OpenAI has launched GPT-5.6, a new family of frontier AI models designed to deliver stronger performance while using fewer tokens and completing complex assignments more efficiently.

The release introduces three distinct model tiers: GPT-5.6 Sol for demanding professional work, GPT-5.6 Terra for balanced everyday use and GPT-5.6 Luna for faster, more economical workloads. GPT-5.6 is now being introduced across ChatGPT, ChatGPT Work, Codex and the OpenAI API.

More than a conventional model upgrade, GPT-5.6 reflects OpenAI’s growing focus on AI systems that can plan, use tools, operate software and complete end-to-end assignments—not simply generate answers.

Three models for three levels of work

OpenAI has divided GPT-5.6 into three capability tiers, giving users and developers more control over the balance between intelligence, speed and cost.

  • GPT-5.6 Sol: OpenAI’s flagship model for advanced coding, professional analysis, cybersecurity, scientific research and complex agentic work.
  • GPT-5.6 Terra: A balanced model intended for everyday professional tasks, offering strong performance at a lower cost than Sol.
  • GPT-5.6 Luna: The fastest and most affordable member of the family, designed for high-volume or cost-sensitive workloads.

The number identifies the model generation, while Sol, Terra and Luna represent continuing capability tiers. OpenAI says these tiers may evolve on their own release schedules, potentially making model selection more consistent as new generations arrive.

Better results from fewer tokens

Efficiency is one of the central themes of GPT-5.6. OpenAI says the models have been trained to extract more useful work from each token, reducing the number of model interactions and the amount of generated content required to complete a task.

This matters because the operational cost of an AI system depends on more than its advertised price per token. A model that requires fewer prompts, shorter outputs, fewer tool calls and less human correction may be less expensive overall, even if its individual token price is higher.

According to OpenAI’s published evaluations, GPT-5.6 Sol achieved a score of 53.6 on Agents’ Last Exam, an assessment covering long-running professional workflows across 55 fields. OpenAI reports that Sol exceeded Claude Fable 5 on this evaluation while requiring fewer tokens and delivering a lower estimated cost.

Benchmark results should still be treated as indicators rather than guarantees. Actual performance will depend on the quality of the prompt, connected tools, source data, workflow design and review process.

Ultra mode introduces parallel AI agents

GPT-5.6 introduces an ultra capability setting for particularly demanding work. Instead of relying on one agent to process an entire assignment sequentially, ultra coordinates four agents in parallel by default.

Each agent can work on a separate part of the problem before the results are combined. For example, a complex market assessment could divide the work across competitor research, customer analysis, financial modelling and regulatory review.

OpenAI says this approach can improve both the quality and completion time of difficult tasks, although it uses more tokens than a single-agent workflow.

The model also supports a max reasoning setting. Max gives GPT-5.6 additional time to explore alternatives, conduct checks and revise its approach before delivering a result. Ultra goes further by adding parallel agent execution.

GPT-5.6 is OpenAI’s strongest coding model so far

OpenAI describes GPT-5.6 Sol as its most capable coding model to date. The company reports improvements across software implementation, command-line workflows, code review and long-running engineering assignments.

On the Artificial Analysis Coding Agent Index, GPT-5.6 Sol with maximum reasoning recorded a score of 80. OpenAI says it completed the evaluated work using fewer output tokens and less time than several competing frontier models.

GPT-5.6 can also write and run lightweight programs that coordinate tools, process intermediate results and determine the next action as work progresses. This can reduce the need for developers to manually script every step of an agentic workflow.

Within the Responses API, a new Programmatic Tool Calling capability allows the model to filter large volumes of intermediate information and retain only the material required for the next step. OpenAI says this can reduce token consumption and the number of model round trips required for tool-heavy applications.

From rough instructions to polished interfaces

OpenAI is also positioning GPT-5.6 as a significant improvement for interface and visual design. The model can use high-level instructions to create functional websites, applications and interactive visualisations.

Its computer-use capabilities allow it to inspect the rendered result rather than evaluating only the underlying code. This means the model can potentially identify visual inconsistencies, usability problems and functional issues before refining its work.

