GPT-6.1 Terra Pricing: 10 Chatbot Routes Tested

GPT-6.1 Terra pricing is $1 per million input tokens and $5 per million output tokens, with cached input at $0.05. That is half the input price and 58% less on output than GPT-5.6 Terra, the model it replaces in OpenAI’s middle tier, which still lists at $2 and $12. GPT-6.1 Terra appeared on OpenRouter on 26 October 2026, and we ran it through a WordPress support chatbot the next morning to see what the new rate card means per answer.

Short version: 1,000 chatbot answers cost $4.49 on GPT-6.1 Terra at list price with no cache, against $8.65 on GPT-5.6 Terra and $8.93 on GPT-6.1 Sol. All 8 answers were correct at the default, low and minimal effort settings. At high effort the model refused 5 of 8 questions that its knowledge base did answer, and cost 22% more for the privilege. GPT-6.1 Terra Pro, which uses the same per-token price, cost $14.82 per 1,000 answers because it billed 3.4 times as many input tokens.

Below: the full price list, how we tested, cost and speed for every route, what each reasoning-effort setting actually did, how caching changes the bill, and two API restrictions that will break a chatbot that uses function tools if you move to Terra without checking.

GPT-6.1 Terra pricing in October 2026

These are the per-token rates OpenRouter listed on 27 October 2026. OpenAI is the only provider for every model in the table. Prices above 272,000 input tokens per request are higher; a chatbot request is nowhere near that, so the first column of rates is the one that applies to you.

Three things stand out. First, the new Terra is priced at exactly half of GPT-6.1 Sol on both input and output, so the tiers are now evenly spaced: Luna, then Terra at ten times Luna, then Sol at twice Terra. Second, the cached-input rate fell from 10% of the input price on GPT-5.6 Terra to 5% on GPT-6.1 Terra, so a cache hit is now a quarter of what it was in dollars. Third, Terra Pro has the same per-token rates as Terra. OpenRouter describes it as the same model with a “pro” reasoning mode and warns that it spends far more tokens. Our test found where those tokens go.

OpenRouter also lists a cache write rate of $1.25 per million for GPT-6.1 Terra, 1.25 times the input price. We saw it on the bill: the first call with a new prompt cost more than the list price, and later identical calls cost far less. We covered that pricing change in detail in our piece on OpenAI’s cache write charge.

How we tested

We built each request the way MxChat builds a chatbot reply. The system message holds our site’s own system prompt plus four knowledge-base passages from mxchat.ai pages, and the user message is the visitor’s question. That came to about 4,070 input tokens on OpenAI’s tokenizer. The four questions are ones our own chatbot gets:

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

Each question went twice to each of 10 routes through OpenRouter on 27 October 2026: GPT-6.1 Terra at its default effort and at minimal, low and high effort; GPT-6.1 Terra Pro; GPT-5.6 Terra; GPT-6.1 Sol; GPT-6 Luna; Claude Haiku 5.5; and Claude Sonnet 5.5. We also sent the four questions straight to OpenAI’s API with our own key, to check that OpenRouter was not changing the token counts. It was not: the direct calls counted the same input tokens to the token. That is 84 priced chatbot calls, plus 19 short calls to check which API settings OpenAI accepts.

We read every answer. The knowledge base answers all four questions: the WooCommerce order lookup needs MxChat Pro and a logged-in customer, Pinecone is optional, Slack and Telegram handoff are both available, and you can bring your own key. An answer counted as correct if it said those things.

Costs below are measured tokens multiplied by each model’s list price, with no cache discount, so the routes compare on equal terms. Cached figures are in their own section.

What 1,000 chatbot answers cost on GPT-6.1 Terra

Input is 90% or more of the bill on every OpenAI route at default settings. A chatbot request is a long prompt and a short answer: about 4,070 tokens in and 85 out on GPT-6.1 Terra. That is why halving the input price nearly halves the cost per answer, from $8.65 to $4.49 (−48%), even though the new Terra thinks a little more by default than the old one did. GPT-5.6 Terra spent no reasoning tokens on these questions at its default setting; GPT-6.1 Terra spent 37 on average.

