Liquid AI LFM 2.5 Pricing: Tested on a Chatbot

Liquid AI’s LFM 2.5 24B has the lowest input price of any model we have put through our chatbot test this autumn: $0.05 per million input tokens and $0.70 per million output tokens on OpenRouter, served by Liquid itself. It is a mixture-of-experts reasoning model with 24 billion parameters in total and about 2.45 billion active per token, built, Liquid says, for agents that call tools. OpenRouter added it on 13 October 2026, next to a smaller sibling, LFM 2.5 2.6B.

A low input price matters a lot for a support chatbot, because every visitor question carries the whole system prompt and a stack of knowledge-base passages with it. So on 2 November 2026 we sent the same support-bot request from our own WordPress server to LFM 2.5 24B at three reasoning-effort levels and with reasoning switched off, tried the 2.6B model, and ran the same request against four rival models. Every answer that came back was checked against our documentation.

The short answer: LFM 2.5 24B cost $0.96 per 1,000 answers with no cache and answered in a 2.96-second median. That is more than Claude Haiku 5.5, which cost $0.85 and took 2.15 seconds, even though Haiku’s input rate is twice as high. The reason is output: LFM 2.5 wrote about 1,070 output tokens per answer, almost all of it reasoning, against 328 for Haiku. It got 22 of 24 answers fully right; the two misses left out a detail rather than getting anything wrong. Reasoning cannot be switched off, and the small 2.6B model was not usable in a live bot on the day we tested it.

LFM 2.5 pricing at a glance

These are the rates on OpenRouter’s model API on 2 November 2026. Liquid is the only host for both models.

Two things in that table are unusual. The smaller model charges seven times more per input token than the larger one, so for a chatbot, where input dominates, the 2.6B is the expensive choice. And the 24B’s output rate is fourteen times its input rate, a much steeper ratio than most models. For comparison, Claude Haiku 5.5 lists at $0.10 input and $0.50 output, Nemotron 3.5 Super at $0.15 and $0.50, and Reflection’s Beam 1.0 at $0.25 and $1.00 on its cheapest host.

Price trackers disagree about what Liquid charges on its own platform; one lists $0.10 and $1.00. We could not confirm a direct price, so this post covers OpenRouter, which is how a WordPress plugin like MxChat reaches the model.

How we tested it

Every request was the one a MxChat support bot actually sends. It contained the site’s real system prompt plus four passages pulled from our own documentation, about 4,300 tokens on LFM’s count, followed by one visitor question. We asked four questions, each twice:

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?

We ran these setups:

• LFM 2.5 24B on Liquid at the default setting, with reasoning effort set to low, set to high, and with reasoning switched off, host pinned with fallbacks disabled.
• LFM 2.5 2.6B, paid and free.
• Claude Haiku 5.5, the cheap default we recommend most often.
• DeepSeek V4.1 Flash and Nemotron 3.5 Super, two cheap open models, with the host left to OpenRouter as MxChat does, and Reflection Beam 1.0 on Nebius, the open model we tested yesterday.
• We also queued Mercury 3, Inception’s diffusion model, but it returned HTTP 429 (“temporarily rate-limited upstream”) on every one of 64 attempts in two separate runs, so it is not in the results.

All calls went through OpenRouter from our WordPress server in the US. They were non-streamed, with usage accounting switched on, so OpenRouter reported the exact tokens and billed cost of every call. The first pass of each setup gives the “cold” cost; the second, identical pass shows what caching does. We checked every answer against the facts in the passages. For example, the WooCommerce answer had to say the feature needs MxChat Pro or its add-on and a logged-in customer, and the Pinecone answer had to say WordPress storage works.

What 1,000 chatbot answers cost

Three things stand out.

The input price is low enough to stop mattering. At $0.05 per million, our 4,300-token prompt costs about $0.21 per 1,000 answers on LFM 2.5. On Claude Haiku 5.5 the same request counts as 6,838 tokens and costs about $0.68. If you only looked at input, LFM would be a third of Haiku’s price.

Output decides the bill, and LFM writes a lot of it. LFM 2.5 produced about 1,070 output tokens per answer at the default setting. The answers the visitor saw averaged 46 words, about 60 tokens, so roughly nine tenths of what we paid for was reasoning. At $0.70 per million, that output cost about $0.75 per 1,000 answers, close to four fifths of the $0.96 total. Haiku 5.5 wrote 328 output tokens per answer at $0.50 per million, about $0.16. That is the whole difference.

The answers were short and mostly right. The visible replies were 43 to 46 words, about the same length as Haiku’s. Every LFM answer said WordPress can store embeddings without Pinecone, named both Slack and Telegram for live handoff, and said you can bring your own API key. The two misses were both on the WooCommerce question: each mentioned the add-on and order tracking but did not say the lookup works only for a logged-in customer. Incomplete, not wrong. At effort high, all eight answers were complete.

