GPT-6 Sol in Codex: budget credits and reviewer time together

GPT-6 Sol is rolling out in Codex and ChatGPT Work for complex coding and agentic tasks. The question for a team is not merely whether Sol can finish a task. It is whether the accepted result is worth the credits, retries, and human review compared with the team's current route. A short pilot can answer that without treating a published rate as a promise about your monthly allowance.

Know which unit you are counting

OpenAI's Standard-speed Codex credit rates, checked on 24 September 2026, are quoted per million tokens:

Those are Codex credit rates, not US-dollar API prices. Included subscription usage, remaining limits, account rules, and some Enterprise agreements need separate checks. API-key usage follows API pricing instead. OpenAI also lists Fast mode at 2.5 times the Standard credit rate for Sol and Luna where available; compare runs at the same speed.

At Standard rates, a hypothetical Sol run with 100,000 uncached input tokens, 100,000 cached input tokens, and 10,000 output tokens would correspond to 5 + 0.5 + 2.5 = 8 credits under the published token-rate formula. It is an illustration, not a bill for a real task or a guarantee that those token categories will appear in your account. Check your Codex usage dashboard for actual limits; the CLI's /status shows remaining limits during a session.

Run a pilot that records accepted work

Choose six to ten tasks before you start. Include a small known edit, a multi-file bug, a task requiring documentation, and a failure that needs recovery. Save the starting commit and acceptance checks. Keep the brief, tools, and review standard comparable across runs. Record each attempt, including failures, with these fields:

The dashboard may show account-level usage rather than an itemized charge for every task. Do not invent per-task credits from a plan allowance. Where attribution is unavailable, compare accepted work and reviewer time, and label the usage gap. If a failed attempt consumes time or credits, keep it in the denominator.

Route the next task by its failure cost

OpenAI positions Sol for complex coding and agentic work, and Luna for focused, high-volume tasks. That is a starting hypothesis. A small edit with clear tests may be economical on Luna. A cross-file change that repeatedly needs rescue may cost less in total staff time on Sol, despite its higher credit rate. Your pilot should decide which work goes where; a single token rate cannot.

Start at the default reasoning effort, then change one setting at a time for a specific failure. Stop or tighten scope when retries rise or the review queue grows. The Sol/Luna task-selection guide explains the initial routing decision, and the Opus 5.5 versus Sol pilot offers a common acceptance scorecard across providers.

Editorial note: This is a planning method, not a hands-on performance claim. Rates and usage guidance were checked against OpenAI's Codex pricing, model guidance, and usage-limit guidance on 24 September 2026. Availability and rates can change.