Mercury 3 Pricing: Diffusion LLM Tested on a Chatbot

Mercury 3 pricing is $0.10 per million input tokens and $0.35 per million output tokens, with cached input at $0.01. That is a 50% launch discount: Inception lists the regular rate as $0.20 / $0.70, and it has not said when the discount ends. Mercury 3 is a diffusion language model. It drafts and refines many tokens in parallel instead of writing one token at a time, and Inception sells it on speed.

We wanted two numbers a WordPress site owner can act on: what Mercury 3 costs per chatbot answer, and whether the speed shows up in a real support-bot request. So on 28 October 2026 we sent the same retrieval-augmented request from our own WordPress server to Mercury 3 at four reasoning settings, to Mercury 2.5 and Mercury 2, and to five rival models. That was 96 calls, all of which returned HTTP 200.

The short answer: at the launch price, 1,000 answers cost $0.47 on Mercury 3, the same as GPT-6 Luna and about half of Claude Haiku 5.5 ($0.89). Mercury 3 was also the fastest model we tested, at a 1.13-second median answer by default and 0.74 seconds with reasoning off, against 2.34 seconds on Haiku 5.5 and 3.02 seconds on Luna. All 32 Mercury 3 answers were correct. When the discount ends, the same 1,000 answers cost $0.93, slightly more than Haiku 5.5.

Mercury 3 pricing at a glance

These are the rates OpenRouter and Inception published on 28 October 2026. OpenRouter passes Inception’s price through with no markup on the model. Its endpoint record for Mercury 3 carries a discount: 0.5 flag, which matches the “50% off at launch” wording on Inception’s model page.

Mercury 3 can return up to 131,072 output tokens, and it supports tools, structured outputs and a reasoning_effort setting. Inception reports 1,068 output tokens per second on NVIDIA GPUs. That is a vendor figure, and we come back to it below.

How we tested it

Every request was the one a MxChat support bot actually sends. It contained our site’s system prompt, four passages pulled from our own documentation and blog (about 4,000 input tokens in total), and one visitor question. We asked four questions that real visitors ask:

1. Does MxChat work with WooCommerce, and can the bot look up an order status for a logged-in customer?
2. Do I need Pinecone to use the knowledge base, or can I store embeddings in WordPress?
3. Can a visitor be handed over to a human on Slack or Telegram when the bot cannot answer?
4. Can I use my own OpenRouter or OpenAI API key, and which models can I choose from?

Each question went to each setup twice, so every setup had 8 calls. All calls went through OpenRouter from our WordPress host in the US, without streaming, so the time we report is the time to a complete answer. Mercury routes were pinned to Inception as the provider, with fallbacks off. We priced every call from its own token counts at the model’s list price with no cache. That gives the cost of a first-time question, because cache hits depend on your traffic. Caching is covered in its own section below.

We graded each answer against the facts in the passages. For example, WooCommerce order lookup needs the Pro add-on and a logged-in customer, and Pinecone is optional.

What 1,000 chatbot answers cost

Three things stand out.

Input dominates the bill, so effort barely moves the price. A retrieval chatbot sends about 4,000 tokens of context to get a 40-word answer. On Mercury 3, output was only 14% of the cost by default. Going from reasoning off to high effort changed the price per 1,000 answers by 7 cents. If you are choosing an effort level for a knowledge-base bot, choose it for speed, not price.

Mercury 3 matches GPT-6 Luna on price and beats it on time. Both models charge $0.10 per million input tokens and used almost the same number of input tokens (4,023 against 4,067). Mercury’s output rate is lower ($0.35 against $0.50), but it wrote more reasoning tokens, so the two ended up at the same $0.47. The difference is time: 1.13 seconds against 3.02.

Claude Haiku 5.5 costs more than its rate card suggests. It has the same $0.10 input rate, but Anthropic’s tokenizer counted the same request as 6,838 input tokens rather than about 4,050, and Haiku wrote 403 output tokens per answer. We covered that tokenizer gap in our Claude tokenizer test and in our Claude Haiku 5.5 pricing test.

