📘 Corporate AI training: our free roadmap for a successful rollout

★★★★★

4,9/5 sur Google

🔌 Connect Claude to Pennylane: our step-by-step setup guide

Download the Claude x Pennylane guide

★★★★★

4,9/5 sur Google

2026-10-28

OpenAI strikes hard: GPT-6.1 Terra, close to Claude Fable 5.1 and half the price of GPT-6.1 Sol

Emma Pariente

On a pink background, a woman with curly hair and mauve sunglasses smiles while holding a smartphone, in front of a GPT-6.1 Terra chat window asking “How can I help you?”.

On 26 October 2026, OpenAI launched GPT-6.1 Terra, the mid-tier model in its GPT-6 range. Its API price is half that of GPT-6.1 Sol. On the Artificial Analysis index, it sits level with Claude Fable 5.1 at medium effort, for a cost per task around seven times lower.

At AI Sisters, an AI consultancy, we help teams choose and use these models. Here is what GPT-6.1 Terra is really worth: what OpenAI announces, what independent evaluators measure, what it costs, and which tasks to choose it for.

In brief

GPT-6.1 Terra was released on 26 October 2026. It replaces GPT-5.6 Terra, between Luna, the lightest, and Sol, the top of the range.
Its API price is $1 per million input tokens and $5 per million output tokens: half the price of GPT-6.1 Sol and of the previous Terra.
On the Artificial Analysis index, it scores 48 points at maximum effort, against 49 for Claude Fable 5.1 at medium effort and 52 for GPT-6.1 Sol.
OpenAI presents it as a model for office work: reading a screen, clicking, filling in forms, entering invoices.
It is available in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu subscribers, and in the API under the name gpt-6.1-terra.
OpenAI rates it at the “critical” level for cybersecurity and applies the same safeguards as GPT-6 Astra.

Also worth reading: Claude, ChatGPT or Gemini: which AI to choose for your company, OpenAI Dots, the arrival of AI colleagues and Claude Duet, Anthropic’s AI agent.

What is GPT-6.1 Terra?

GPT-6.1 Terra is OpenAI’s mid-range model. OpenAI describes it as “a fast and efficient model”, designed for vision, precise computer use and more natural writing.

Since September, OpenAI’s range has had four tiers, from the cheapest to the most powerful: Luna, Terra, Sol and Astra. Sol and Luna moved to generation 6 in September; Terra stayed on version 5.6. GPT-6.1 Terra fills that gap.

OpenAI’s GPT-6 range, from the cheapest to the most powerful, with API input and output prices per million tokens. GPT-6 Luna, the fastest and cheapest: 0.10 dollars and 0.50 dollars. GPT-6.1 Terra, new, for office work, vision and computer use: 1 dollar and 5 dollars. GPT-6.1 Sol, top of the range for complex code: 2 dollars and 10 dollars. GPT-6 Astra, OpenAI’s most powerful model: 10 dollars and 50 dollars. — Terra sits between Luna and Sol, at a tenth of the price of Astra.

Its knowledge cutoff is 1 June 2026, more recent than that of Sol and Astra (30 April 2026), according to The Register. Ted Sanders, a researcher at OpenAI, summed up its strengths on Hacker News: Terra is “probably better” than Astra and Sol for technical communication, comparable in vision and computer use but much faster, and “a bit worse at most other things”.

In plain terms. OpenAI’s four models work like the tiers of a company vehicle fleet. Luna is the city car, Astra the lorry. Terra is the van: it does not carry everything, but it handles most everyday journeys at a much lower cost than the lorry.

Half the price of GPT-6.1 Sol

Price is the main selling point of this release. OpenAI charges $1 per million input tokens and $5 per million output tokens for GPT-6.1 Terra, half the rate of GPT-6.1 Sol and GPT-5.6 Terra.

API prices for OpenAI’s GPT-6 range as of 27 October 2026 (per million tokens)
Model	Input	Output	Positioning
GPT-6 Luna	$0.10	$0.50	Fast, simple tasks
GPT-6.1 Terra	$1	$5	Office work, vision
GPT-6.1 Sol	$2	$10	Complex code
GPT-6 Astra	$10	$50	Most powerful

The price per token does not tell the whole story. Artificial Analysis also measures the real cost of a task, which depends on how many tokens the model produces to reason. At maximum effort, a task from its index costs $0.40 with Terra, against $1.40 with GPT-5.6 Terra: less than a third.

