OpenAI has announced the general availability of GPT-5.6, its new flagship model family. The launch follows a limited preview period and introduces three distinct models: Sol, the top-tier flagship; Terra, a mid-range option for everyday work; and Luna, the fastest and most affordable of the three.
The timing matters. AI model competition has intensified sharply in 2026, with Anthropic's Claude Fable 5 and Google's Gemini 3.1 Pro all vying for the same enterprise and developer budgets. OpenAI's answer with GPT-5.6 is not just raw performance, but efficiency. The company says it trained the model to extract more useful work from every token generated, which translates directly into lower costs for businesses running large-scale AI workflows.
GPT-5.6 is available now across ChatGPT, Codex, and the OpenAI API, with a global rollout completing over the first 24 hours after launch.
Three models, one architecture
The GPT-5.6 family is structured around three capability tiers, each designed for a different use case and budget:
- Sol is the flagship, designed for complex, demanding tasks. Priced at $5 per million input tokens and $30 per million output tokens.
- Terra is the balanced middle option, competitive with GPT-5.5 at $2.50 input / $15 output per million tokens.
- Luna is the fastest and cheapest, at $1 input / $6 output per million tokens.
OpenAI also introduced two new performance settings. “max” gives the model more time to reason through difficult problems, explore alternatives, and check its own work. “ultra” goes further by running four agents in parallel by default, trading higher token usage for faster results on demanding tasks. On benchmarks like BrowseComp and SEC-Bench Pro, adding parallel agents consistently produced stronger results in less time.
Benchmark performance and what it means in practice
The numbers OpenAI cites are significant. On Agents' Last Exam, a benchmark measuring performance across 55 professional fields over long-running workflows, GPT-5.6 Sol scores 53.6, which is 13.1 points above Claude Fable 5's adaptive reasoning score. Even at medium reasoning settings, Sol beats Fable 5 by 11.4 points at roughly one-quarter the estimated cost.
For coding specifically, GPT-5.6 Sol with max reasoning scores 80 on the Artificial Analysis Coding Agent Index, 2.8 points above Fable 5, while using less than half the output tokens and costing about one-third less. The model also sets new records on Terminal-Bench 2.1 and DeepSWE, which test complex command-line work and real-world software engineering tasks.
On computer use, GPT-5.6 Sol hits 62.6% on OSWorld 2.0, surpassing Claude Opus 4.8 while using 85% fewer output tokens. On BrowseComp, it reaches 92.2%.
The efficiency gains extend down the product line too. Terra and Luna both outperform Claude Fable 5 on the Agents' Last Exam benchmark at around one-sixteenth the estimated cost, which is a meaningful advantage for companies running high volumes of routine tasks.
A smarter approach to coding and design
GPT-5.6 introduces Programmatic Tool Calling in the Responses API, which lets the model write and run lightweight programs that coordinate tools, process intermediate results, and decide the next action without needing a developer to script every step. This reduces the number of model round trips required for complex tool-heavy tasks and makes the overall workflow more efficient.
On the design side, OpenAI says GPT-5.6 can take high-level instructions and produce functional, polished user interfaces. Its computer use capabilities also allow it to inspect the rendered output, not just the underlying code, so it can catch visual and layout issues before handing results back. This is a practical step forward for anyone building front-end products with AI assistance.
Early customers reported real productivity gains. Lovable, an app-building platform, said GPT-5.6 completes production-grade app workflows with roughly 25% fewer steps and 35-48% fewer tool calls than the previous model, while reducing stuck runs by 15%.
Knowledge work and document output
OpenAI has put clear effort into making GPT-5.6 useful for professional document work. The model can pull context from Slack, Notion, Microsoft 365, and Google Drive and turn it into polished outputs. It also creates fully editable presentations from scratch, inferring design systems, layouts, typography, and spacing conventions from reference decks and applying them consistently to new material.
