OpenAI’s Sol/Terra/Luna Naming and Codex Merge: Making GPT‑5.6 a Deployable Platform

Clear Naming Redefines Model Hierarchy

OpenAI adopts a three‑tier naming scheme—Sol, Terra, Luna—that mirrors Anthropic’s Opus/Sonnet/Haiku. The new system separates generations by numbers and tiers by names, allowing users to infer that Sol > Terra > Luna indicates increasing capability. This resolves the previous confusion caused by labels such as o4‑mini, high, o3, and o1.

From Chat App to Codex Execution Shell

The ChatGPT app’s main screen now functions as a task‑oriented entry point for Codex. The traditional Chat mode is hidden behind a toggle, indicating that OpenAI has moved the execution engine to the primary interface. Codex, once a tool for programmers, becomes the underlying engine for all ChatGPT interactions, turning the product from a simple Q&A bot into a platform that can “receive a task and keep driving it to completion.”

Strategic Consolidation of Resources

Removing the Codex desktop app and folding its capabilities into the ChatGPT desktop app is not a cosmetic redesign but a strategic convergence. With limited compute and GPU resources, maintaining multiple entry points would increase tuning costs, user‑education overhead, and risk capability fragmentation. A single, unified entry point reduces these costs, stabilizes user habits, and concentrates compute power.

GPT‑5.6 Demonstrates End‑to‑End Task Completion

A video test shows GPT‑5.6 Sol creating a complete game in a single conversation: it designs rules, sound effects, interaction logic, and builds a playable version. This moves the model’s value from “strong language understanding” to “realizable task execution.”

New Evaluation Criteria for AI Product Managers

This ability can complete a full end‑to‑end task.

It can close the loop “goal → execution → result.”

The value lies more in “getting things done” than in sounding human.

Consequently, the next competitive edge will be measured by who can reliably deliver tasks at lower cost.

Agents’ Last Exam Benchmark

The newly introduced “Agents’ Last Exam” benchmark evaluates agents across 55 industry‑long‑thread workflows. Sol scores 53.6, Claude Fable 5 scores 40.5, and Terra/Luna also outperform Fable 5 while using only 1/16 of the cost. The key insight is that cost‑effective, stable task delivery matters more than raw model strength for enterprise adoption.

Atlas Shutdown Signals Resource Focus

The gradual shutdown of the Atlas AI browser indicates OpenAI’s decision to concentrate resources on a single, high‑traffic product rather than spreading effort across experimental entry points. This reflects a platform‑centric strategy that avoids internal competition and maximizes the impact of limited compute.

Competition Shifts from Model Scores to Landing Experience

The author likens the model to an engine and the product to a car: a powerful engine matters only if the car drives well. The real competition is not against Anthropic’s models but against products that deliver a smooth, reliable “landing experience.” By embedding Codex’s execution logic into the broader ChatGPT interface, OpenAI upgrades the product foundation rather than merely adding a feature.

For AI product managers, the decisive question becomes: is the product a “chatbot” or an “executable‑task platform”? Prioritizing task delivery over dialogue polish will determine who gains the next competitive advantage.