Where the tokens go: reasoning you cannot switch off

OpenRouter’s model page says LFM 2.5 24B “always reasons before answering”, and the API holds to that. Sending reasoning: {enabled: false} returned HTTP 400 on all eight calls with the message “Reasoning is mandatory for this endpoint and cannot be disabled.” That is the third model in three days of our tests that will not let you turn reasoning off, after StepFun’s Step 5 Preview and Reflection’s Beam 1.0.

The effort setting does not help much either:

Setting effort to low changed almost nothing: 2% fewer tokens, one cent less per 1,000 answers, and a slightly slower median in our sample. Setting it to high added 43% more output tokens and 34% more cost, and it did fix the two incomplete WooCommerce answers. If you use LFM 2.5 for a support bot, leave effort at the default; there is no cheaper mode to switch to.

The thinking also varies a great deal from one question to the next. At the default setting the shortest call used 475 output tokens and the longest 2,044, for answers of similar length. That spread shows up in the timings below and in the bill: you pay for the long ones too.

Speed: fast generation, slow answers

Liquid’s host is quick at producing tokens. At the default setting the median call generated about 365 output tokens per second, against about 165 for Claude Haiku 5.5 and 148 for DeepSeek V4.1 Flash. But LFM has to write three to five times as many tokens before the answer appears, so the time a visitor waits is longer: a 2.96-second median against 2.15 for Haiku and 0.75 for DeepSeek V4.1 Flash.

The range matters more than the median for a chat window. LFM’s fastest answer took 1.12 seconds and its slowest 5.50 seconds at the default setting, and 8.15 seconds at effort high. If your chatbot does not stream its replies, a visitor will sometimes stare at a typing indicator for five seconds. Nemotron 3.5 Super, on hosts OpenRouter chose, was slower still at a 5.25-second median with one 53.8-second outlier.

Caching: listed, but it did not show up

OpenRouter lists a cached-input rate of $0.025 per million for LFM 2.5 24B, half the normal input price. On our second, identical pass the API reported zero cached tokens on every call, and the repeat bill was $0.91 against $1.02 on the first pass; the difference came from shorter reasoning, not from the cache.

Even a working cache would not change much here. Input is only about $0.21 of the $0.96, so halving it would save about 11 cents per 1,000 answers. Compare DeepSeek V4.1 Flash, which cached most of the repeated prompt and fell from $0.49 to $0.28, or Beam 1.0 on Nebius, which fell from $1.28 to $0.93. Claude Haiku 5.5 reported no cached tokens on our request either.

The smaller LFM 2.5 2.6B

We also tried to test the small model, and it is worth knowing what happened.

• The paid 2.6B endpoint would not run on our account. OpenRouter returned HTTP 404 with “0 endpoints out of 1 requested are available”, giving the reason as “Paid model training violation (account settings)”. Our OpenRouter account is set to avoid providers that may use prompts for training, and OpenRouter excluded Liquid’s paid 2.6B endpoint under that setting. The 24B model ran normally under the same setting. If you keep the same privacy setting, you will hit the same wall.
• The free 2.6B answered 9 of 16 requests. The other 7 returned HTTP 429, rate-limited upstream, across two runs. That is fine for a quick experiment and not acceptable for a live chatbot.
• When it answered, it was quick but wordy. The median answer took 1.21 seconds and ran to about 90 words, roughly twice as long as the 24B’s. Eight of nine answers were complete; one WooCommerce answer left out the Pro or add-on requirement.

And because the paid 2.6B charges $0.35 per million input tokens, seven times the 24B, it would cost more per answer than the larger model on an input-heavy chatbot even if it ran. For a WordPress support bot, the 24B is the LFM 2.5 to consider.

Should you run your WordPress chatbot on LFM 2.5?

Probably not as your default, but it depends on what the bot does.

• For a plain knowledge-base support bot, LFM 2.5 24B gives correct answers at a fair price, but it is beaten on every measure we took. DeepSeek V4.1 Flash cost half as much, answered four times faster and got all eight answers right. Claude Haiku 5.5 cost slightly less, was faster, and was fully correct.
• If your prompts are very long, the maths shifts in LFM’s favour. A bot that sends 20,000 tokens of product catalogue with every question (on LFM’s count) would pay about $1.00 per 1,000 answers for input on LFM. Claude’s tokenizer counts the same text as about 60% more tokens, so Haiku 5.5 would pay about $3.20 before output. As long as LFM’s reasoning does not grow with the prompt, the longer the prompt, the better its $0.05 rate looks.
• If your bot calls tools, such as order lookups or bookings, LFM 2.5 is built for that work and is worth a test of its own. Our test covered answers from a knowledge base, not tool calls.
• If you need fast replies, avoid it. A 3-second median with a 5.5-second tail is slow next to the 0.75-second median DeepSeek V4.1 Flash managed on the same request.

