OpenAI has made its GPT-5.6 model family generally available following a limited preview, launching three models for different performance and cost requirements: Sol, Terra, and Luna.
Sol is the flagship model, Terra is intended for general-purpose workloads, and Luna is the lowest-cost option. OpenAI said the family was trained to complete tasks with fewer tokens across coding, professional workflows, cybersecurity, and scientific research.
GPT-5.6 Sol also introduces new reasoning configurations. The max setting gives the model more time to explore alternatives, run checks, and revise its approach, while ultra coordinates four agents in parallel by default.
OpenAI said ultra runs several workstreams in parallel. Developers can build similar multi-agent workflows through a beta capability in the Responses API.
OpenAI reported a score of 53.6 for GPT-5.6 Sol on Agents’ Last Exam, an evaluation of long-running professional workflows across 55 fields. The company said this was 13.1 points above Claude Fable 5 using adaptive reasoning.
At medium reasoning, Sol exceeded Fable 5 by 11.4 points at approximately one-quarter of the estimated cost. Terra and Luna also outperformed Fable 5 on the same evaluation at about one-sixteenth of the estimated cost, according to OpenAI.
On the Artificial Analysis Intelligence Index, which covers agentic work, coding, scientific reasoning, and other capabilities, GPT-5.6 Sol with maximum reasoning finished within one point of Fable 5 while completing tasks in 61% less time and at roughly half the estimated cost.
GPT-5.6 can also write and execute lightweight programs that coordinate tools, process intermediate results, monitor progress, and determine subsequent actions. Through Programmatic Tool Calling in the Responses API, these programs can filter intermediate data before returning relevant information to the model.
OpenAI described GPT-5.6 Sol as its highest-performing coding model to date. On the Artificial Analysis Coding Agent Index, Sol with maximum reasoning scored 80, 2.8 points above Fable 5.
The company said Sol used less than half the output tokens, completed tasks in less than half the time, and cost about one-third less than the comparison model. Terra scored slightly above Fable 5, while Luna outperformed Opus 4.8.
OpenAI also reported its highest results to date on Terminal-Bench 2.1 and DeepSWE, which test command-line workflows and long-running engineering tasks in real codebases.
GPT-5.6 can inspect rendered interfaces after generating them and make further revisions to visual or functional elements.
It can also turn natural-language instructions into interactive explanations and visualisations within ChatGPT Work. OpenAI did not provide a separate numerical benchmark for these design capabilities in the material released.
GPT-5.6 can work with documents, presentations, spreadsheets, and workplace data from services including Slack, Notion, Microsoft 365, and Google Drive.
GPT-5.6 Sol scored 92.2% on BrowseComp and 62.6% on OSWorld 2.0. On OSWorld, OpenAI said the model exceeded Opus 4.8 while using 85% fewer output tokens.
OpenAI said Luna nearly matched GPT-5.5’s highest performance at less than half the estimated cost, while Terra surpassed it at a lower cost.
For presentation work, GPT-5.6 can create editable decks from prompts and source files while identifying recurring elements in reference material, including layouts, typography, spacing, colours, content patterns, and Slide Master rules.
OpenAI also reported improvements in following reference formats for documents and spreadsheets, including handling equations, financial models, typography, page layouts, and worksheet structures.
On ExploitBench2, which measures progress from identifying vulnerable code through arbitrary code execution, GPT-5.6 scored 73.5%, compared with 47.9% for GPT-5.5 at a similar output-token budget.
On ExploitGym3, which tests whether agents can turn real-world vulnerabilities into working exploits, GPT-5.6 reached a 24.9% peak pass rate under a two-hour limit, up from 15.1% for GPT-5.5. With six hours, GPT-5.6 reached 33.7%.
The model also scored 71.2% on SEC-Bench Pro, compared with 45.8% for GPT-5.5. The benchmark evaluates proof-of-concept generation for complex software vulnerabilities.
OpenAI said GPT-5.6 can support defensive work including code review, patching, threat modelling, vulnerability triage, malware analysis, detection engineering, and patch validation.
Some advanced cyber capabilities remain available through OpenAI Daybreak’s Trusted Access for Cyber programme, which gives verified individuals and organisations additional access for authorised defensive work.
Individual participants must enable Advanced Account Security using hardware-backed passkeys by September 1 to retain access to OpenAI’s most cyber-capable frontier models. Those who do not meet the requirement will return to the default level of access.
OpenAI said it is also applying additional restrictions to high-risk entities and jurisdictions.
OpenAI reported gains for GPT-5.6 Sol across scientific research evaluations covering biology, life-science research workflows, and chemistry, although it did not provide individual scores for each benchmark in the material released.
The company is also using GPT-5.6 internally for AI research tasks including diagnosing failures, optimising training systems, running experiments, interpreting results, tuning kernels, and improving training recipes.
During internal testing, average daily output tokens per active researcher were more than double the highest level previously recorded for GPT-5.5, according to OpenAI.
Over the preceding six months, the share of research compute allocated to internal coding inference increased 100-fold, while internal agentic token use rose approximately 22-fold. OpenAI said these figures measure adoption rather than research progress.
OpenAI said the GPT-5.6 family underwent its most extensive safety evaluation programme to date before general availability.
Testing included human red teaming, automated evaluations, work with external experts, and approximately 700,000 A100-equivalent GPU hours of black-box automated red teaming.
The company said GPT-5.6 is more capable than earlier models in cybersecurity and biology but does not cross its Critical capability threshold in either field.
In cybersecurity, OpenAI’s testing found that GPT-5.6 performed better at identifying and fixing vulnerabilities than at reliably conducting autonomous end-to-end attacks against hardened targets. In biology, the company said the model can assist legitimate research but does not provide the complete capabilities required to create, engineer, or synthesise a highly dangerous novel biological threat.
OpenAI has introduced layered safeguards that combine protections trained into the model with real-time checks, continuous monitoring, reasoning-based review, and account-level enforcement.
A reasoning monitor reviews conversations for potential harm rather than relying only on classifier flags. OpenAI said the monitor can be updated without retraining classifiers from scratch.
The company said cyber safeguards on GPT-5.6 Sol block roughly 10 times more potentially harmful activity than those used on previous models. Users encountering unnecessary blocks in ChatGPT and Codex can retry prompts on lower-capability models.
OpenAI said safeguards will continue to be revised based on evaluation results, researcher findings, monitoring, and observed misuse. The company has also introduced a rapid-remediation process alongside its existing security and biology bug bounty programmes.
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events, click here for more information.
Tech Wire Asia is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。