Speed: is a diffusion model faster in a real chatbot?

Yes, and by a wide enough margin that visitors would notice. Every Mercury 3 setting had a faster median than every non-Mercury model in the test. With reasoning off or effort low, the median was 0.74 to 0.75 seconds, and the slowest of those 16 calls took 2.15 seconds. Claude Haiku 5.5 took 2.34 seconds at the median, GPT-6 Luna 3.02 and MiniMax M3.1 Flash 3.74. The slowest MiniMax and DeepSeek calls took over 10 seconds.

Do not expect to see Inception’s 1,068 tokens per second in a chatbot, though. Dividing output tokens by total time, Mercury 3 delivered about 170 to 210 tokens per second end to end. That is still several times the 27 to 80 we measured on GPT-6 Luna, GPT-6.1 Terra, MiniMax and DeepSeek. The gap from the headline figure is not a broken promise. A chatbot answer is short, so most of the wall time goes on network round trips and on reading 4,000 tokens of context before the first word, and generation speed cannot shrink either of those. The longer the answer, the closer you get to the headline. Mercury 2.5, which wrote about 1,000 tokens per answer, reached 407 tokens per second.

Reasoning effort: low is the setting to use

Mercury 3’s default is medium effort. Sending reasoning: {effort: "medium"} produced the same 132 reasoning tokens per answer as sending nothing. Here is what the other settings did on our four questions:

For questions that the knowledge base answers directly, reasoning added time without improving the answers. Low effort was as fast as reasoning off and still gave the model a short think. That is the setting we would use for a support bot that also takes product questions. Note that Mercury 3 accepted every setting we sent, including turning reasoning off completely. That is not true of every model: in yesterday’s test, OpenAI’s GPT-6.1 Terra rejected minimal with an HTTP 400, and MiniMax M3.1 Flash rejects reasoning off.

There is one oddity. High effort finished slightly faster at the median than medium effort while writing more tokens. With 8 calls per setting, that is within the noise, so read it as “about the same”.

Mercury 2.5 and Mercury 2: should you use the older models?

Mercury 2.5 is the cheapest model in the test at $0.31 per 1,000 answers, and it got all 8 answers right. But it is a heavy reasoner. It spent 958 reasoning tokens on each 29-word answer, its median time was 2.65 seconds, and every answer began with two blank lines. Those lines are harmless in most chat widgets, but they are worth trimming if yours renders them. Choose it if cost is the only thing that matters. Mercury 3 at its launch price costs 15 cents more per 1,000 answers and is more than twice as fast.

Mercury 2 costs $1.19 per 1,000 answers, more than either newer model, and it was the only setup in the test that got an answer wrong. Asked about WooCommerce order lookups, it left out that the feature needs the Pro add-on. The same answer also garbled a phrase (“real statusorder status”) and added a citation link the passages did not contain. If you are on Mercury 2 today, Mercury 3 is cheaper and better.

Prompt caching on Mercury 3

Cached input on Mercury 3 costs $0.01 per million tokens, 10% of the input rate, and caching is automatic. You do not mark anything in the request. But it caught less of the prompt than OpenAI’s caching did.

Mercury 3’s cache matched about half the prompt on a repeat question and none of it on the high-effort repeats. Caching like this is best-effort, so plan on the no-cache column. On a busy bot where the same system prompt and passages repeat all day, GPT-6 Luna’s full cache hits pull it well below Mercury 3 on price. Our OpenAI prompt caching test covers how those hits are billed.

What it costs per month

Here is what the measured costs come to at three traffic levels, before any cache savings:

For most WordPress sites, the model bill is not the deciding factor at these prices. A store answering 10,000 questions a month pays less than $10 on any of the first four columns. The deciding factors are answer quality, which was tied on our questions, and how long a visitor waits, where Mercury 3 won clearly.