In plain terms. The price per million tokens is the price per litre. The cost per task is the fuel for the whole journey. A model that is cheaper per litre can use more if it reasons for a long time; that is why cost per task is the figure to compare.

Close to Claude Fable 5.1? What the figures say

OpenAI claims that Terra matches GPT-6 Astra, its most powerful model, on the Agents’ Last Exam test, at a tenth of the price per token. This result comes from OpenAI and covers a single test. The ranking from Artificial Analysis, an independent evaluator, gives a more nuanced picture.

Score and cost per task on the Artificial Analysis Intelligence Index, as of 27 October 2026. GPT-6.1 Terra at maximum effort: 48 points, 0.40 dollars per task. Claude Fable 5.1 at medium effort: 49 points, 2.98 dollars. GPT-6.1 Sol at maximum effort: 52 points, 0.72 dollars. GPT-6 Astra at maximum effort: 53 points, 3.26 dollars. Claude Fable 5.1 at maximum effort: 53 points, 7.63 dollars. Claude Opus 5.5 at maximum effort: 58 points, 5.98 dollars. Claude models are measured with fallback, and Claude Fable 5 is not ranked. — Terra matches Claude Fable 5.1 at medium effort, for a cost per task around seven times lower.

GPT-6.1 Terra against competing models: who measured what
Result	Figure	Measured by
Intelligence Index, maximum effort	48 (Sol: 52, Astra: 53, Fable 5.1: 47 to 53 depending on effort)	Artificial Analysis
Cost per index task	$0.40 (Sol: $0.72, Fable 5.1 max: $7.63)	Artificial Analysis
Agents’ Last Exam, maximum effort	60.0% (Sol: 58.6%, Opus 5.5: 63.2%)	OpenAI
Terminal-Bench 4.0, maximum effort	50.3% (GPT-5.6 Terra: 21.5%)	OpenAI
Error rate on trick questions, low effort	16.7% (GPT-5.6 Terra: 53.6%)	OpenAI
Rate of guessed answers	38% (GPT-5.6 Terra: 88%)	Artificial Analysis

Three takeaways. Terra is not the strongest model: Claude Opus 5.5 keeps a 10-point lead on the index. However, it matches Claude Fable 5.1 set to low or medium effort (47 to 49 points), for a cost per task six to seven times lower. And Claude Fable 5, the previous version, does not appear in the ranking: no direct comparison has been published.

OpenAI says so itself: for the most complex code, it is better to stick with GPT-6.1 Sol or GPT-6 Astra.

Choose the right model for each task

Our training starts from your real use cases: which model, which setting, what cost.

ChatGPT training for companies Book a call

What GPT-6.1 Terra can do at the office

OpenAI presents Terra as the model for “everyday office work”. Its headline strength is computer use: it reads what is on the screen, clicks, types and switches from one application to another.

The examples cited by OpenAI, reported by The Next Web: entering incoming invoices into accounting software, filling in forms, drafting proposals, reading technical drawings and editing video. In a demonstration, Terra turned architectural plans into a 3D model and produced a timestamped log of an animated film, shot by shot.

OpenAI also reports tasks completed more than three times faster than with GPT-6 Astra and GPT-6.1 Sol. This speed matters most for agents, which chain dozens of actions.

In plain terms. Computer use is what an intern does when shown the screen: they look for where to click, fill in the fields, move on to the next piece of software. The model does not go through a technical connection between applications; it uses the screens as a human would.

Three business use cases to start with

1. Accounting: entering supplier invoices

This is the example highlighted by OpenAI. Terra reads the invoice, opens the accounting software and fills in the fields. First task to hand over: a batch of invoices already entered, so you can compare. Watch out: check every entry before approval, as you would for a new starter. Our AI training for finance covers this kind of check.

