Skip to main content

Mercury 3 Review: A Diffusion Language Model With a 1M-Token Window

Pick Mercury 3 for latency-sensitive jobs like support replies and classification; check speed yourself, as the 1,000+ tokens per second is Inception's.

Open the Playground

Spicy Index

71.8/ 100

Preliminary

Freedom Score

Not yet tested

Price
$0.325 per 1M mixed tokens (3 input : 1 output)
Output speed
238 tok/s
Success rate
Collecting data

Mercury 3 is Inception's diffusion-based language model, released on 19 October 2026. Instead of writing one token after another, it refines whole passages in parallel, which Inception credits for its speed. On our API it is one OpenAI-compatible chat endpoint: up to 1,000,000 tokens of text in, up to 131,072 tokens out, tools, JSON output and streaming, at $0.20 per million input tokens and $0.70 output.

The verdict

This is a catalogue review. We checked the model against Inception's launch post and our endpoint limits and have not run our own speed, quality or content tests on it. Inception reports 1,068 tokens per second on NVIDIA GPUs and says the model outperforms Claude Haiku 5.5 and GPT-6 Luna on its benchmarks; those are the publisher's figures, and one independent write-up notes that Haiku 5.5 wins two of the benchmarks in Inception's own table.

Pick it where waiting hurts: voice, search, classification, support. Choose something else for deep reasoning or for images, which this endpoint does not take.

Try Mercury 3 →

Key facts

  • Tested by us: catalogue check and launch smoke run only, no quality or content tests; speed and benchmark numbers are Inception's
  • Price: $0.20 per million input tokens, $0.02 cached and $0.70 output
  • Limits: 1,000,000-token context, replies up to 131,072 tokens (default 4,096), text-only input, up to 128 tools
  • Reasoning: none, low, medium or high; low is the default; temperature 0.5 to 1
  • Sources: the scoring rules are on our review methodology page, and the full parameter list is on the Mercury 3 model page.

What we liked and what held it back

What we liked

  • Built for speed. Inception reports 1,068 tokens per second on NVIDIA GPUs; measure it on your own workload, since speed here depends on load, prompt length and reasoning effort.
  • A 1,000,000-token window. Up from about 260K in Mercury 2.5 according to Inception.
  • A low price. $0.20 in and $0.70 out per million tokens, with cached input at $0.02.
  • Reasoning defaults to low. reasoning_effort takes none, low, medium and high, and the default favours speed.
  • Tools and JSON. Up to 128 tool definitions per request and response_format for a JSON object.

What held it back

  • Text only. No image, audio or file parts.
  • A narrow temperature range. The endpoint accepts 0.5 to 1.
  • Speed is not guaranteed. The headline figure was measured by Inception, not on this endpoint.
  • Benchmarks are the publisher's. Third-party composite scores favour Claude Haiku 5.5, so run your own prompts before switching.
  • Max output and knowledge cutoff are not stated by Inception.

What we tested

We have not run Mercury 3 through our capability sets or content levels. This page uses Inception's launch post of 19 October 2026, independent coverage of it, the limits and parameters of our endpoint and the live price on the model page. We quote no benchmark scores because the launch post publishes none.

Prompts that worked

Starting points based on the endpoint's shape and Inception's stated use cases, not prompts we benchmarked.

1. A short support reply. The kind of short, latency-sensitive job the model is aimed at.

text
Draft a concise support reply that answers the question and states the next action.

Run this prompt on Mercury 3 →

2. A tool step. One function, tool_choice set to auto.

text
Call report_status with status ready.

Run this prompt on Mercury 3 →

Settings that matter: keep reasoning_effort at low or none for classification and replies, move to medium or high only for multi-step problems; use stream for anything long; set temperature between 0.5 and 1.

Roleplay and how far it writes

We have not tested Mercury 3 on our content levels yet, so this review makes no claim about how far it writes or how often it softens or refuses. The model page carries no content-test results for it either. On our API we add no content filter; whatever comes back is the model's own behaviour.

Not tested yet: all five content levels, roleplay quality over many turns, and languages other than English. We will add results here once the runs are done.

Every character in a request should be an adult, keep real, identifiable people out of it, and you're responsible for the law where you publish.

