OpenAI has just released their next-generation large language model family: GPT-5.6.
In this article, we will explore GPT- 5.6 and go into depth on Sol, Terra and Luna.
The models are designed to give for more flexibility in regards to maximum capability, balanced performance, and cost efficiency.
GPT-5.6 Sol is the flagship model whilst Terra and Luna allow options for those with specific performance and pricing requirements.
The tech giant claims that the release of GPT-5.will advance AI capabilities across software engineering, computer use, scientific research, cybersecurity, and professional knowledge work.
The model is now fully released to the public following a delay from the US government over safety concerns.
What is GPT‑5.6?
GPT-5.6 is OpenAI’s newest family of large language models.
The release introduces three models under the 5.6 umbrella names Sol, Terra and Luna. Each model has slightly different goals meaning it’s loaded with differing levels of capability, speed, and cost.
Rather than focusing exclusively on increasing model size, OpenAI says that GPT-5.6 has been developed to get more useful work from every token. This means the models can complete complex tasks while using fewer tokens, reducing the overall running costs.
These efficiency improvements extend across the entire GPT-5.6 family. The three models provide different levels of performance and affordability, allowing developers and organisations to select the model that best fits their needs.
The new models also introduce more capable AI agents that write and run lightweight programs that coordinate tools, process intermediate results, monitor progress, and actively decide what actions to take as the task progresses.
This allows GPT-5.6 to complete tool-heavy workflows with fewer model interactions and less guidance. This will help reduce the burden of developers manually scripting each step.
The models also feature higher capability for additional reasoning. This means the model can spend more time exploring, checking and refining before giving answers.
The ultra setting takes this further by coordinating four agents in parallel by default which lets the model to tackle complex tasks across multiple workstreams.
On the technical side, the Agents’ Last Exam evaluation, which measures long-running professional workflows across 55 fields, found that GPT-5.6 Sol achieved a score of 53.6. OpenAI claims that this surpassed Anthropics Claude Fable 5’s adaptive reasoning score by over 13 points.
On the Artificial Analysis Intelligence Index, GPT-5.6 Sol with maximum reasoning came within one point of Fable 5 while completing tasks in 61% less time and at roughly half the estimated cost.
What is GPT-5.6 Sol?
GPT-5.6 Sol is OpenAI’s latest flagship model. It is the capable model in the GPT-5.6 release.
The model is designed for users who require the highest level of performance across complex workloads, including coding, knowledge work, cybersecurity, and scientific research.
OpenAI claims that the GPT-5.6 Sol mode will set a ‘new standard; for intelligence and efficiency. They say it will deliver a stronger performance per dollar by completing more successful work for the same spend and achieving comparable results at a lower cost.
The model introduces new capabilities for handling more demanding tasks. Through stronger computer use and improved design judgment, GPT-5.6 Sol can inspect, refine, and deliver more polished results beyond generating initial outputs.
What is GPT-5.6 Terra
GPT-5.6 Terra is OpenAI’s balanced model in the GPT-5.6 release. It is designed for everyday work while maintaining strong performance across a range of tasks.
The model is designed for users who require a balance between capability and cost efficiency across everyday workloads, including coding, knowledge work, and AI-powered workflows.
OpenAI claims that GPT-5.6 Terra will deliver performance competitive with GPT-5.5 while offering improved cost efficiency. They say the model provides users with a more affordable option without sacrificing the capabilities needed for complex tasks.
The model gives developers and organisations another option within the GPT-5.6 family, allowing them to select a model that better matches their performance and pricing requirements.
What is GPT-5.6 Luna?
GPT-5.6 Luna is OpenAI’s most cost-efficient model in the GPT-5.6 release. It is designed to provide strong AI capabilities at a lower price point.
The model is designed for users who require an affordable option for running AI workloads while still benefiting from the capabilities of the GPT-5.6 family.
OpenAI claims that GPT-5.6 Luna delivers strong capability at its lowest price point. They say the model provides a more accessible option for developers and organisations looking to scale AI applications efficiently.
The model completes the GPT-5.6 family by offering a cost-efficient alternative alongside the flagship capabilities of Sol and the balanced performance of Terra.
Balancing Scaling and Safety
As AI models become more capable across the board it is vital that safety and security measures must evolve alongside them.
OpenAI reassures audiences that its new advanced GPT-5.6 models have the robust safety system to date that are designed to match the capabilities of each model while applying greater scrutiny for application in higher-risk applications.
In the cybersecurity space, OpenAI says that GPT-5.6 is stronger at identifying and fixing vulnerabilities than carrying out autonomous, end-to-end attacks against hardened targets.
This means the model can be used to support defensive security work, helping researchers find weaknesses, validate fixes, and improve protection against potential threats.
They apply the same logic to health and biology, saying that whilst GPT-5.6 can support research activities, the models cannot be used to create, engineer, or synthesise highly dangerous novel threats.
The tech giant has also taken a more conservative approach to applying safeguards. Compared with previous models, the company claim the new models safeguards block roughly ten times more potentially harmful activity.
For the release, OpenAI conducted its most extensive safety evaluation period to date, including human red teaming, external expert testing, and around 700,000 A100e GPU hours of automated testing. These evaluations were used to identify weaknesses, test jailbreak attempts, and strengthen protections before release.
Comments ( 0 )