2. Sales: commercial proposals

OpenAI highlights “clearer and more natural” writing. Terra can draft a first version of a proposal from a brief and past offers. First task: a template proposal for a fictitious client. Watch out: prices and commitments must still be written by the team.

3. Engineering office: reading drawings

Terra’s vision lets it interpret technical drawings. First task: extract the list of components from a drawing you already know and compare it with the existing bill of materials. Watch out: a misread drawing is expensive, so always keep a human review.

Security and reliability: what to know

OpenAI rates Terra at the “critical” level for cybersecurity in its risk assessment framework, the Preparedness Framework. In other words, the model is judged capable enough to assist with serious cyberattacks. OpenAI therefore applies the same safeguards as for GPT-6 Astra, according to The Next Web.

On reliability, the progress is clear. According to Artificial Analysis, Terra guesses in 38% of cases where it does not know, against 88% for the previous version. OpenAI also says it stays better within the scope of the task requested and flags its limits more often.

For company data, the rule does not change with the model: keep sensitive documents for ChatGPT’s business plans, and write down what can and cannot go into them.

How to get started this week

Plus, Pro, Enterprise and Edu subscribers can select Terra in ChatGPT Work and Codex. Developers call it in the API under the name gpt-6.1-terra. The launch articles do not specify access for free accounts or for the Business plan.

Choose a repetitive task you know how to check: data entry, a form, a document summary.
Test Terra at two effort levels, medium then maximum, and note the time taken and the quality obtained.
Compare the cost per task, not the price per token, with the model you use today.
Keep Sol or Astra for complex code, as OpenAI recommends.
Train your teams to choose the model according to the task. That is the purpose of our ChatGPT training for companies.

Where to start in your company?

A 30-minute conversation to identify the tasks where a model like Terra saves your teams time.

Book a call AI diagnostic

Prefer a full programme? Our corporate AI training courses start from your own tools and real files.

Key takeaways

GPT-6.1 Terra, released on 26 October 2026, is OpenAI’s mid-tier model, between Luna and Sol.
It costs half as much as GPT-6.1 Sol: $1 and $5 per million tokens.
On the Artificial Analysis index, it matches Claude Fable 5.1 at medium effort for around seven times less per task, but remains behind Sol, Astra and Opus 5.5.
It is designed for office work: computer use, vision, writing.
The favourable comparisons with Astra come from OpenAI and cover a single test.
Test it on a task you know how to check, and compare the cost per task.

Sources

OpenAI, Introducing GPT-6.1 Terra, 26 October 2026.
Artificial Analysis, GPT-6.1 Terra pushes the cost efficiency frontier, 26 October 2026.
Artificial Analysis, model leaderboard, accessed on 27 October 2026.
The Register, OpenAI rolls out budget-priced mid-tier GPT-6.1 Terra, 26 October 2026.
BleepingComputer, OpenAI launches GPT-6.1 Terra with 3x faster AI agents, 50% lower pricing, 26 October 2026.
The Next Web, OpenAI launches GPT-6.1 Terra, a cheaper model for everyday computer work, 26 October 2026.
Interesting Engineering, OpenAI’s GPT-6.1 Terra matches GPT-6 Astra at 10% of the token cost, 26 October 2026.
9to5Mac, OpenAI upgrades ChatGPT and Codex with new GPT-6.1 Terra model, 26 October 2026.

FAQ

Answers to your questions about AI in business.
How to train your teams, identify the right use cases and roll out AI effectively.

Book a call

How can I train my teams in AI without it taking months?

It is the question we hear most often. The short answer: with the right method, a single day is enough for your teams to really start changing their habits. At AI Sisters, we have built a three-step approach, tested on more than 350 projects:

Step 1 - Understand, not just "discover"
We start with an AI Show: a live, immersive session with real-time demos on real cases. No generic slides. Your teams leave with a clear view of what AI can (and cannot) do for them.

Step 2 - Practise on real topics
Hands-on workshops are organised by role: HR, Finance, Marketing, Sales, Legal, Ops. We use your real documents, your real emails, your real tools. Prompts are tested live.

Step 3 - Embed new habits at D+30
One month later, we run an impact review: what has been adopted, what is still blocking, what needs adjusting. This follow-up is what makes the difference.

