Skip to content
Call wiki

Which Codex model, reasoning effort, and access settings matter?

Last updated 2026-08-17 · Confidence: documented — OpenAI’s Codex docs on models, sandboxing, approvals, and pricing; the after-medium plateau and the picker gap are reported.

Three settings do most of the work: model (GPT-5.6 Sol), reasoning effort (medium), and the sandbox’s network-access setting — whose default silently blocks web research.

Model: Sol is the flagship of the GPT-5.6 lineup

Section titled “Model: Sol is the flagship of the GPT-5.6 lineup”

Since the GPT-5.6 release (July 9, 2026), the picker offers three variants: Sol, Terra, and Luna — Sol has “the strongest capability for complex coding, computer use, research, and cybersecurity”; Terra is the balanced everyday model; Luna the fast, cheap one. A garbled model name on a call (“soul”) is Sol. Plans differ in what the picker shows, and it can glitch: an August 2026 issue reports Sol missing from the desktop picker on a Pro account.

Effort: medium — the default, and roughly the plateau

Section titled “Effort: medium — the default, and roughly the plateau”

Each model runs at Low through Ultra effort; Medium is the shipped default, with the docs advising more only “when the task needs deeper planning or analysis”. Alejo recommends medium with Sol: benchmark plots circulating on Twitter showed performance plateauing sharply after medium, so higher effort mostly buys latency.

Network access: off by default — the silent research-killer

Section titled “Network access: off by default — the silent research-killer”

The default workspace-write sandbox disables network access; the agent must ask to escalate, and on work accounts internet access is a Data-controls toggle. An agent that “couldn’t find” public documents and invents a reason (wrong language, paywall) often just lacked this setting — ask it to report which tools it had. A forum thread reports new chats launching network-disabled despite full access being enabled (reported).

Approvals and limits

Section titled “Approvals and limits”

Approval modes run from on-request (default in git folders) through an auto_review reviewer agent to danger-full-access — no sandbox, no approvals; see Isolation models and Permission modes. Plan limits: Plus ($20) gives Sol 10–100 messages per shared 5-hour window (Terra 25–200, Luna 250–2,000) plus weekly caps; Pro is 5x that at $100 or 20x at $200, and extra credits are purchasable. Training controls are covered in Codex training privacy.

Sources

Section titled “Sources”