Against the rest of OpenAI’s line, Terra now sits where its name suggests. It costs about half of GPT-6.1 Sol for answers that were just as correct on this task, and about ten times GPT-6 Luna, which also got all 8 right. For a support bot that answers from a knowledge base, Luna remains the cheaper choice by a wide margin. Terra makes sense if your bot has to reason across several passages, read screenshots, or handle conversations where Luna’s answers fall short. Test it on your own questions before you pay ten times more.

The Claude rows count more input tokens for the same text (6,838 against 4,067) because Anthropic’s tokenizer splits text differently. So Claude Sonnet 5.5, which has the same $2 / $10 rate card as GPT-6.1 Sol, cost almost twice as much per answer. Claude Haiku 5.5 is the closest rival to Luna on price.

Reasoning effort: high effort made the answers worse

GPT-6.1 Terra takes a reasoning effort setting. We tried four settings through OpenRouter. Minimal and low behaved the same: no reasoning tokens, the fastest answers, the lowest cost, and 8 correct answers out of 8. The default spent 37 reasoning tokens per answer, took about 0.7 seconds longer, and was also 8 for 8.

High effort was different. It spent 251 reasoning tokens per answer, took more than twice as long (4.63 seconds median), and in 5 of 8 calls it replied with our system prompt’s fallback line, “I don’t have enough information in my knowledge base to answer that question accurately.” The information was there. Every other route found it, and high effort itself answered the same Pinecone and WooCommerce questions correctly on its first run and refused them on its second. Our reading: with more room to think, the model became stricter about whether the passages answered the exact question, and our system prompt tells it to use the fallback line when unsure. That is a reasonable instruction, and a reasonable model can take it too literally.

The lesson for a site owner is simple. Do not raise reasoning effort on a support chatbot on the assumption that it makes answers better. On this task it cost 22% more per answer, doubled the wait, and turned correct answers into refusals. Low effort was the best setting we tested: cheapest, fastest, and fully correct.

Minimal works on OpenRouter, not on OpenAI’s API

One detail matters if you call OpenAI directly. OpenRouter accepted “minimal” and returned answers with no reasoning tokens. OpenAI’s own Chat Completions API rejected it with HTTP 400: “‘reasoning_effort’ does not support ‘minimal’ with this model.” The values OpenAI accepts for GPT-6.1 Terra are none, low, medium, high and xhigh. Use “none” or “low” if you call OpenAI directly. OpenAI also rejects any temperature other than the default of 1 on this model, which matters if your plugin sends a lower temperature for more consistent answers.

GPT-6.1 Terra Pro: same price per token, 3.3 times the bill

Terra Pro has the same rate card as Terra, and its answers were as good: 8 for 8, about 30 words each. But every request billed between 13,450 and 14,373 input tokens for a prompt that Terra counted as about 4,070. The extra input tokens are not in the request we sent; they come from the extra work pro mode does on OpenAI’s side, and you pay for them at the input rate. Output was also higher, 210 tokens against 85.

The result is $14.82 per 1,000 answers with no cache, 3.3 times plain Terra, and more than GPT-6.1 Sol. Most of those extra input tokens were cached on repeat calls, which brings the actual bill down (see the next section), but a cache does not help with a visitor’s first question. For a chatbot, Terra Pro buys nothing we could measure. Keep it for hard one-off tasks where a better answer is worth several times the cost.

How caching changes the GPT-6.1 Terra bill

OpenAI caches long prompt prefixes automatically. When the same system prompt and passages come back within a few minutes, the cached part is billed at $0.05 per million on GPT-6.1 Terra instead of $1. On our second run of each question, 4,064 of 4,067 input tokens were cached on every OpenAI route.

The fully cached figure is the best case: the same prompt, repeated within the cache window. A real chatbot gets less. The system prompt is the same on every request, so that part caches well. The knowledge-base passages change with each question, so they only cache when visitors ask about the same topic close together. And the first request with a new prefix pays the cache write rate of $1.25 per million, 25% more than the normal input price. We saw that on OpenRouter’s bill: the first call of a new question cost about $5.23 per 1,000 answers before the repeats came in at a fraction of that.

