GPT-5.6 Is Already Running in Some Codex Sessions
Developers have found a technique to detect whether GPT-5.6 Sol is already serving their Codex sessions — and some sessions are returning signals consistent with the new model before any public rollout.
Image by OpenAI
Three days after OpenAI restricted GPT-5.6 to a small group of government-vetted partners, developers have found a way to check whether the new model is already serving their Codex sessions — and some sessions are returning signals consistent with GPT-5.6 Sol, not the GPT-5.5 they expected to be running.
The detection method is built around a hidden system-prompt parameter OpenAI embeds in each model's context: a numeric value developers are calling the Juice value. GPT-5.5 at maximum reasoning intensity returns a different Juice value from GPT-5.6 Sol. Community posts, first surfaced by TechTimes on June 29, describe differing results across subscription tiers — consistent with an A/B test or a phased rollout that was not publicly disclosed.
What the Juice Value Is
OpenAI embeds a numeric parameter inside each model's hidden system prompt — the instruction layer that operates beneath any developer-set context. The Juice value is not publicly documented. Developers discovered it by constructing a specific XML-structured prompt that causes the model to surface the parameter in its response.
By comparing the arithmetic result returned across sessions, developers can distinguish which model is actually handling their requests. GPT-5.6 Sol and GPT-5.5 at maximum reasoning return different values. The Codex analytics panel and the /status CLI command also surface model information, though same-day usage data may lag by up to 24 hours.
What Developers Found
Community reports on developer forums and X indicate that some ordinary Codex users — not government-vetted partner organizations — are receiving GPT-5.6 Sol responses in sessions nominally set to GPT-5.5 at maximum reasoning. The pattern appears limited to Codex; the web-based ChatGPT interface was not returning the GPT-5.6 Juice value in the same tests, which is consistent with OpenAI's stated rollout position.
Results varied by subscription tier, pointing toward an A/B test being run across user cohorts rather than a deliberate, targeted deployment to specific accounts.
OpenAI's Stated Position
OpenAI announced on June 26 that GPT-5.6 — a three-model family named Sol (flagship), Terra (balanced), and Luna (low-cost) — was available only through the API and Codex to a select group of trusted partners. Access to each organization in the cohort was shared with the U.S. government, following a June 2 executive order directing agencies to build evaluation frameworks before broad model releases.
OpenAI described the arrangement as temporary and said it did not think government-gated access "should become the norm." Broader availability across ChatGPT, Codex, and the open API was described as "in the coming weeks," with no specific date given. OpenAI did not respond to a request for comment on the Juice-value detection findings before the TechTimes report was published.
Why This Matters for Developers
The detection raises a structural question about model serving transparency. If a provider can silently swap models beneath a session pinned to a specific version, developers building workflows around model-specific behaviors — Codex agents, evaluation pipelines, regression tests — need reliable ways to verify what they are actually running.
For developers currently using Codex on GPT-5.5 at maximum reasoning intensity, the practical question is whether GPT-5.6 Sol's behavior is already influencing their results without an explicit opt-in or notification.
What's Unconfirmed
OpenAI has not confirmed whether any rollout to non-partner Codex users is intentional. The pattern could represent an A/B test, infrastructure bleed-over from the partner deployment, or test traffic that was not properly isolated from production sessions. No formal post-incident statement or changelog entry has been issued as of publication.
GPT-5.6 Sol costs $5 per million input tokens and $30 per million output tokens when it becomes broadly available — pricing confirmed at the June 26 announcement alongside Terra ($2.50/$15) and Luna ($1/$6). Broader general availability is expected in the weeks following the restricted preview, with mid-July cited by analysts as a plausible window.
