OpenAI has officially launched three new GPT-5.6 series models built on a fresh architecture, while also introducing a multi-step, long-context intelligent agent tool named "ChatGPT Work" within its unified, upgraded system ecosystem. Industry analysts note that major AI research organizations are rapidly converting their foundational technological advantages into high-barrier commercial vertical applications. The deep linkage between "reasoning compute investment" and per-call costs signifies the formal entry of enterprise AI applications into an era of meticulous cost control.
Release data for the new models indicates the GPT-5.6 series comprises three core models: the flagship "Sol," the balanced "Terra," and the fast "Luna." Regarding commercial access rights, the flagship Sol model is exclusively available to paid subscribers (Plus, Pro) and enterprise users, with no free access. The Terra model aims to balance processing speed with computational efficiency, primarily serving free-tier and basic paid users. The Luna model is specifically designed with a lightweight architecture for high-speed response scenarios.
Addressing specific application scenario overhauls, OpenAI engineering lead Thibault Sottiaux stated that the newly launched ChatGPT Work agent tool is now fully integrated across desktop and mobile ecosystems. It can directly call external applications and structured files, independently executing complex, long-form workflows such as research retrieval, data analysis, document creation, presentation development, and website building. Sottiaux emphasized that while the new platform offers a more cost-effective office alternative for non-coders, the dedicated Codex application for professional software developers will continue to receive long-term technical support and updates.
In terms of industry pricing and operational costs, technological evolution is driving profound structural changes. Industry research expert Simon Willison conducted benchmark cost analyses for the GPT-5.6 series and core models from competitor Anthropic. The data shows that, excluding additional computational overhead, the Luna model is priced at $1/$6 per million input/output tokens. The Terra model costs $2.50/$15, while the flagship Sol model reaches $5/$30. This pricing positions it in direct competition with Anthropic's Claude Opus series at $5/$25 and Claude Fable 5 at $10/$50.
Willison pointed out that as large models fully integrate an adjustable "Reasoning Effort" mechanism, the traditional static per-token pricing model no longer accurately reflects the actual expenditure for enterprise users. In the latest Pelican benchmark tests, due to variations in user-set reasoning depth and computational involvement for the same prompt input, the cost for a single generation could surge from 0.71 cents to 48.55 cents—a nearly 70-fold difference.
In response, OpenAI's official strategy team recommends that both enterprises and individual users, when deploying new AI workflows, should initially test with the most suitable model set to the lowest reasoning effort. They should then gradually increase the reasoning compute level based on output quality to avoid unnecessary budget overruns within compute bottlenecks and call limits. Mainstream technology think tanks widely believe that accurately assessing the "cost-performance inflection point" of AI tools in specific application scenarios will become a core challenge for enterprises seeking to achieve digital cost reduction and efficiency gains in the next phase.