OpenAI announced on Thursday that its GPT-5.6 family of models is now available broadly following a limited preview. The release comprises three models aimed at different use cases: Sol, the company’s new flagship; Terra, positioned as a balanced option for routine professional tasks; and Luna, designed to be the most economical choice.
OpenAI highlighted performance metrics comparing GPT-5.6 Sol to competing systems across multiple independent evaluations. On Agents’ Last Exam - an evaluation focused on sustained professional workflows spanning 55 fields - GPT-5.6 Sol achieved a score of 53.6. OpenAI reported this outperformed Claude Fable 5 with adaptive reasoning by 13.1 points. With medium reasoning enabled, Sol was said to top Fable 5 by 11.4 points while operating at roughly one-quarter of the estimated cost, according to OpenAI’s comparisons.
The company also cited results on the Artificial Analysis Intelligence Index, which aggregates performance across agentic work, coding, scientific reasoning, and broader capabilities. There, GPT-5.6 Sol with maximum reasoning came within one point of Fable 5 while finishing tasks in 61% less time and at around half the estimated cost.
For coding-specific evaluations, OpenAI reported that GPT-5.6 Sol at maximum reasoning scored 80 on the Artificial Analysis Coding Agent Index, 2.8 points higher than Fable 5. The company said Sol used fewer than half the output tokens, required less than half the time to complete tasks, and cost about one-third less than the competitor in these coding trials.
OpenAI introduced a capability setting labeled ultra, which by default coordinates four agents in parallel. The firm described this mode as trading increased token consumption for stronger results and quicker completion on demanding tasks, effectively prioritizing performance and speed over token efficiency.
Independent assessments referenced by OpenAI included feedback from vendors and users. Itamar Friedman, co-founder and CEO of Qodo, was quoted saying that GPT-5.6 was the strongest model his company evaluated on agentic code-review tests. Friedman noted it outperformed GPT-5.5 on F1 while consuming about three times fewer tokens per pull request and achieving roughly half the median latency.
On knowledge work measures, GPT-5.6 Sol set new highs on BrowseComp at 92.2% and on OSWorld 2.0 at 62.6%. OpenAI said that on OSWorld, Sol surpassed Opus 4.8 while using 85% fewer output tokens.
OpenAI also provided cybersecurity benchmark results. GPT-5.6 scored 73.5% on ExploitBench1 under a given output-token budget, compared with 47.9% for GPT-5.5 at a similar token level. In the ExploitGym2 evaluation, Sol reached 24.9% under a two-hour cap, nearly double GPT-5.5’s peak pass rate of 15.1%.
The company stated that GPT-5.6 models show improved capabilities in both biology and cybersecurity relative to earlier versions but noted they do not cross the Critical threshold in either domain. To manage risks, OpenAI described layered safeguards that combine protections baked into the models with real-time checks, continuous monitoring, and account-level enforcement.
Prior to release, the models underwent approximately 700,000 A100e GPU hours of automated, black-box red teaming and were subject to extensive human expert red teaming. OpenAI presented these steps as part of its safety and deployment process.
GPT-5.6 is being rolled out across ChatGPT, Codex, and the OpenAI API starting Thursday, with a global rollout expected to continue toward full availability over the following 24 hours.
OpenAI published API pricing for the new family. For Sol the rates are $5 input and $30 output per 1 million tokens. Terra is priced at $2.50 input and $15 output per 1 million tokens. Luna is set at $1 input and $6 output per 1 million tokens. Cache writes are billed at 1.25 times the model’s uncached input rate, and cache reads receive a 90% cached-input discount.
Availability and immediate use cases
The announcement makes the three GPT-5.6 models available to developers and customers via multiple OpenAI products and the company’s API. The set of claimed benchmark advantages and the new ultra mode are positioned to appeal to users prioritizing speed and outcome quality on complex workflows, while Terra and Luna offer cost-to-performance trade-offs for routine tasks and budget-sensitive applications.
Technical and safety summary
OpenAI emphasized both performance and safety: the firm quantified pre-launch red teaming investment in GPU hours, reported human expert testing, and described layered mitigation measures. It also reported performance gains in areas including agentic workflows, coding, knowledge work, and cybersecurity benchmarks while explicitly stating the models do not exceed Critical thresholds in sensitive domains according to its internal assessments.
Key takeaways, risks and market sectors affected
See the boxed sections following this article for concise key points and supported uncertainties related to the GPT-5.6 launch.