Daily · Article 02 · 5 min read
- OpenAI
GPT-6.1 Terra launches: the cost and checks behind computer work
Official announcement dated October 26; source check October 27 JST
Summary
OpenAI introduced GPT-6.1 Terra on October 26 for visual reasoning, computer use and writing. Standard API token rates are half those of GPT-6.1 Sol; permissions and result checks still shape an actual workflow.
OpenAI introduced GPT-6.1 Terra on October 26 to help read screens and documents and carry out computer work. Repetitive tasks such as transferring quantities and amounts from a delivery note into a business application connect that reading to actual data entry.
Key facts
OpenAI says Terra improves visual reasoning, computer use and writing over GPT-5.6 Terra. It gives interpreting drawings and filling out forms from documents as examples. The company announced availability in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and through the API as gpt-6.1-terra. Official announcement
Standard API rates per million tokens are $1 for ordinary input and $5 for output, half those of GPT-6.1 Sol. Tokens measure the volume of supplied material and generated responses; a cache that reuses repeated material has its own rates. Terra model documentation
1. The AI that reads a screen and the tool that clicks
A model produces answers and proposed actions from text or images. An API is the connection through which a company's software sends it requests. In a computer workflow, the model chooses the next action and a connected tool carries out permitted clicks or typing.
The tool returns the changed screen, which the model reads before choosing its next action. Data entry proceeds through this repeated cycle. Terra's tool calling uses the Responses API; the application connecting the tools maintains the working environment, manages permissions and cleans up afterward. Screenshots sent to the API count as image input. Computer-use tools
In this fictional delivery-note example, staff first set the source material and the scope of data entry. Until the draft quantities and amounts are ready, the AI reads the screen and chooses the next entry, then examines the new screen returned after the tool acts.
%% Fictional delivery-note entry · repeat screen work until the draft is ready
flowchart TD
A["AI: read source<br/>and current screen"] --> B["AI: choose<br/>field and value"]
B --> C["Tool: enter;<br/>return new screen"]
C --> AOnce the draft is ready, a staff member compares it with the original and approves saving. The tool saves only the approved scope, and staff check the saved result.
%% After the draft is ready · proposed approval, save and result checks
flowchart TD
A["Staff: compare<br/>and approve"] --> B["Tool: save<br/>approved scope"]
B --> C["Staff: check<br/>saved result"]The cat asks Q&A
The researcherHow can I check the entries before they are saved?
Pocket NotebookMake the approval step in this example a condition for the tool to save. If a value differs from the original, stop the save and compare the corrected entries again. Check the saved result afterward too.
2. Using a subscription or building with the API
ChatGPT Work is an AI work environment; Codex is an agent for development tasks. You can use connected tools in an existing product, or have a developer build an API integration for your business. API usage is billed separately from the product subscription. Subscription and API pricing
| Access route | What to prepare first | Where to check usage |
|---|---|---|
| ChatGPT Work / Codex | Whether the account can select Terra and connect the material and tools needed | The account's subscription or credit usage and limits |
| API in your application | Model ID, input material, connected tools and execution permissions | Per-request API usage and billing records |
- Access route
ChatGPT Work / Codex
- What to prepare first
- Whether the account can select Terra and connect the material and tools needed
- Where to check usage
- The account's subscription or credit usage and limits
- Access route
API in your application
- What to prepare first
- Model ID, input material, connected tools and execution permissions
- Where to check usage
- Per-request API usage and billing records
The cat asks Q&A
- The researcher
What can I try before involving a developer?
Pocket NotebookIf the account lets you select Terra, you can ask it to read a delivery note with confidential information removed and make a table of quantities and amounts. Entering those values directly into a business application depends on the product's connected tools and permissions.
3. What does the same amount of text cost?
Tokens do not always match characters or words. Repeated screen work accumulates input for the material and screens sent, and output for actions and answers, on each request. Output billing includes the model's internal reasoning tokens as well as the answer you can see. Reasoning tokens and billing
This illustrative calculation assumes one Standard API request without caching, using 100,000 text input tokens and 10,000 total output tokens. That output includes internal reasoning; this is not a measured delivery-note workflow cost. Sol model rates
| Model | 100,000 input tokens | 10,000 total output tokens | Sum of both items |
|---|---|---|---|
| GPT-6.1 Terra | 0.1 × $1 = $0.10 | 0.01 × $5 = $0.05 | $0.15 |
| GPT-6.1 Sol | 0.1 × $2 = $0.20 | 0.01 × $10 = $0.10 | $0.30 |
- Model
GPT-6.1 Terra
- 100,000 input tokens
- 0.1 × $1 = $0.10
- 10,000 total output tokens
- 0.01 × $5 = $0.05
- Sum of both items
- $0.15
- Model
GPT-6.1 Sol
- 100,000 input tokens
- 0.1 × $2 = $0.20
- 10,000 total output tokens
- 0.01 × $10 = $0.10
- Sum of both items
- $0.30
At equal token volumes, Terra's sum for these two items is 50% lower. Sending screenshots adds their image input tokens. The calculation excludes tool fees, the execution environment and regional-processing charges.
A cache reuses an unchanged prefix, the beginning of the input, across requests. Each input token is billed in one applicable category below. The cache-write rate is not added to ordinary input pricing; the chosen minimum retention period is reflected in that write rate. How caching and pricing work
| Terra input processing | Standard rate per million tokens | When it applies |
|---|---|---|
| Ordinary input | $1.00 | Material neither read from nor written to cache |
| Cache write, at least 30 minutes | $1.25 | Writing reusable material with the 30-minute option |
| Cache write, at least 12 hours | $2.50 | Writing reusable material with the 12-hour option |
| Cache reuse | $0.05 | Reusing a matching prefix on a later request |
- Terra input processing
Ordinary input
- Standard rate per million tokens
- $1.00
- When it applies
- Material neither read from nor written to cache
- Terra input processing
Cache write, at least 30 minutes
- Standard rate per million tokens
- $1.25
- When it applies
- Writing reusable material with the 30-minute option
- Terra input processing
Cache write, at least 12 hours
- Standard rate per million tokens
- $2.50
- When it applies
- Writing reusable material with the 12-hour option
- Terra input processing
Cache reuse
- Standard rate per million tokens
- $0.05
- When it applies
- Reusing a matching prefix on a later request
Request size and processing speed also change the rates.
- Above 272,000 input tokens, input and cache rates double and output rates become 1.5 times the base rate for that entire request. The example stays below this threshold.
- Fast mode, the API's faster-processing option, costs twice Standard.
The cat asks Q&A
- The researcher
What should I record when processing several delivery notes?
Pocket NotebookCollect usage for every request and the number of retries until each task finishes. Separate ordinary input, cache writes and cache reuse, and count all output including internal reasoning. Recording the time people spend correcting values also helps compare the workflow before and after adoption.
To close
Compare results and cost on one repetitive task
Terra's useful role spans reading material, entering values and reviewing the result. Tracking API usage and correction time until a delivery note's quantities and amounts are saved accurately shows what lower token rates mean for the actual work.
Sources
- Official announcementopenai.com
- Terra model documentationdevelopers.openai.com
- Computer-use toolsdevelopers.openai.com
- Subscription and API pricinglearn.chatgpt.com
- Reasoning tokens and billingdevelopers.openai.com
- Sol model ratesdevelopers.openai.com
- How caching and pricing workdevelopers.openai.com