that Inception says can hit 1,107 tokens per second. That sounds wild, but for now it is still a vendor claim with no independent validation.
Advertisement 728 × 90
CompanyInception
Context260K
Released2026-08
Updated2026-09-03

Mercury 2.5 Preview Overview

Mercury 2.5 Preview launched on August 31, 2026. It is a diffusion large language model, or dLLM, available through APIs including OpenRouter.

GPT, Claude, and most other LLMs generate tokens in order. One comes out, then the next, then the next. Mercury works differently: it creates text in parallel and repeatedly refines that draft through a denoising process.

That is where the speed claim comes from.

Mercury 2.5 supports roughly 256K to 260K tokens of context, with up to about 65K tokens of output. It also has adjustable reasoning levels, parallel tool calling, and schema-based JSON output. It is strictly text-only, so there is no image, audio, or video support.

Inception is not pitching this as a rival to the biggest flagship models. The comparison set is closer to GPT-5.6 Luna at lower reasoning levels, Gemini 3.5 Flash-Lite, and Claude Haiku 4.5.

In other words, the bet here is not “smartest model wins.” It is fast enough, capable enough, and cheap enough.

There is a catch, and it is a fairly big one. The weights are closed, independent benchmark results are still missing, and platforms such as BenchLM do not yet have public scores for the model.

So for now, we are working with vendor numbers, not proof.

Mercury 2.5 Preview Pricing

PlanPriceDescription
Input List price $0.20 / 1M tokens; launch price $0.04 80% off through September 8
Output List price $0.75 / 1M tokens; launch price $0.15 80% off through September 8
Cache reads $0.02 / 1M tokens Standard price

At the discounted rate, it is genuinely cheap. $0.04 for input and $0.15 for output puts it near the floor for this class of API model.

That makes the preview easy to experiment with. It does not make the long-term economics clear.

The discount disappears after September 8, bringing pricing back to $0.20 for input and $0.75 for output. And because Mercury is closed and API-only, there is no option to download the weights and take control of inference costs yourself.

Mercury 2.5 Preview Key Features

1. Diffusion Is the Whole Point

Mercury does not generate text the way most LLMs do.

Think of a normal LLM as an extremely fast typist: one token comes out, then another, and the answer grows from left to right.

Mercury is closer to laying down a rough page and cleaning the whole thing up over several passes.

Inception says that gets Mercury 2.5 to 1,107 tokens per second on standard GPUs.

If independent testers can reproduce that number under realistic conditions, it is a big deal. Long answers could feel nearly instant, and agent systems that spend their day generating and regenerating text could shave meaningful time off each run.

Voice is another obvious use case. Nobody enjoys waiting two seconds for an AI to start talking back.

Still, 1,107 tokens per second is not a fact we would put in ink just yet. It is Inception’s benchmark. We need someone else to run the stopwatch.

2. You Can Trade Reasoning for Speed

Mercury 2.5 lets developers adjust reasoning intensity.

For easy requests, turn it down and get the answer faster with fewer tokens. For harder coding or logic work, raise it and give the model more room to reason.

Nothing revolutionary there. Plenty of newer models expose some version of this control.

It matters more on Mercury because speed is the product. If a basic lookup spends ages thinking before those 1,107 tokens per second kick in, the headline number stops being quite so exciting.

3. Parallel Tools Make More Sense Here

Mercury supports parallel tool calls and schema-aligned JSON output.

The JSON part is useful but fairly standard developer plumbing. Parallel tool use is more interesting.

Imagine an agent that needs prices from one API, account data from another, and a database lookup before it can answer. Running those requests one after another wastes time. Running them together does not.

Pair that with fast generation and Mercury starts to make more sense as an agent model than as a chatbot you use to ask random questions.

That is probably where we would test it first.

4. Plenty of Context, Plenty of Output

Depending on the listing, Mercury 2.5 is shown with a 256K or 260K context window. Maximum output is around 65K tokens.

That is enough for large codebases, long documents, or agents carrying a substantial amount of history.

There are models with bigger context windows, but that feels beside the point here. Mercury is selling speed and price. A quarter-million-plus tokens is already enough for most of the workloads it is targeting.

Summary

Pros

  • Diffusion-based generation rather than standard autoregressive decoding
  • Inception claims up to 1,107 tokens per second
  • Very cheap launch pricing: $0.04 input and $0.15 output per million tokens
  • Parallel tool calling
  • Structured JSON output
  • Roughly 256K–260K context
  • About 65K maximum output
  • Potentially very attractive for low-latency agents

Limitations

  • Speed and performance claims still lack independent verification
  • Closed-source and API-only
  • No self-hosting
  • Text only
  • The 80% discount ends September 8, 2026
  • It is aimed at the fast-and-affordable tier, not the absolute top of the intelligence charts

Recommended For

  • Voice, search, and coding-agent developers who care heavily about latency
  • Researchers curious about diffusion-based language models
  • Teams that value generation speed more than maximum reasoning quality
  • Anyone who wants to run a cheap proof of concept before the launch discount disappears

Not Recommended For

  • Teams that need the model weights
  • Companies that require fully local deployment
  • Apps that need image, audio, or video understanding
  • Critical production workloads that cannot afford to discover later that the vendor benchmarks were optimistic

Comments (0)

Leave a comment

Advertisement 728 × 90

Similar Models