In short: training your teams in AI does not take months. With an approach built on immersion, role-based practice and impact follow-up,

Can AI training really be adapted to my sector (healthcare, finance, industry...)?

Yes, and that is exactly what separates useful training from training people forget.

Before each project, AI Sisters carries out a pre-audit: we analyse your typical documents (contracts, emails, reporting, product sheets), your tools (Office suite, Google, CRM, ERP) and your workflows.
We identify where AI can create value in your specific context. Then we choose, from our 40 business experts, the one who knows your sector best.
An HR team in an industrial group will not see the same examples as a B2C sales team. Prompts, exercises, use cases - everything is adapted.

‍In short: AI Sisters adapts each training programme to the reality of your sector, thanks to a pre-audit and 40 business experts. Healthcare, finance, industry, services: content is built on your real documents and your real constraints, not on generic examples.

Can AI training be funded by my OPCO?

Yes, in most cases.
AI Sisters training is Qualiopi-certified, which makes it eligible for OPCO funding (French training funds). The level of funding depends on your skills operator, the scheme used and your available budget, but in many cases a significant part of the cost can be covered.

In practice, we support you through the whole application: detailed quote, compliant programme, supporting documents, follow-up until reimbursement.

‍In short: thanks to Qualiopi certification, AI Sisters training is OPCO-fundable. The amount varies with your operator and your situation, but it can mean substantial savings.

Is AI training better on-site or remote?

Both work. What matters is the intention behind the format.
‍
On-site is ideal for a strong kick-off: a seminar day, a moment of collective co-creation, a team starting out together.

Remote is perfect for multi-site or international teams - our online sessions use training assistants via chat to keep things interactive (Q&A, co-writing prompts, live exercises).

Either way, the method stays the same: AI upskilling + practice + follow-up. And every session is led by a business expert, not a generalist trainer.

‍In short: on-site or remote, the format adapts to your organisation. AI Sisters keeps the same level of quality and interactivity in both cases, with business experts dedicated to each session.

Which tasks can I really hand over to AI in my team today?

The best way to answer is by role.

‍HR: screen CVs, generate job descriptions, write emails to candidates, prepare interviews.
‍Finance: explain an Excel table, write a commentary on results, prepare a forecast budget.
‍Marketing / Comms: create an editorial calendar, adapt a message into several formats, A/B test hooks.
‍Sales: write prospecting emails, personalise sequences, prepare responses to tenders.
‍Legal: analyse an NDA, flag risky clauses, generate a first draft to review.
‍Leadership: summarise committee meetings, prepare strategy notes, build impact scenarios.

In our AI workshops, we do not show fictional cases. We take your real documents and test AI on them, live.

Before AI: 2 hours to write meeting minutes.
With AI: 20 minutes.
‍
Before AI: 3 hours to analyse an Excel file.
With AI: 30 minutes.

‍In short: AI can speed up dozens of practical tasks in every department. AI Sisters helps you identify the ones with the most impact for your team, and test them straight away on your own documents.

My team is not technical. Can they still use AI?

Yes. In fact, that is the whole point.

Our approach is 100% role-based, with no technical jargon. We start from your teams' real problems (lost time, repetitive tasks, heavy reporting) and show how to turn them into simple prompts.

At the end of the workshop, each participant leaves with ready-to-use prompt templates adapted to their job.

The goal is not to train data scientists. It is for every employee to be able to say: "This is how AI helps me in a typical day."

‍In short: You do not need technical skills to use AI every day. AI Sisters turns the tools into accessible, role-based practices, and gives every participant prompts they can use the very next day.

How can I be sure AI does not make mistakes in a sensitive context (legal, medical, financial)?

It is a more than legitimate question, and we address it directly in our training.
We do not just show what AI can do. We also stress what should not be delegated to it, how to review and validate its answers, and how to combine a clear business brief with a well-built prompt to get the best quality.

Our experts know the constraints of your sector: compliance, deadlines, accuracy, GDPR, approved tools.
Nothing is left to improvisation.

‍In short: AI Sisters builds the limits and risks of AI into its training. Your teams learn to use AI reliably, knowing what they can hand over to it and how to check the results.