New Delhi: OpenAI has moved its GPT-5.6 model family out of limited preview and into general availability, rolling it out globally across ChatGPT, ChatGPT Work, Codex, and the OpenAI API starting July 9, with full availability within 24 hours.
Rather than shipping one flagship model, OpenAI split the release into three tiers. Sol is the top-of-line model built for heavy reasoning, coding, science, and long-running agentic work. Terra is a mid-tier model that OpenAI says matches GPT-5.5-level performance at half the cost. Luna is the fastest and cheapest option in the lineup, designed for lighter, high-volume tasks.
On the Artificial Analysis Coding Agent Index, Sol running at maximum reasoning reportedly set a new high score of 80, which is 2.8 points above Anthropic’s Claude Fable 5, while using under half the output tokens, less than half the time, and roughly a third less cost.
OpenAI extended that comparison across the rest of the family, claiming Terra performs just above Fable 5 and Luna outperforms Claude Opus 4.8, both while using roughly a third of the time and about half the output tokens at close to a quarter of the cost.
On a broader benchmark, the Artificial Analysis Intelligence Index, Sol at maximum reasoning reportedly landed within a point of Fable 5’s score while finishing tasks 61 percent faster and at about half the estimated cost. Independent coverage also noted a gap that cuts the other way. On SWE-Bench Pro, Sol’s 64.6 percent score trails Claude Mythos 5’s 80.3 percent by roughly 15 points, a reminder that OpenAI’s comparisons don’t hold across every benchmark.
The release also brings a few notable technical additions. A new multi-agent mode called “ultra” runs four agents in parallel, which OpenAI says lifted its Terminal-Bench 2.1 score from 88.8 percent to 91.9 percent. A feature called Programmatic Tool Calling lets the model write and run its own JavaScript inside an isolated environment with no network access, allowing it to chain tool calls, loop, and process intermediate results before answering.
Prompt caching has also changed, with explicit cache breakpoints and a 30-minute minimum cache life, and cache writes are now billed at 1.25 times the uncached input rate while cache reads keep a 90 percent discount. OpenAI is also positioning the model for producing editable presentations, documents, spreadsheets, interfaces, and frontend prototypes with tighter layout accuracy.
On cybersecurity, OpenAI is calling GPT-5.6 its strongest model to date, built for defensive uses such as threat modeling, code review, patch development, and simulated internal attacks meant to expose weaknesses before real attackers find them. OpenAI says it briefed the U.S. government on the models’ capabilities ahead of launch, and at the government’s request, began with a limited preview for a small group of trusted partners before widening access. This follows earlier friction with the Trump administration over concerns about potential misuse of increasingly capable models.
Pricing per million tokens is set at 5 dollars input and 30 dollars output for Sol, 2.50 dollars and 15 dollars for Terra, and 1 dollar and 6 dollars for Luna. Access varies by plan. ChatGPT Plus, Pro, Business, and Enterprise users get Sol through medium and higher reasoning effort settings, with Pro and Enterprise able to select the higher capability Sol Pro, while Free and Go users on ChatGPT Work and Codex get Terra by default. The models are also rolling out to developer tools, with GitHub Copilot beginning to make Sol, Terra, and Luna available to its Pro Plus, Max, Business, and Enterprise tiers.
Previous
