Skip the cover

OpenAI

GPT-6.1 Terra

OpenAI's fast model for high-volume professional work, at $1 per million input tokens and rated Critical for cybersecurity.

By

Total
Undisclosed
Active
Undisclosed
Experts
Undisclosed
Context
1.05M tokens
Max output
128k tokens
Input
text, image
Output
text
Licence
Proprietary
Weights
API only

Not disclosed by OpenAI

Contents1 to 3

What is disclosed

OpenAI released GPT-6.1 Terra on 26 October 2026 and describes it as a fast, efficient model for high-volume professional work.11OpenAI, API changelog, entries of 20 and 26 October 2026, read 30 October 2026. https://developers.openai.com/api/docs/changelog 22OpenAI, Addendum to GPT-6 Astra System Card: GPT-6.1 Terra, 26 October 2026, sections on introduction, model data and training, HealthBench and cybersecurity. https://deploymentsafety.openai.com/gpt-6-1-terra It takes text and images and returns text. Audio and video are not supported.33OpenAI, GPT-6.1 Terra model page, read 30 October 2026. https://developers.openai.com/api/docs/models/gpt-6.1-terra

The window is 1,050,000 tokens, of which 922,000 can be input. A response can run to 128,000 tokens. The knowledge cutoff is 1 June 2026.33

The system card addendum says it uses "the same types of data and training as GPT-6 Astra".22 Its size and architecture are not published.

What is new

Terra sits between GPT-6 Luna and GPT-6.1 Sol in price. OpenAI compares it with its predecessor, GPT-5.6 Terra, and highlights image analysis and computer use.22

On HealthBench Professional, length-adjusted, it scores 62.7, against 57.7 for GPT-5.6 Terra.22 On the plain HealthBench its length-adjusted score is 55.1, which is 1.9 points lower than its predecessor's.22

OpenAI treats the model as Critical in cybersecurity and High in biological and chemical capability.22 It runs behind the same safeguards stack as Sol and Astra, with probes updated for compatibility.22

ExploitBench
An OpenAI test of whether a model can turn a known software vulnerability into working code execution.

On the Internal Port test of ExploitBench it reaches arbitrary code execution in 24.5% of attempts. GPT-6.1 Sol reaches 21.5% and GPT-6 Astra 31.5%.22

ExploitBench
An OpenAI test of whether a model can turn a known software vulnerability into working code execution.

On Sandbox Bench it solves 5 of 22 targets, against 9 for GPT-6.1 Sol and 10 for GPT-6 Astra.22

Using it

In the API the model is gpt-6.1-terra.33

Up to 272K input tokens, Standard pricing is $1 per million input tokens and $5 per million output tokens. Cached input costs $0.05.44OpenAI, API pricing, Standard table, read 30 October 2026. https://developers.openai.com/api/docs/pricing. Arithmetic: 1 / 2 = 0.5; 1 / 0.10 = 10; 5 / 0.50 = 10; 1 x 2.5 = 2.5. Above 272K, input doubles to $2 and output rises to $7.50.44

That is half the price of GPT-6.1 Sol, at $2 and $10 for input and output.44 It is ten times the price of GPT-6 Luna, at $0.10 and $0.50.44

On 20 October OpenAI added 12-hour prompt caching, with cache writes at 2.5 times the input price.11 For Terra a 30-minute cache write costs $1.25 and a 12-hour write $2.50.44

Notes

  1. OpenAI, API changelog, entries of 20 and 26 October 2026, read 30 October 2026. https://developers.openai.com/api/docs/changelog 2

  2. OpenAI, Addendum to GPT-6 Astra System Card: GPT-6.1 Terra, 26 October 2026, sections on introduction, model data and training, HealthBench and cybersecurity. https://deploymentsafety.openai.com/gpt-6-1-terra 2 3 4 5 6 7 8 9

  3. OpenAI, GPT-6.1 Terra model page, read 30 October 2026. https://developers.openai.com/api/docs/models/gpt-6.1-terra 2 3

  4. OpenAI, API pricing, Standard table, read 30 October 2026. https://developers.openai.com/api/docs/pricing. Arithmetic: 1 / 2 = 0.5; 1 / 0.10 = 10; 5 / 0.50 = 10; 1 x 2.5 = 2.5. 2 3 4 5

More from OpenAI

All models by OpenAI