This combination of code generation and visual inspection is important. Previous AI models could produce technically valid interfaces that still required substantial human work to improve spacing, responsiveness, navigation and overall presentation.

Stronger documents, spreadsheets and presentations

GPT-5.6 is designed to transform information from documents and connected workplace services into editable business outputs. OpenAI specifically identifies Slack, Notion, Microsoft 365 and Google Drive as examples of sources that can contribute context to a task.

Through ChatGPT Work, the model can use these materials to create presentations, reports and spreadsheets. OpenAI says GPT-5.6 is more capable of following existing templates and reproducing their layouts, typography, colours, spacing and recurring design patterns.

This could be particularly useful for organisations that need AI-generated work to follow established brand and reporting standards. Potential applications include:

  • Producing management presentations from financial and operational data
  • Updating an existing slide deck while preserving its design system
  • Creating financial models with editable formulas and structured worksheets
  • Combining research from several sources into an executive report
  • Turning a business brief into an interactive explanation or visualisation

The ability to produce polished material does not eliminate the need for human review. Users should verify calculations, source attribution, assumptions and conclusions before relying on AI-generated outputs for important decisions.

Major gains in cybersecurity capability

OpenAI says GPT-5.6 is also its strongest cybersecurity model. On ExploitBench 2, the company reports that GPT-5.6 Sol scored 73.5%, compared with 47.9% for GPT-5.5 at a comparable output-token budget.

The model can support authorised defensive activities such as secure code reviews, vulnerability analysis, patching, threat modelling and blue-team exercises. However, stronger cyber capabilities also create greater potential for misuse.

OpenAI has therefore introduced more restrictive safeguards for sensitive cyber activity. Certain advanced capabilities are available through OpenAI Daybreak’s Trusted Access for Cyber program, which requires verification for eligible individuals and organisations.

OpenAI says GPT-5.6 combines safeguards trained into the model with real-time checks, continuous monitoring and account-level enforcement. A reasoning monitor examines the context of potentially harmful requests, with the aim of distinguishing legitimate defensive work from malicious activity.

Availability across ChatGPT, Codex and the API

GPT-5.6 is being introduced globally across OpenAI’s major products, although the available models and capability settings depend on the user’s subscription.

  • ChatGPT: Plus, Pro, Business and Enterprise users can access GPT-5.6 Sol through medium and higher reasoning settings. Pro and Enterprise customers can also select GPT-5.6 Sol Pro for demanding tasks.
  • ChatGPT Work and Codex: Free and Go users receive access to GPT-5.6 Terra. Paid users can choose between Sol, Terra and Luna and select an effort level.
  • Max mode: Available to users who have GPT-5.6 access in ChatGPT Work and Codex.
  • Ultra mode: Available in ChatGPT Work for Pro and Enterprise users, and in Codex for Plus and higher subscriptions.
  • OpenAI API: Developers can access all three model tiers and use capabilities including Programmatic Tool Calling and beta multi-agent execution.

GPT-5.6 API pricing

OpenAI has announced the following API prices per one million tokens:

ModelInput tokensOutput tokensGPT-5.6 SolUS$5.00US$30.00GPT-5.6 TerraUS$2.50US$15.00GPT-5.6 LunaUS$1.00US$6.00

GPT-5.6 also introduces more predictable prompt caching. Cache writes are charged at 1.25 times the model’s standard input-token rate, while cache reads receive a 90% discount from the normal input price.

What GPT-5.6 means for businesses

The most important aspect of GPT-5.6 may not be a single benchmark result. It is the combination of reasoning, tool use, computer control, document generation and parallel agent execution within one model family.

For businesses, the three-tier structure creates a practical deployment strategy. Luna can support high-volume routine work, Terra can manage everyday professional workflows, and Sol can be reserved for tasks where stronger reasoning justifies the additional cost.

Organisations should avoid automatically deploying the most powerful model for every use case. A more sustainable approach is to test each workflow against defined measures for accuracy, completion time, token use, human review and total cost.

GPT-5.6 moves OpenAI closer to an AI platform that does not merely help users think about work, but can actively plan, execute, inspect and refine that work. Whether it delivers meaningful productivity gains will ultimately depend on how well organisations connect these capabilities to trusted data, controlled workflows and accountable human decision-making.