GPT-5.6 Sol looks like a work model, not a coronation

Zvi Mowshowitz describes GPT-5.6 Sol as a practical work model alongside the cheaper Terra and Luna. He lists Sol at $5 and $30 for input and output units, Terra at $2.50 and $15, and Luna at $1 and $6, while giving Fable the edge in raw intelligence.

Sol is framed as a worker for tasks, not a universal sage

The post contrasts Sol plus Codex or Work with Fable plus Claude Code or Cowork. In Zvi's view, Fable feels stronger as an architect and planner, while Sol is better at practical execution, computer use and web search.

Zvi also quotes OpenAI saying GPT-5.6 improved AI research work and that internal testing pushed average daily output tokens per active researcher to more than twice the highest level seen with GPT-5.5. OpenAI also cites a 100-fold increase in the share of research compute devoted to internal coding inference and about a 22-fold increase in agentic token usage over 6 months.

Cost per task is becoming as important as model IQ

The shift is that model debates are no longer only about who gives the best single answer. When an agent takes dozens of steps, calls tools and runs long jobs, cost and reliability in repeated work start to matter more than a one-off wow moment.

For developers, that changes workflow. A stronger model may design the plan, while a cheaper and more obedient one handles refactors, research, migrations, tests and repetitive edits. The practical question is where Sol saves time without needing more supervision than a human.

Early reactions are room temperature, not an eval

Zvi explicitly says he is collecting a gestalt from reactions and that early hype is biased. That caution matters. Expert reactions can reveal patterns, but they do not replace reproducible evals, company test suites and error measurements in agentic runs.

Claims that Sol helped post-train Luna also need precision. A controlled subtask in a supervised environment is not the same thing as autonomously training another model end to end.

The useful answer is a division of labor across models

The best signal will not be the absolute chat winner. It will be a task map: what goes to Fable, what goes to Sol, what goes to Terra or Luna and when to run two models side by side. Teams that measure that split on their own work will learn more than teams watching the leaderboard.

Lilith's verdict

Sol does not look like a throne for one model king. It looks like a new shift worker you send out with the cart, while the architect stays near the plans and the red pencil.