Where Mercury 3 falls short

• The price is temporary. At list price, Mercury 3 costs about the same as Claude Haiku 5.5 and twice GPT-6 Luna. Inception has not published an end date or a notice period for the discount. If you switch for the price, recheck it monthly.
• One provider. On OpenRouter, Inception is the only host for Mercury 3, and Inception says the model is API-only with no open weights. GPT-6 Luna and Claude Haiku 5.5 each have more than one route. If Inception has an outage, you have no second host for the same model.
• Weaker caching. Half the prompt cached at best on our repeats, against all of it on OpenAI.
• A small test. Four questions and 8 calls per setup show cost, speed and basic accuracy on a retrieval bot. They do not show how Mercury 3 handles long multi-turn chats, tool calls or hard reasoning. Inception’s benchmark claims against GPT-6 Luna and Haiku 5.5 were run by Inception itself.

How to use Mercury 3 in a WordPress chatbot

MxChat reaches Mercury 3 through OpenRouter:

1. Create an OpenRouter API key and add it under MxChat → Settings → API Keys.
2. In the chatbot’s model setting, choose inception/mercury-3 from the OpenRouter model list. That list is read live from OpenRouter, so new models appear without a plugin update.
3. Ask your bot the five questions your visitors ask most, and check the answers against your knowledge base before you switch it on for everyone.

MxChat does not send a reasoning setting on the OpenRouter route today, so Mercury 3 runs at its default medium effort. In our test, that was still the fastest setup apart from the other Mercury 3 settings, at 1.13 seconds. For a comparison of OpenRouter’s fees with going direct to each vendor, see our OpenRouter pricing test. The MxChat documentation covers knowledge-base setup, and MxChat Pro adds the WooCommerce and live-agent features our test questions asked about.

Mercury 3 against the other cheap models we have tested

We have run the same request against most of this year’s low-cost models. If you are choosing between them:

• GPT-6 Luna costs the same as Mercury 3 at launch prices and becomes far cheaper with a warm cache, but it is about three times slower.
• MiniMax M3.1 Flash cost $0.86 per 1,000 answers in its own test, with prices varying by host. It was the slowest model here.
• GPT-6.1 Terra is the mid-tier option at about ten times the price, for bots that need stronger reasoning than a knowledge-base lookup.

FAQ

How much does Mercury 3 cost?

$0.10 per million input tokens, $0.01 per million cached input tokens and $0.35 per million output tokens at Inception’s launch price, as of 28 October 2026. The list price is $0.20 / $0.02 / $0.70.

When does the Mercury 3 launch discount end?

Inception has not said. Its model page describes the price as 50% off at launch with no end date. We will add a dated note to this post when the price changes.

What is a diffusion LLM?

A language model that writes by drafting a block of tokens and refining them in parallel over several passes, instead of predicting one token after another. The advantage is generation speed, which is why Mercury 3 finished every answer in our test faster than the conventional models we tested against.

Is Mercury 3 cheaper than GPT-6 Luna?

Not in practice. Both cost $0.47 per 1,000 of our chatbot answers without caching. With a warm cache, GPT-6 Luna was cheaper ($0.10 against $0.31) because OpenAI cached the whole prompt. Mercury 3 was 2.7 times faster.

Which Mercury 3 reasoning setting should a chatbot use?

Low. In our test it was as fast as reasoning off (0.75 against 0.74 seconds), cost a cent less per 1,000 answers than the default, and answered all 8 questions correctly.

Does Mercury 3 work with MxChat?

Yes, through OpenRouter. Add an OpenRouter key and pick inception/mercury-3 from the model list.

Method note: 96 chat-completion calls through OpenRouter from mxchat.ai’s WordPress server on 28 October 2026, 12 setups × 4 questions × 2 runs, all HTTP 200. Costs are measured token counts at each model’s list price (Mercury 3 at its launch price unless marked) with no cache. Times are the median wall-clock time to a complete, non-streamed answer. Prices read from OpenRouter’s model API and Inception’s model page on 28 October 2026.