GPT-6.1 Terra: 10x cheaper, the October 2026 AI recap
October 2026 reminded us of one thing: the price of a frontier model no longer tracks its power. On October 26, OpenAI launched GPT-6.1 Terra at $1 per million input tokens and $5 for output, while GPT-6 Astra asks $10 and $50. Here is the recap: what came out, what the figures say, and what you still need to check yourself.
All prices below are API rates read from the OpenRouter catalogue on November 2, 2026; scores come from Epoch AI’s Epoch Capabilities Index (ECI), the same snapshot as the site’s model pages.
The story of the month: GPT-6.1 Terra, top-tier quality at entry-level prices
OpenRouter describes GPT-6.1 Terra as the mid-tier model of the GPT-6.1 series, positioned below GPT-6.1 Sol. It accepts a context of about 1.05 million tokens.
What stands out is not the price alone but its position on Epoch AI’s index:
| Model | Release date | Price in / out (per million tokens) | ECI (interval) |
|---|---|---|---|
| GPT-6 Astra | September 4 | $10 / $50 | 169.06 (165.88 – 174.29) |
| GPT-6.1 Terra | October 26 | $1 / $5 | 168.11 (164.94 – 172.94) |
| GPT-6.1 Sol | September 29 | $2 / $10 | 167.82 (164.29 – 172.51) |
One point apart on the index, almost fully overlapping intervals, and a price divided by ten. So you cannot claim Astra is better than Terra on this measure. You cannot claim the opposite either: the ECI aggregates tests, and an agent juggling tools over twenty steps does not behave like a maths exam.
The useful reasoning for a small business is simple. If you pay for a top model today to sort emails or write meeting notes, replay ten of your real tasks on Terra. If the result is equivalent, your bill falls tenfold on those tasks. The guide to LLM API pricing explains how to turn a per-million-token price into a monthly budget, and the article on the monthly cost of an AI agent covers the other items.
Small models move up
Terra is not an isolated case. On October 7, Anthropic released Claude Haiku 5.5 at $0.10 input and $0.50 output per million tokens, with one million tokens of context. Epoch AI gives it an ECI of 160.37 (157.33 – 163.84), about 9 points below GPT-6 Astra, for a hundredth of its input price.
Google, for its part, opened access to Gemini 4 Argon on October 13, at $2 / $10, with an ECI of 166.07 (163.14 – 170.27). It lands in the same range as Sol and Terra, at Sol’s price.
What this sequence shows: the quality gap between the top model and the affordable one is narrowing, while the price gap remains very large. To see where that gap actually costs you, The cheapest LLM for each use ranks models by type of task.
Open-weight models: closer and closer to closed ones
October also brought several models whose weights are published on Hugging Face:
- Mistral Large 4, October 6: $0.70 / $2.10, ECI 150.59 (147.40 – 153.49);
- Muse Spark Open, from Meta, October 15: $1.25 / $4.25, ECI 155.05 (153.13 – 157.31), described as optimized for agentic coding;
- DeepSeek V4.1 Flash, October 17: $0.30 / $1.20, ECI 158.83 (156.71 – 162.20);
- MiniMax M3.1 Flash, October 19: $0.15 / $0.60, not yet scored by Epoch AI.
The best of this group, DeepSeek V4.1 Flash, sits about 10 points below GPT-6 Astra on the ECI, but at a thirtieth of its input price, and its file can run on your own hardware if you can afford it. Two caveats. First, “published weights” does not mean “free to use”: the licence is on the model’s page and varies by publisher. Second, running a model with hundreds of billions of parameters takes servers few companies own; most will use it through a host. The choice between the two families is covered in Open-source or closed models for your agents?, and the Mistral AI and DeepSeek articles detail those two publishers.
During the week of October 19–24 we already covered MiniMax and Kimi in the October 24 AI roundup; the table above reuses their up-to-date figures.
A catalogue that grows faster than anyone can read
The November 2 snapshot counts 27 entries created in October on OpenRouter. That figure overstates the number of truly new models: it includes variants, test models and summarization or extraction services. But it gives the pace: more than one release per working day. Nobody can test everything.
The right strategy is not to follow every release but to keep a small set of real tasks (ten to twenty examples from your documents, with the correct answer known) and replay it when an interesting model comes out. Public rankings are for shortlisting; they do not replace that test. To learn to read a score without being misled, see LLM benchmarks: how to read model rankings.
What to take from October
- A pricier model is not always a better fit: try the economical option on your tasks first.
- A one-point ECI gap with overlapping intervals is not a measurable difference.
- Open-weight models are catching up, but read the licence and cost out the hosting before committing.
- Prices quoted are those of November 2, 2026 and may change: the site’s model pages are refreshed weekly, and the This week page lists tracked news.
For the state of agents themselves this month, also read AI agents in October 2026: the state of play.
Frequently asked questions
Does GPT-6.1 Terra replace GPT-6 Astra?
No. They are two offers from the same company at very different prices. Terra is the mid-tier model of the 6.1 series; Astra remains the top model of the previous series. Epoch AI places them very close on its index, but that index summarizes several tests: on your specific task the gap may be larger, or nil. Compare on your own cases.
Is a cheaper model necessarily a worse one?
Not always. October shows models under $1 per million input tokens that come close to top models on published tests. Price also reflects commercial positioning and competition, not only quality. Only a trial on your own documents settles it.
Can I use an open-weight model however I like?
Not automatically. "Open weights" means the model file can be downloaded; the licence says what you may do with it (commercial use, user thresholds, redistribution). Read the licence on the model's page before any business use.
What does Epoch AI's ECI mean?
The Epoch Capabilities Index combines a model's results on several benchmarks into one scale. Epoch publishes a value and an interval: when two models' intervals overlap, you cannot say which is better. The index measures tests, not your use case.
Add Info Agent IA as a preferred source on Google Our articles show up more often in your Google “Top stories”.