OpenAI’s Codex now routes work across three GPT-5.6 variants named Sol, Terra, and Luna, according to a post from eric provencher, a developer who writes about AI coding tools, published July 16 on X. Sol targets ambiguous problems where the payoff justifies extra compute. Terra covers routine implementation work. Luna handles fast tasks with a narrow, well-defined scope. A fourth option, Sol Ultra, adds a heavier reasoning pass and coordinates work across multiple agents on a single job.
The split turns Codex into a small menu of price points rather than a single model applied to everything. Sending a one-line bug fix to a reasoning-heavy tier burns tokens and adds latency for no benefit. Sending a multi-file refactor with unclear requirements to a bounded, fast-response tier risks a wrong answer delivered quickly. Provencher’s post frames the choice as a matching problem: the ambiguity and stakes of the task should decide which model handles it, not habit.
This mirrors a broader shift already visible in other coding agents, where a single “smart” model is being replaced by a router that picks strength based on task shape. Anthropic’s Claude Code and GitHub’s Copilot have both moved toward model selection inside the product rather than a fixed backend. Codex’s three named tiers make that internal routing explicit and put the decision in the user’s hands instead of hiding it behind an automatic classifier.
That choice has a direct cost consequence. Reasoning-heavy tiers like Sol and Sol Ultra are almost certainly priced and metered higher than Luna, since deeper reasoning and coordination across agents consume more compute per request. Teams that default to the highest tier for every task will pay a premium that most of their work does not need. Teams that default to the cheapest tier will see more failures on the fraction of tasks that actually require planning across files or systems.
Provencher’s guidance for prompting reflects this same logic. He advises that a clear prompt should specify the goal, the relevant context, the boundaries of the change, and a way to check the work is complete. Under a tiered system, an underspecified prompt is now a routing error as much as a communication failure: it forces the model, or the developer choosing the model, to guess at ambiguity that should have been resolved before submission.
The operator takeaway is a triage habit, not a technical one. Before assigning a Codex task, an operator should ask whether the problem’s difficulty is genuinely unclear or whether it only feels that way because the request is underspecified. Genuinely ambiguous, high-stakes work belongs on Sol or Sol Ultra. Bounded, well-understood tasks belong on Luna, where speed and lower cost carry no real risk. Getting that triage wrong repeatedly is now a line item, not just a workflow inefficiency.
eric provencher (@pvncher) described GPT-5.6’s three-tier Codex system in a post on X on July 16, 2026.
