Pareto AI Pricing: 26.10 vs 26.9 on a Chatbot

Pareto is the model family sold by Unbiased, and its October release made it a lot cheaper. Pareto 26.10 lists on OpenRouter at $0.80 per million input tokens and $3.20 per million output tokens. The September release, Pareto 26.9, is still listed at $2.50 and $7.50. Input is 68% cheaper, output 57% cheaper, and the context window grew from 262,144 tokens to just over a million. Unbiased describes Pareto as a multimodal “composite” model for research, coding and agent work. It is the only host for both versions.

Rate cards don’t tell you what a support chatbot will actually cost, so on 3 November 2026 we sent the same support-bot request from our own WordPress server to Pareto 26.10, Pareto 26.9 and the pareto-latest alias, and to five rival models. We timed every answer, recorded the tokens each one billed, and checked each answer against our documentation.

The short answer: on the same request, Pareto 26.10 cost $3.71 per 1,000 answers at list price with no cache, against $11.78 on Pareto 26.9, so the new version is about 68% cheaper per answer. It still costs about four times as much as Claude Haiku 5.5, which came in at $0.87 and answered faster (2.03 s median against 3.08 s). Pareto’s automatic prompt cache narrows the gap a lot: once the cache was warm, 26.10 billed $0.99 per 1,000 answers. All 24 Pareto answers were correct.

Pareto pricing at a glance

These are the rates on OpenRouter’s model API on 3 November 2026. Unbiased serves both models; OpenRouter shows no other host.

The cached-input line is the one to notice. On 26.10, a cached token costs $0.03 per million, under 4% of the normal input price. On 26.9 a cached token costs a tenth of the normal price. For a chatbot that sends the same system prompt with every question, that difference decides most of the bill.

Neither Pareto model lists a reasoning setting on OpenRouter, so there is no effort level to tune and no hidden thinking tokens to pay for. In our test both models reported zero reasoning tokens on every answer.

How we tested it

Every request was the one a MxChat support bot actually sends: the site’s real system prompt plus four passages pulled from our own documentation, about 4,150 tokens on Pareto’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 eight setups: Pareto 26.10 and Pareto 26.9 pinned to Unbiased with fallbacks off; the pareto-latest alias; and five rivals, all through OpenRouter with the same prompt: Claude Haiku 5.5, DeepSeek V4.1 Flash, GLM-5.3 FlashX, GPT-5.6 Luna and GPT-6.1 Terra. That is 64 requests, and all 64 came back with HTTP 200 on the first try.

We report cost two ways. No cache is the tokens each answer used multiplied by the list price, which is what a low-traffic site pays when every question arrives cold. Warm is what OpenRouter actually billed on the second pass, after the first pass had filled the provider’s prompt cache. A busy site sits somewhere close to the warm figure. For DeepSeek V4.1 Flash, which OpenRouter routes across several hosts, we use the billed first-pass cost as the no-cache figure.

What 1,000 chatbot answers cost

Three things stand out.

The version change is real money. Both Pareto versions used exactly the same number of input tokens (4,149), so the gap is the rate card. At list price, 26.9 spends $10.37 of its $11.78 on input; 26.10 spends $3.32 of its $3.71. A site handling 10,000 questions a month with no cache would pay about $118 on 26.9 and about $37 on 26.10. If you set up a Pareto bot in September and pinned the version, you are still paying the old rates.

Input is most of the bill. Pareto’s answers were short, 37 to 42 words on average, and output was only about a tenth of the cost. That is why Pareto loses to Haiku 5.5 even though Haiku used 65% more input tokens on the same text: Haiku’s input rate is $0.10, an eighth of Pareto’s.

The cache changes the ranking. Warm, Pareto 26.10 billed $0.99 per 1,000 answers, close to Haiku 5.5’s $0.87. Haiku reported no cached tokens on either pass, so its warm price was the same as its cold one. GPT-5.6 Luna and DeepSeek V4.1 Flash cached automatically too and dropped to about $0.17.

Pareto’s cache kicks in sooner than most

Pareto 26.10 started reading from cache during the first pass. Its first question went in cold, but by the second question 3,111 of 4,122 input tokens were billed as cached, because every request starts with the same system prompt and the same documentation passage. Across the first pass, an average of 2,597 tokens per request came from cache, and the first-pass bill was $1.84 per 1,000 answers, half the no-cache figure.

On the second pass, every Pareto request had 3,375 cached input tokens. Here is how that played out on each version:

Two caveats. First, we ran setups side by side, so the 26.10 first pass may have been helped by cache entries from the pareto-latest alias, which ran at the same moment and resolved to the same model. The second-pass figures don’t have that problem. Second, a cache only lasts a few minutes without traffic. A site that gets one question an hour will pay something close to the no-cache figure, whatever model it uses.

On both OpenAI models, the first pass billed more than their reported tokens at list price would suggest: $1.10 against $0.90 for Luna and $5.58 against $4.67 for Terra. We report the reported-token figure as the no-cache cost so all models are compared the same way, but your OpenAI bill for cold traffic may come in higher.

Is pareto-latest the same as 26.10?

On billing, yes: the alias was charged at the 26.10 rates on every request. Its answers were longer, though, 229 output tokens on average against 123 when we called 26.10 directly, and its median answer time was a second slower (4.08 s against 3.08 s). With eight answers each, that may just be ordinary variation. Since the cost difference is small, the real question is whether you want your bot to switch to a new model when Unbiased ships one.

