GPT-5.6 Sol makes runtime strategy part of model selection

OpenAI’s July 9 product post positions GPT-5.6 Sol as the flagship of a three-tier family, alongside lower-cost Terra and Luna variants. The practical change for technical users is therefore broader than a single model upgrade: the company presents a choice among capability, latency, and token cost across ChatGPT, Codex, and the API. OpenAI lists Sol at $5 per million input tokens and $30 per million output tokens, versus $2.50 and $15 for Terra and $1 and $6 for Luna. That pricing ladder makes routing and evaluation policy part of the adoption decision, rather than an afterthought after choosing one default model.

OpenAI’s comparisons with GPT-5.5 point to gains that may matter most in tool-using engineering workflows. Its published table reports Sol at 72.7% on DeepSWE and 88.8% on Terminal-Bench 2.1, compared with 67% and 85.6% respectively for GPT-5.5. The same page describes Sol as the family’s flagship, while characterizing Terra as a lower-cost option competitive with GPT-5.5 and Luna as the fastest, most affordable tier. Those numbers are useful as a starting hypothesis for an evaluation plan, but they are provider-published results rather than a substitute for testing on an organization’s own repositories, tools, and failure modes.

The more consequential product shift may be in how work is executed. OpenAI says Programmatic Tool Calling in the Responses API can run in-memory programs to coordinate tools and process intermediate results, while its multi-agent capability remains in beta. It also describes an `ultra` setting that coordinates four agents in parallel by default, alongside a `max` setting that spends more time on reasoning and checking. For teams already paying for multiple model turns and tool calls, this creates a new tradeoff: orchestration can reduce hand-written control flow, but higher-effort or multi-agent modes should be budgeted and measured separately from ordinary single-agent requests.

There are material reasons not to read the launch table as a neutral scoreboard. OpenAI’s own July 8 audit says it found evidence that about 30% of SWE-Bench Pro tasks are broken, even though the GPT-5.6 comparison table includes SWE-Bench Pro results. The product page also mixes external benchmarks, internal evaluations, and partner-reported results, so individual rows do not carry the same evidentiary weight. The system card further documents real-time monitoring and restrictions for higher-risk cyber use cases, meaning capability tests may not predict every production interaction. The careful operational takeaway is to compare Sol, Terra, and existing models on representative tasks, with the intended tools, reasoning settings, safeguards, and cost limits enabled.