Alternatives

  • A measured alternative at a similar tier? Claude Haiku 5.5 has a full review with capability scores and reads text and images.
  • Even lower price, same shape? Ling 3.1 Flash is text-only with a 262,144-token window.
  • Images and a 1M window? Muse Wonder 1.0 accepts images, video and audio.

How Mercury 3 earns its Spicy Index

Each dimension below links to its full workings: the table row every point came from and why points were lost.

Public leaderboards 40% of quality
Not yet tested

No exact match on a tracked public leaderboard yet. How it works

Role-play 15% of quality
Not yet tested

Our role-play suite hasn’t run on this model yet. How it works

Creative writing 15% of quality
53.0
Instruction following 15% of quality
100.0
Long-context memory 15% of quality
Not yet tested

Our long-context suite hasn’t run on this model yet. How it works

Capability 10.5% weight
45.0
Freedom Score 30% weight
Not yet tested

Freedom evidence is untested or insufficient. The Index retains the original quality and capability score, is marked preliminary and still receives a rank. Missing evidence is not zero. How it works

Freedom evidence is untested or insufficient. The Index retains the original quality and capability score, is marked preliminary and still receives a rank. Missing evidence is not zero.

Open the full test report →

Capability, item by item

Points from the public scoring table. Upscaled tiers and limits we set ourselves don’t count.

Context window (creator-published)Not declared
0.0 / 30
Input modalities on SpicyAPIText only
0.0 / 25
Tool useTools and tool_choice
15.0 / 15
Structured outputAccepted, schema adherence verified
15.0 / 15
Reasoning control4 levels: none · low · medium · high
15.0 / 15

What Mercury 3 can do on SpicyAPI

Straight from the live model specification on SpicyAPI. Some creators offer more in their own apps; this is what you can call through our API today.

Capabilities by endpoint
FeatureChat
Context window1,000,000 tokens
Max output per request131,072 tokens
Reasoning levelsnone · low · medium · high
Tool callingYes
Structured outputYes
See every parameter and code sample →

How long it really takes

Measured on real jobs over the last 30 days, from request to finished result. Times include our own test runs.

Output speed
238 tok/sIncludes time to first token
Success rate
Collecting data

A time appears once at least 10 completed jobs are in. Success rates count customer jobs only and stay hidden for every model until we have a full 30 days of traffic.

Models to consider instead of Mercury 3

What do you think?

Was this review useful?

Log in to vote.

Comments

Comments are checked automatically and by our team. We remove spam, offensive content, personal attacks and anything that names a real person without consent. Report a comment and we review it within one business day.

Tried this model? Share what worked. You need a free account to comment, so we can keep spam out.

Mercury 3 FAQ

Try Mercury 3 yourself

Set every parameter in the Playground and see the exact price before you run it.

$0.325 per 1M mixed tokens (3 input : 1 output)

More reviews

See the leaderboards →
  • Small blank marble tiles race along a short brass rail on a basalt plinth into a three-slot marble gate, where a gold filament splits into three channels ending in three neat stacks

    GPT 6.1 Terra

    Pick GPT 6.1 Terra for high-volume screens, documents and writing where reasoning_effort none helps; so far we have run only launch checks, no scored tests.

    Spicy Index
    79.9
    Freedom Score
    —

    Tested Oct 28, 2026

    Read the review
  • Small blank marble tiles race along a short brass rail on a basalt plinth into a three-slot marble gate, where a gold filament splits into three channels ending in three neat stacks

    Kimi K3.1 Pro

    Pick Kimi K3.1 Pro for long-context work with screenshots and clips when you accept always-on thinking; so far we have run only launch checks, no scored tests.

    Spicy Index
    80.4
    Freedom Score
    —

    Tested Oct 28, 2026

    Read the review
  • A burst of golden light filaments fanning low across a shallow glass dish against pure black

    Grok 4.3

    Pick Grok 4.3 for a huge context window, a genuine reasoning off switch and uncensored writing at a low price; skip it if you need file uploads.

    Spicy Index
    77.0
    Freedom Score
    98.9

    Tested Sep 24, 2026

    Read the review