The improvement is especially visible in financial modeling and equity research outputs. GPT-5.6 handles equations and financial models with greater precision and follows complex formatting templates more faithfully than GPT-5.5, which is important for repeatable, professional-grade work.
Cybersecurity: stronger offense and defense, with safeguards to match
GPT-5.6 is OpenAI's strongest cybersecurity model to date. On ExploitBench2, which tracks progress from locating vulnerable code through to arbitrary code execution, it scores 73.5% versus GPT-5.5's 47.9% at a comparable token budget. On ExploitGym3, which tests whether agents can turn real-world vulnerabilities into working exploits, GPT-5.6 nearly doubles GPT-5.5's peak pass rate, going from 15.1% to 24.9% under a two-hour cap, and reaching 33.7% with six hours.
OpenAI is careful to frame this as a dual-use capability. The same skills that could help an attacker find a vulnerability can help a defender reproduce it and build a fix. Overblocking, the company argues, creates its own security risk by preventing legitimate defensive work while bad actors continue using other tools.
To manage this, OpenAI is expanding its Trusted Access for Cyber program through OpenAI Daybreak. Verified individuals and organizations can access more of the model's defensive capabilities for tasks like vulnerability triage, malware analysis, and detection engineering. Individual members will need to enable Advanced Account Security with hardware-backed passkeys by September 1 to keep access to the most capable cybersecurity features.
Science and self-improvement
On life sciences benchmarks, GPT-5.6 Sol shows clear gains over GPT-5.5 across genomics, biology research workflows, and chemistry. On GeneBench Pro, it scores 28.7%, up from GPT-5.5's 12%, while using fewer tokens and less time.
OpenAI also reports that GPT-5.6 is accelerating its own internal research. During internal testing, average daily output tokens per active researcher more than doubled compared to the peak observed with GPT-5.5. Over the past six months, the share of research compute devoted to internal coding inference grew 100-fold, and internal agentic token usage increased roughly 22-fold. On the company's internal RSI (recursive self-improvement) index, GPT-5.6 Sol scores 16.2 points higher than GPT-5.5.
Safety architecture and what's new
OpenAI says GPT-5.6 has its most extensive safety system yet. The approach layers protections trained into the model with real-time checks, continuous monitoring, and account-level enforcement. A key addition is a reasoning monitor that reviews conversations in context to assess potential for harm, rather than relying solely on lower-intelligence classifier models.
Before launch, the company ran approximately 700,000 A100e GPU hours of black-box automated red teaming, alongside human red teaming and external expert testing. GPT-5.6 Sol's cyber safeguards now block roughly ten times more potentially harmful activity than the previous model's safeguards did.
The system is designed to be updated quickly. Because some protections use test-time reasoning, OpenAI says it can close gaps without retraining classifiers from scratch, which matters in a field where new jailbreaks appear regularly.
OpenAI acknowledges that no security system is perfect and says it will continue pairing its existing bug bounty programs with a new rapid-remediation process and ongoing monitoring. GPT-5.6 does not cross the “Critical” threshold in either biology or cybersecurity under OpenAI's internal safety evaluations.
How to access GPT-5.6
Access varies by plan and platform:
- ChatGPT: Plus, Pro, Business, and Enterprise users get GPT-5.6 Sol through medium and higher effort settings. Pro and Enterprise can also select Sol Pro for the most demanding tasks.
- ChatGPT Work and Codex: Free and Go users access Terra. Plus and above can choose between Sol, Terra, and Luna, and set effort levels for each. max is available to all GPT-5.6 users; ultra is available to Pro and Enterprise in ChatGPT Work, and Plus and above in Codex.
- API: Developers can access all three models. Programmatic Tool Calling is Zero Data Retention compatible. Multi-agent support is available in beta through the Responses API.
GPT-5.6 also introduces more predictable prompt caching. Cache writes are billed at 1.25x the uncached input rate, while cache reads continue to receive a 90% discount. A minimum 30-minute cache life and support for explicit cache breakpoints are also included.
