AI tools use tokens to read information and create answers. Longer conversations, larger documents and more advanced models generally use more tokens.
For most people, managing AI token costs comes down to three simple habits:
- Choose the right model for the task.
- Give the AI relevant information.
- Keep conversations focused.
Theta Assist supports models from several AI providers and provides tools to help organisations understand and manage usage.
Model information in this article is current as of 8 September 2026 and may change.
What are AI tokens?
Tokens are small units of information processed by an AI model. They can represent complete words, parts of words or punctuation.
Tokens are used when the AI:
- Reads your instructions
- Reads documents and earlier messages
- Uses connected tools
- Works through a problem
- Creates its response
The final answer is therefore only one part of the total usage.
Choose the right AI model for the task
Different AI models offer different levels of capability and cost.
Using an advanced model for a simple rewrite may cost more than necessary. Using a basic model for difficult work can also become expensive if the result needs several corrections.
Choose the lowest-cost model that can complete the work accurately and reliably.
Theta Assist supports models from multiple providers. OpenAI’s current models provide a useful example of how to match the model to the work.
GPT-5.6 Luna: straightforward tasks
GPT-5.6 Luna is designed for cost-sensitive, high-volume work.
Use it for tasks such as:
- Formatting content
- Rewriting emails
- Sorting or classifying information
- Extracting details
- Creating short summaries
- Producing simple first drafts
GPT-5.6 Terra: everyday business work
GPT-5.6 Terra balances capability and cost. For many people, it will be a sensible everyday choice.
It is suited to:
- Business writing
- Research and summaries
- Meeting preparation
- Drafting reports and proposals
- Analysing several sources
- Developing recommendations
GPT-5.6 Sol: complex professional work
GPT-5.6 Sol is designed for work requiring deeper analysis or stronger technical skills.
It can help with:
- Strategy development
- Detailed analysis
- Advanced coding
- Complex research
- Executive or customer-facing material
- Work where a strong first result can reduce rework
GPT-6 Astra: the most demanding work
GPT-6 Astra is OpenAI’s most capable model. It is designed for advanced research, coding, problem-solving and longer multi-step tasks.
It is also significantly more expensive. Reserve it for work where its additional capability can improve the outcome, save substantial staff time or reduce business risk.
For routine writing, extraction and summaries, GPT-5.6 Luna or Terra will usually offer better value.
Consider cost per successful task
The price per token tells only part of the story.
A lower-cost model may need several attempts to produce an acceptable result. A more capable model may cost more per request but complete the work correctly the first time.
Consider:
- The quality of the result
- The number of attempts required
- The time spent reviewing and correcting it
- The importance and risk of the task
- The value of completing the work faster
The best measure is often the total cost of completing the task successfully.
AI provider pricing varies
Theta Assist supports models from several AI providers. Each provider has its own features and charging structure.
Costs can be influenced by:
- The selected model
- The amount of information provided
- The length of the answer
- The level of reasoning required
- Repeated or cached information
- Connected tools and searches
- Large images, audio, video or documents
- The speed or processing option selected
This makes direct price comparisons difficult. A model with a low advertised price may still cost more for a particular task.
For a wider view of how AI pricing has changed, read AI API prices tumble, but GenAI subscriptions hold their ground.
Keep conversations focused
Long conversations gradually collect earlier questions, answers, documents and tool results. Later requests may need to process much of this history again.
Start a new conversation when moving to an unrelated task. Keep work together when previous decisions or information remain important.
For example, a conversation about preparing a board report has little useful context for planning a staff event.
How Theta Assist manages growing conversations
Theta Assist already uses prompt caching where supported. This can lower the cost of repeatedly using the same instructions, reference material and conversation history.
Theta Assist is also adding conversation compaction. This will reduce the amount of older material carried into later requests while aiming to retain the information needed to continue the work. Administrators will be able to configure when compaction occurs because context limits and pricing rules vary between models.
Provide relevant information
Large documents can be useful, but include them when they directly support the task.
Where possible:
- Upload the relevant document rather than an entire folder.
- Provide the required section instead of a complete report.
- Remove duplicated material.
- Explain which information should guide the answer.
- Use document-search features to retrieve relevant passages.
Asking the AI to focus on one part of a document can improve relevance, but the full document may still be processed if it was submitted in full.
Ask for the result you need
Clear instructions can reduce unnecessary output and repeated attempts.
For example:
- “Summarise this in five bullet points.”
- “Write a 150-word email for staff.”
- “Create a table showing risks, owners and next steps.”
- “Recommend one option and explain it briefly.”
- “Identify the five issues that need a decision.”
Include the audience, purpose and required format in your request.
Use analytics to understand AI usage
Theta Assist analytics show usage across users, assistants and models.
These reports can help identify:
- Routine work being sent to premium models
- Assistants using large instructions or knowledge sources
- Conversations becoming unnecessarily long
- Requests producing more detail than users need
- Tasks requiring repeated attempts
- Complex work where a stronger model may save time
Usage reports provide a starting point for investigation. Combining them with workflow reviews and user feedback gives a clearer view of whether AI use is delivering value.
Focus on value
The goal is to achieve reliable business outcomes at an appropriate cost.
Use GPT-5.6 Luna for straightforward work, GPT-5.6 Terra for everyday business tasks, GPT-5.6 Sol for complex professional work and GPT-6 Astra for the most demanding assignments.
Models from other providers may be a better fit for particular tasks, budgets or organisational requirements. Theta Assist gives organisations the flexibility and visibility to make those choices.
Book a demo to see how Theta Assist can help your organisation improve model selection, manage AI token costs and get more value from enterprise AI.
