OpenAI has pushed its GPT-5.6 family into general availability, splitting the lineup between flagship Sol and mid-priced Terra, two models built for very different workloads.
Key Points:
- Sol costs $5 per million input tokens and $30 per million output, double Terra's rate.
- Sol tops coding and agent benchmarks, while Terra targets routine work at lower cost.
- Early testing shows Terra can burn more output tokens than Sol on long jobs.
GPT-5.6 Sol And Terra Diverge On Pricing
The company announced the rollout across ChatGPT, Codex and its API on Jul. 9, ending a limited preview that ran for roughly two weeks. Sol, the flagship, targets complex coding, cybersecurity and long-running agent work, and OpenAI calls it its "best coding model yet." Terra sits one tier below, delivering performance competitive with GPT-5.5 at about half the cost.
Pricing draws the sharpest line between the two. Sol costs $5 per million input tokens and $30 per million output tokens, while Terra charges $2.50 and $15 for the same volumes, with a third model, Luna, priced at $1 and $6 for high-volume jobs.
Access splits along similar lines inside ChatGPT. Paid subscribers reach Sol through the medium and higher reasoning settings, while free and Go users get Terra in ChatGPT Work and Codex. GitHub has already added all three models to Copilot, framing Sol as the pick for reasoning over large codebases and Terra as the balanced default for everyday agentic coding.
Also Read: Is Charles Hoskinson Leaving Cardano? He Calls The Claim A Complete Lie
Sol Outpaces Terra In Coding Benchmarks
Sol scored 80 on the Artificial Analysis Coding Agent Index, which OpenAI says puts it 2.8 points above Anthropic's Claude Fable 5. Chief executive Sam Altman said the model is 54% more token efficient on coding tasks than the company's previous models, a pitch aimed squarely at enterprise budgets.
The numbers are not uniform, though. Sol's 64.6% score on SWE-Bench Pro still trails Claude Mythos 5 by roughly 15 points. Independent reviewers found that Sol nonetheless stays on task through long, messy repository work, while Terra suits scoped implementation and first-pass code review, with escalation to the bigger model kept in reserve.
Those same tests carried a warning for budget-minded teams. Terra's long coding run consumed more output tokens than Sol did, meaning cheaper rates do not always produce cheaper finished tasks, and teams should measure cost per solved job before moving heavy traffic.
The staged debut marked a shift in how OpenAI ships frontier models.
The company briefed the U.S. government on the family's capabilities ahead of launch and restricted early access to a small circle of vetted partners whose participation was shared with officials.
The generation also scraps the firm's old naming habits, since the number now marks the generation while Sol, Terra and Luna act as durable capability tiers that can advance on their own schedules.
Read Next: Palantir Sheds 50 Engineers And A Quarter Of Its Value: What Went Wrong?