GPT‑5.6 is here and promises frontier power at a fraction of the cost

OpenAI announces in an official blog post that GPT‑5.6 is now generally available, following the limited preview it launched several weeks earlier. The family consists of three tiers: the flagship model Sol, the balanced option for everyday work Terra; and the most cost-efficient version Luna. According to OpenAI, all three models outperform rivals such as Anthropic’s Claude Fable 5 while using fewer tokens and costing less to run.

Three tiers built for different budgets

OpenAI positions performance per dollar as the central selling point of this release. On the company’s Agents’ Last Exam benchmark, which tests long-running professional workflows across 55 fields, Sol scores 53.6, which OpenAI says beats Fable 5 by 13.1 points. At medium reasoning effort, Sol still leads by 11.4 points while costing roughly a quarter as much, the company states. Terra and Luna reportedly match or beat Fable 5 at around one-sixteenth the cost.

Beyond the standard models, OpenAI introduces “ultra,” a setting that coordinates four agents working in parallel by default, and can scale to sixteen for tasks such as security testing. A new feature called Programmatic Tool Calling lets GPT‑5.6 write small in-memory programs that manage tool calls and filter data itself, reducing the number of round trips through the model.

Coding, design and defensive cybersecurity

OpenAI calls Sol its best coding model to date. On the Artificial Analysis Coding Agent Index, Sol reaches a score of 80, which the company says is 2.8 points above Fable 5 while using less than half the output tokens and about a third less cost. The company also highlights stronger “computer use,” meaning the model can inspect rendered interfaces and documents, not just generate code, and correct visual or functional issues before handing work back.

For knowledge work, OpenAI reports state-of-the-art results on BrowseComp (92.2 percent) and OSWorld 2.0 (62.6 percent), along with improvements in presentations, spreadsheets and financial models that follow existing templates more faithfully.

In cybersecurity, OpenAI states that GPT‑5.6 nearly doubles GPT‑5.5’s performance on exploit-development benchmarks. The company frames this primarily as a defensive gain, noting that the model is better at finding and patching vulnerabilities than at autonomously executing attacks against hardened targets. Access to the most sensitive capabilities is limited to vetted users through OpenAI’s Trusted Access for Cyber program.

Safety measures and rollout

OpenAI says it built its most robust safety system yet for this release, combining trained-in safeguards with real-time monitoring and account-level checks. The company reports that cyber-related safeguards for Sol now block roughly ten times more potentially harmful activity than for previous models, while offering a retry option on lower-capability models to reduce friction for legitimate users.

The rollout begins immediately and will reach full global availability within 24 hours, according to OpenAI. Pricing per million tokens is set at $5 input / $30 output for Sol, $2.50 / $15 for Terra, and $1 / $6 for Luna. Free and low-tier ChatGPT users receive Terra by default, while paid plans can choose between all three tiers and adjust reasoning effort.