Anthropic’s models did not cache at all in our test, because Anthropic only caches what the request marks for caching and our test request did not mark anything. That is a setting your plugin controls, not a property of the model.

On speed, GPT-6.1 Terra at low effort answered in a median 1.43 seconds, close to GPT-5.6 Terra at 1.21 seconds. The default effort setting added about 0.7 seconds. Luna was slower than either Terra in this run (3.78 seconds), because it spent 78 reasoning tokens per answer at its default setting; our Luna test shows how to bring that down.

Function tools: the 400 error to expect

If your chatbot uses function calling, for example to look up a WooCommerce order, check this before you switch. We sent a request with one tool, a lookup_order function, to OpenAI’s Chat Completions endpoint with a reasoning effort set. GPT-6.1 Terra returned HTTP 400: “Function tools with reasoning_effort are not supported for gpt-6.1-terra in /v1/chat/completions. To use function tools, use /v1/responses or set reasoning_effort to ‘none’.” GPT-5.6 Terra and GPT-6 Luna give the same error.

Every call that went through returned the right tool call, lookup_order with order 4512. So the model handles tools fine; the restriction is in the Chat Completions endpoint. For Terra and Luna, sending effort “none” with tools works. For Sol, the only route for tools with these models is the Responses API, because Sol does not accept “none” at all. If your plugin or integration still uses Chat Completions, this is the error that will show up the day you move a tool-using bot to a newer OpenAI model.

Should you switch your WordPress chatbot to GPT-6.1 Terra?

If you are on GPT-5.6 Terra today, yes, once you have checked two settings. The new model cut our cost per answer by 48% with the same accuracy and nearly the same speed at low effort. Before switching, make sure your integration sends a reasoning effort OpenAI accepts for this model (none or low, not minimal) and a temperature of 1, and check the tools restriction above if your bot calls functions.

If you are on GPT-6.1 Sol for a support bot, Terra did the same job here for half the price. If you are on Luna or Claude Haiku 5.5, Terra is an upgrade you pay for at roughly ten and five times the cost respectively, so only move if your own questions show Luna or Haiku getting things wrong.

MxChat lets you use your own OpenAI or OpenRouter API key, so the provider bills you at these rates with no markup from us; see our comparison of OpenRouter and direct API pricing for when the router is worth it. MxChat’s model list includes GPT-5.6 Terra, GPT-6.1 Sol and GPT-6 Luna today. GPT-6.1 Terra is newer than the current release. Setup guides are in the MxChat documentation, and MxChat Pro adds the WooCommerce order lookup our first test question asked about.

Frequently asked questions

How much does GPT-6.1 Terra cost?

$1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens, as listed on OpenRouter on 27 October 2026. Requests over 272,000 input tokens cost $2 and $7.50. Batch requests cost half.

Is GPT-6.1 Terra cheaper than GPT-5.6 Terra?

Yes. Input is half the price ($1 against $2) and output is 58% cheaper ($5 against $12). On our chatbot request that worked out to $4.49 per 1,000 answers against $8.65, a 48% saving.

How much does GPT-6.1 Terra cost per chatbot answer?

About $0.0045 per answer with a 4,000-token prompt and no cache, or $4.49 per 1,000 answers. With the prompt fully cached it fell to $0.63 per 1,000 answers.

What is the difference between GPT-6.1 Terra and Terra Pro?

They are the same model at the same per-token price. Pro runs a heavier reasoning mode that billed about 3.4 times as many input tokens in our test, so each answer cost 3.3 times as much. The answers were not measurably better for a support chatbot.

Why does GPT-6.1 Terra return a 400 error with reasoning_effort?

Two common causes. OpenAI’s API does not accept “minimal” for this model (use none, low, medium, high or xhigh), and Chat Completions does not accept function tools together with a reasoning effort other than “none”. Use the Responses API if you need both.

Which reasoning effort should a chatbot use on GPT-6.1 Terra?

Low. In our test it was the fastest (1.43 seconds median), among the cheapest, and answered all 8 questions correctly. High effort refused 5 of 8 questions it should have answered.

Prices read from OpenRouter on 27 October 2026; all tests run the same day from our WordPress server. Model prices change often, so check the provider’s pricing page before you commit.