# Mercury 2.5 upgrades the calls you don't see

> Inception's new model provides a 40% intelligence gain over Mercury 2 while running at 1,107 tokens per second, with launch pricing at just $0.04 per million input tokens. The interesting question is what that does to the steps in your pipeline that run ten thousand times a day.

- Published: 2026-09-11
- Article slug: mercury-2-5-calls-you-dont-see
- Author: [Glevd](https://x.com/glevd)
- Reading time: 5 minutes
- Topics: Mercury 2.5, Inception, diffusion LLM, LLM latency, LLM pricing, sponsored
- Data policy: Dated analysis; static benchmark claims are retained.
- Canonical URL: https://benchlm.ai/blog/posts/mercury-2-5-calls-you-dont-see

*Sponsored by Inception. All figures come from Inception's Mercury 2.5 announcement and its published customer examples. We did not run these benchmarks ourselves.*

## What shipped

Mercury 2.5 is Inception's most capable model to date.

| **Mercury 2.5** | **Announced figure** |
| --- | --- |
| Output speed | 1,107 tokens/s |
| Context window | 260K tokens |
| Standard price, per million tokens | $0.20 in / $0.75 out |
| Launch price 80%, per million tokens | $0.04 in / $0.15 out |

Across a suite of 10 benchmarks, Mercury 2.5 scores on average over 10 points higher than Mercury 2, with the largest gains on agentic or long-context tasks. Inception reports that Mercury 2.5's quality is comparable with GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite, and Claude Haiku 4.5, which is the tier every frontier lab now maintains for high-volume work.

The context window roughly doubles, from 128K to 260K, and [launch rates](https://www.inceptionlabs.ai/models#pricing) are at an 80% discount on standard rates.

A single intelligence figure won't tell you how many of your tasks will pass, so use the chart below as a guide to the kind of applications for which Mercury 2.5 is suited. The overall takeaway is that Mercury 2.5 is far more intelligent than Mercury 2, and that diffusion models can now take on considerably more complex workloads.

![Mercury 2.5 against Mercury 2 on the ten benchmarks Inception reports. Mercury 2.5 averages 51% against 38%, with the largest gains on Tau3Bench Telecom, AA-LCR, DSQA, and Omniscience Non-Hallucination.](/images/blog/mercury-2-5-vs-mercury-2-benchmarks.svg)

*Mercury 2.5 against Mercury 2, ten benchmarks. Source: Inception.*

## The economics of the repeated call

Most AI products have one or two calls the user notices and dozens they don't, including compaction, routing, query rewriting, tool selection, classification, extraction, and guardrail checks. Those invisible calls set the latency floor for everything downstream, and at scale they account for a surprising share of the bill.

They're also the calls most likely to be running on a model chosen two quarters ago because it was the only cheap option that mostly worked. That's the position Mercury 2.5 is aimed at. At 1,107 tokens per second, a 2,000-token output lands in under two seconds. At launch pricing, a million tokens of input costs four cents. The combination means you can put a call in the hot path of a request and stop budgeting around it.

## Where it fits

The workloads that benefit most share a shape: high call volume, latency felt directly by a user or by the agent loop waiting on it, and a job narrow enough that a frontier model is overkill.

**Context compaction.** Long agent sessions and long chats have to be summarized to keep going. The material is already on hand and only needs condensing, so the task rarely calls for frontier-level reasoning. It is also pure overhead from the user's point of view.

**Tool selection and MCP search.** Agents with large tool inventories spend real time deciding what to call. Matching a request against a set of tool descriptions is closer to retrieval than to reasoning, and a small model handles it well enough to keep the loop moving.

**Retrieval preparation.** Query rewriting, expansion, and reranking sit directly in front of the search the user is waiting on. Inception notes that a single [search request can fan out into dozens of model calls](https://www.inceptionlabs.ai/blog/mercury-2-for-search). No one of them needs a large model, but together they determine how long the search takes.

**Bulk extraction and classification.** Document processing, tagging, and structured extraction over large corpora are well-specified jobs with a fixed output shape, which is where a smaller model gives up the least. It is also where the price per million tokens stops being a rounding error.

Inception's published customer examples show these workloads in practice.

Augment Code uses Mercury for context compaction, model routing, and MCP tool search inside a coding agent, with a different model handling the primary task. Inception reports compaction latency falling 82%, from roughly 150 seconds to 27 seconds, alongside a 90% cost reduction with quality maintained. OpenCall, a voice application developer, reports that after switching to Mercury, its [median model response latency halved](https://www.inceptionlabs.ai/blog/mercury-2-the-first-reasoning-model-fast-enough-to-pick-up-the-phone), from 400 milliseconds to under 200.

## How to test it in an afternoon

Mercury is OpenAI API compatible, access is available through the [Inception API](https://platform.inceptionlabs.ai/), Baseten, and [OpenRouter](https://openrouter.ai/inception/mercury-2.5), **with 100 million free API tokens to start**. You can try the model directly in [Mercury Chat](https://chat.inceptionlabs.ai/) or go straight to the [API docs](https://docs.inceptionlabs.ai/get-started/get-started). The 80% launch discount makes now the cheapest time to run the comparison.

## Mercury Voice and Mercury Router

Alongside 2.5, Inception previewed two new products aimed at low-latency, high-volume applications.

**Mercury Voice** is tuned for voice agents on the tightest latency budgets, with reported time-to-first-token under 170 milliseconds. Voice is unforgiving about latency in a way text is not, and time-to-first-token is the figure closest to the pause a caller actually hears.

**Mercury Router** reads an incoming prompt and dispatches it to whichever open or closed model offers the best mix of quality, speed, and cost. Routing only pays off if the decision itself is close to free in both time and money, which makes it a natural job for a fast model.

Both are in preview; you can [contact Inception](mailto:sales@inceptionlabs.ai) for early access. The [Mercury 2.5 model page](/models/mercury-2-5) collects the current listing data.

## Frequently asked questions

### How fast is Mercury 2.5?

Inception reports 1,107 output tokens per second for Mercury 2.5. At that rate a 2,000-token output lands in under two seconds. The figure comes from Inception's announcement, not from a BenchLM run, and end-to-end latency on your route also depends on prompt length, region, and time to first token.

### How much does Mercury 2.5 cost?

Inception lists Mercury 2.5 at $0.20 per million input tokens and $0.75 per million output tokens. Launch pricing takes 80% off both rates, to $0.04 in and $0.15 out per million tokens. Inception has not published an end date for the launch discount, so check the pricing page before budgeting.

### How does Mercury 2.5 compare with Mercury 2?

Across the ten benchmarks Inception reports, Mercury 2.5 averages 51% against 38% for Mercury 2, with the largest gains on agentic and long-context tasks such as Tau3Bench Telecom, AA-LCR, and DSQA. The context window roughly doubles, from 128K to 260K tokens. All figures are Inception's own.

### Which workloads suit Mercury 2.5?

The best fits share a shape: high call volume, latency that a user or an agent loop feels directly, and a job narrow enough that a frontier model is overkill. Context compaction, tool selection and MCP search, query rewriting and reranking, and bulk extraction or classification are the examples Inception's customers describe.