Mercury 2.5 Says It Is 40 Percent Smarter and Publishes No Benchmark to Prove It

Inception shipped Mercury 2.5 on 8 September 2026, and calls it the largest diffusion language model ever trained. The headline number is 1,107 tokens per second, which works out to 66,420 tokens a minute, on hardware the company describes only as "widely-available NVIDIA GPUs."

Diffusion models are worth paying attention to because they break the one assumption every other LLM is built on. Instead of emitting tokens one after another, they produce and refine many at once, which is why the speed numbers look implausible. We do not have an Inception or OpenRouter key in our tooling, so we could not run the model. What we could do is pull OpenRouter's public model catalog and check the announcement against the live listing, line by line.

What the catalog actually says

A single unauthenticated request to https://openrouter.ai/api/v1/models returns every model OpenRouter serves, including its context length and per-token price. Both Mercury generations are in there.

$ curl -s https://openrouter.ai/api/v1/models | jq '.data[]
    | select(.id | test("mercury"))
    | {id, context_length, pricing}'

The two listings are 188 days apart, Mercury 2 on 4 March 2026 and Mercury 2.5 on 8 September. Here is what changed between them, straight from the catalog rather than the blog post.

Field	Mercury 2	Mercury 2.5
Context length	128,000	260,000
Input, standard	$0.25 / M	$0.20 / M
Output, standard	$0.75 / M	$0.75 / M
Input, live launch price	$0.25 / M	$0.04 / M
Output, live launch price	$0.75 / M	$0.15 / M
Published benchmark scores	None found	None found

The pricing claim holds up exactly. OpenRouter is currently serving Mercury 2.5 at $0.04 per million input tokens and $0.15 per million output tokens, which is precisely the 80 percent launch discount Inception advertises. The 260,000 token context is real. The context window has grown 103 percent in six months.

The quieter finding is the standard rate. At list price the input cost fell 20 percent between generations and the output cost did not move at all. When the promotion ends, Mercury 2.5 costs the same per output token as the model it replaces, with more context and, allegedly, more capability. That is a good deal. It is also not the deal the announcement is selling, which is the discounted one.

The two numbers we could not check

Now the part the catalog cannot settle.

1,107 tokens per second. The announcement gives the figure and the phrase "widely-available NVIDIA GPUs." It does not name the GPU, the batch size, or whether the number is output-only throughput or total. Those three variables can move a throughput figure by an order of magnitude in either direction. This is not a made-up number, and diffusion architectures genuinely do produce results in this range, but as published it is not reproducible by anyone outside Inception.

A 40 percent increase in intelligence. The announcement states "Quality: 40% increase in intelligence from Mercury 2. Comparable to cost-optimized frontier models like GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite, and Claude Haiku 4.5." We went looking for the table behind that sentence and there is not one. No named benchmark, no score, no evaluation harness, no methodology. "Intelligence" is not a quantity with a unit, so a percentage attached to it is a marketing figure wearing a lab coat.

The announcement also carries a customer anecdote: P99 latency dropping from several minutes to one second on a production workload. The customer is not named and the workload is not described. We are reporting it because it is in the source, not because we can stand behind it.

What else is in the box

Three capabilities are new in 2.5: tunable reasoning, parallel tool calls and schema-aligned JSON. The third is the one that matters for anyone building agents. Diffusion models refine a whole span of output at once, which is structurally well suited to filling a schema, and it removes the most common failure mode of cheap fast models, which is valid-looking JSON that does not parse.

Availability is the Inception API, Baseten and OpenRouter, plus enterprise deployments with dedicated capacity. That is a genuinely broad launch footprint for a company this size.

Almost nobody noticed

We looked for the argument and could not find one. The Hacker News submission for the launch sat at 1 point with 1 comment. There is no thread to quote, no dissent to weigh, no maintainer of an adjacent tool poking holes in the throughput figure. On the same day, a media server release took 601 points and 343 comments.

That silence is itself the story. Diffusion LLMs have been the most interesting architectural bet in inference for over a year, and the model that claims to be the largest one ever trained landed with no discussion at all. Part of that is launch fatigue. Part of it, we suspect, is the missing benchmark: there is nothing in the announcement for a skeptic to argue with, so nobody argued.

The pattern worth watching

Sequential generation is the reason your agent sits there for eleven seconds before the first useful line appears. Every technique the industry has thrown at that problem, speculative decoding, smaller draft models, aggressive caching, works around the constraint rather than removing it. Diffusion removes it.

Latency is the last thing an adjective can fix, which is why speed is the one claim Inception put a real number on.

If you are running something latency-bound, an autocomplete, an inline edit, a tool-calling loop where the round trip is the cost, Mercury 2.5 is worth an afternoon and roughly nothing in tokens while the discount holds. If you are choosing a model on reasoning quality, there is currently no evidence to choose it on, and asking Inception to publish one table is a reasonable thing to want before you migrate. We would revisit this the day they do.