For a support chatbot, we’d pin the version. Unbiased has released two Pareto versions a month apart, and the rates moved a lot between them. An alias means your bot’s costs, tone and answer length can change on a day you didn’t choose. Pinning unbiased/pareto-26.10 keeps the bot predictable, and you can move it deliberately when the next release appears.

Speed

Pareto 26.10 had a 3.08-second median to a complete, non-streamed answer, with a range of 2.02 to 5.92 seconds. Pareto 26.9 was about the same at 2.99 seconds. Both were slower than Claude Haiku 5.5 (2.03 s), GPT-5.6 Luna (1.71 s) and DeepSeek V4.1 Flash (1.00 s), and faster than GLM-5.3 FlashX (4.98 s).

Pareto is not fast at writing. 26.10 produced about 32 output tokens per second, measured over the whole request, against 170 for Haiku 5.5. Its answers are short enough that this barely matters for a support bot, but if you ask it for long replies, expect them to take a while.

Answer quality

All 24 Pareto answers were correct against our documentation. Every one said order lookups need the Pro add-on and a logged-in customer, that Pinecone is optional and embeddings can stay in WordPress, that Slack and Telegram handoff both exist, and that you can bring your own API key. Here is 26.10’s first answer to the WooCommerce question, word for word:

Yes, MxChat integrates with WooCommerce via the Pro add-on, and the bot can look up order status for logged-in customers. The Order History action retrieves order details and status, secured for logged-in users only. You’d need an active MxChat Pro license for this.

That is a good support answer: direct, accurate, and it names the condition a site owner needs to know. The answers were also short. 26.10 averaged 42 words and 26.9 averaged 37, so neither pads its replies.

Among the rivals, Claude Haiku 5.5, DeepSeek V4.1 Flash, GLM-5.3 FlashX and GPT-5.6 Luna all got 8 of 8. GPT-6.1 Terra got 6 of 8: on one WooCommerce question and one Pinecone question it replied that it did not have enough information in its knowledge base, even though the passages it was given contained the answer. That is the safe kind of miss, but a visitor still leaves without an answer.

Who should use Pareto for a WordPress chatbot

The honest summary: Pareto 26.10 gives good, short, correct answers, and its cached-input discount is unusually deep. For a typical support bot, though, it costs more than the cheap tier of 2026 models and isn’t faster, so it’s hard to recommend over Haiku 5.5, Luna or DeepSeek’s Flash unless you have a reason to want Pareto. If you’re on 26.9, switching to 26.10 is a straightforward saving.

How to use Pareto in MxChat

1. Create an API key on OpenRouter and add credit to the account.
2. In WordPress, open MxChat’s settings and paste the key into the OpenRouter field. Our documentation covers where each key goes.
3. Pick the model. Type or select unbiased/pareto-26.10 rather than the alias, so the bot stays on the version you tested.
4. Send your bot a few real visitor questions and check the answers against your own pages before you switch it on for everyone.
5. Watch the OpenRouter activity page for the first few days. It shows each request’s token usage and cost, which tells you whether you are paying closer to the cold or the warm price.

If you are deciding between OpenRouter and going direct, our OpenRouter vs direct API pricing guide covers what the router adds. Order lookups for logged-in WooCommerce customers need MxChat Pro.

How Pareto compares with our other tests

We run the same chatbot request against new models as they arrive. For the cheap end of the market, see our Claude Haiku 5.5 pricing test and DeepSeek V4.1 Flash pricing. For OpenAI’s mid tier, our GPT-6.1 Terra pricing test goes into Terra in more depth. If you are weighing Z.ai’s models, our GLM-5.3 pricing guide compares its hosts. And yesterday’s Liquid AI LFM 2.5 test shows the opposite problem to Pareto’s: very cheap input, but so much hidden reasoning that it still cost more than Haiku.

Frequently asked questions

How much does Pareto cost?

On OpenRouter on 3 November 2026, Pareto 26.10 costs $0.80 per million input tokens, $0.03 per million cached input tokens and $3.20 per million output tokens. Pareto 26.9 costs $2.50, $0.25 and $7.50. On our support-bot request, that came to $3.71 per 1,000 answers on 26.10 with no cache and $0.99 with a warm cache.

Is Pareto 26.10 cheaper than Pareto 26.9?

Yes. On the same request, with the same token counts, 26.10 cost 68% less per answer with no cache ($3.71 against $11.78) and 78% less warm ($0.99 against $4.43). Both versions got all eight answers right.

Who makes Pareto?

Pareto is sold by Unbiased, which describes it as a multimodal composite model for research, coding and agent work. Unbiased is the only host listed for it on OpenRouter.

Does Pareto support prompt caching?

Yes, automatically. You don’t need to mark anything in the request. In our test Pareto 26.10 began reading from cache on the second question and cached 3,375 of about 4,150 input tokens per request once warm.

Should I use pareto-latest or a pinned version?

For a live chatbot, pin a version such as unbiased/pareto-26.10. The alias follows each new release, and Pareto’s rates and behaviour have changed between versions a month apart.

Is Pareto good for a WordPress chatbot?

It answered all of our support questions correctly and briefly. It costs more than Claude Haiku 5.5, GPT-5.6 Luna and DeepSeek V4.1 Flash on the same request and was slower than all three, so it suits sites that want Pareto specifically or need its million-token context more than it suits a typical support bot.

Method note: tests run 3 November 2026 from mxchat.ai’s server in the US through OpenRouter, 4 questions × 2 passes per setup, temperature 1. Prices are OpenRouter list prices read the same day. Costs per 1,000 answers are means over the eight answers in each setup. Your costs depend on your prompt size, answer length and traffic pattern.