How to use LFM 2.5 with MxChat

1. Create an OpenRouter account and an API key, and add credit.
2. In MxChat, paste the key into the OpenRouter field of the API keys settings.
3. In the chatbot’s model setting, choose liquid/lfm-2.5-24b-a2b from the OpenRouter model list. The list is read live from OpenRouter, so new models appear without a plugin update.
4. Ask your bot the five questions your visitors ask most, check the answers against your knowledge base, and time them before you switch it on for everyone.

If your OpenRouter account blocks providers that may train on prompts, the 24B model still works; the paid 2.6B does not. Our OpenRouter pricing test covers its fees against going direct. The MxChat documentation covers knowledge-base setup, and MxChat Pro adds the WooCommerce and live-agent features our test questions asked about.

LFM 2.5 against the other models we have tested

We have run the same chatbot request against most of this year’s models:

• Claude Haiku 5.5: $0.85 per 1,000 answers and 2.15 seconds today. Still the safest cheap default.
• DeepSeek V4.1 Flash: $0.49 and 0.75 seconds today, the cheapest and fastest model in this test.
• Nemotron 3.5 Super: $0.75 today but 5.25 seconds on hosts OpenRouter chose, another reasoning-heavy open model.
• Reflection Beam 1.0: $1.28 and 1.49 seconds on Nebius today; it also refuses to switch reasoning off, but thinks a quarter as much as LFM.
• MiniMax M3.1 Flash: another cheap reasoning model where turning reasoning off returns an error.
• Mercury 3: $0.47 per 1,000 answers and about a second in our 28 October test, but rate-limited all day today.

LFM 2.5 24B is a good illustration of a rule that applies to every reasoning model: the price per token tells you little until you know how many tokens it writes. Its $0.05 input rate is real, and for very long prompts it pays off. For a typical support bot, the reasoning it insists on doing costs more than the input saves.

FAQ

How much does Liquid AI’s LFM 2.5 cost?

On OpenRouter on 2 November 2026, LFM 2.5 24B cost $0.05 per million input tokens and $0.70 per million output tokens, with cached input listed at $0.025. LFM 2.5 2.6B cost $0.35 input and $0.50 output, and a rate-limited free version was also listed. On our chatbot request the 24B came to $0.96 per 1,000 answers with no cache.

Is LFM 2.5 cheaper than Claude Haiku 5.5?

Per input token, yes: half the price. Per chatbot answer, no. LFM 2.5 24B wrote about 1,070 output tokens per answer, mostly reasoning, against 328 for Haiku 5.5, so it cost $0.96 per 1,000 answers against $0.85.

Can you turn off reasoning on LFM 2.5?

No. Sending reasoning disabled returned HTTP 400 with “Reasoning is mandatory for this endpoint and cannot be disabled.” Effort low barely changed the bill in our test; effort high cost 34% more.

How fast is LFM 2.5 for a chatbot?

The median complete answer took 2.96 seconds at the default setting, with a range of 1.12 to 5.50 seconds. Liquid generated about 365 tokens per second, but the model writes a lot of reasoning before the answer.

Why can’t I use LFM 2.5 2.6B on OpenRouter?

If your OpenRouter account is set to avoid providers that may train on prompts, OpenRouter excludes the paid 2.6B endpoint and returns HTTP 404 with “Paid model training violation (account settings)”. The free version ran, but 7 of our 16 requests were rate-limited.

Does LFM 2.5 work with MxChat?

Yes, through OpenRouter. Add an OpenRouter key and pick liquid/lfm-2.5-24b-a2b from the model list.

Method note: chat-completion calls through OpenRouter from mxchat.ai’s WordPress server on 2 November 2026, 4 questions × 2 runs per setup. LFM 2.5 24B pinned to Liquid with fallbacks disabled; 24 answers at three effort levels, 8 HTTP 400 with reasoning off. LFM 2.5 2.6B paid: 8 HTTP 404 (account data policy). LFM 2.5 2.6B free: 16 requests over two runs, 9 answers, 7 HTTP 429. Mercury 3: HTTP 429 on all 64 attempts, excluded. Other models on OpenRouter’s default routing except Beam 1.0, pinned to Nebius. Costs are measured token counts at list price with no cache, except DeepSeek V4.1 Flash and Nemotron 3.5 Super, which are first-pass billed costs because their hosts varied. Repeat-pass costs are what OpenRouter billed. Times are the median wall-clock time to a complete, non-streamed answer. Prices, context windows and output limits read from OpenRouter’s model API on 2 November 2026. Parameter counts are Liquid’s, as listed on OpenRouter.