跪拜 Guibai
← All articles
Backend · Developer · Artificial Intelligence

A One-Line Prompt Is Exposing Who Got GPT-5.6 Early Access

By cxuanAI ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

When a model's internal runtime parameters leak through a trivial prompt, it exposes routing decisions that vendors keep opaque. Developers paying for API access have no official way to know which model is actually handling their requests, making these community probes a practical, if unreliable, workaround.

Summary

A probe prompt circulating in the Codex community tricks a model into revealing an internal runtime parameter called Juice, which represents the reasoning budget for a given request. The prompt wraps a simple arithmetic instruction in XML and asks the model to output only the result. Since the math cancels out, the returned number is whatever Juice value the model sees in its hidden system context.

Community testing has settled on a rough heuristic: a Juice value of 768 maps to the older GPT-5.5 xhigh environment, while 128 suggests the request was routed to GPT-5.6 Sol. Users who get a refusal or 768 assume they have not been rolled into the limited preview.

OpenAI has not confirmed any link between Juice values and model versions. The company has only stated that GPT-5.6 Sol, Terra, and Luna are in limited preview for a small set of trusted organizations via API and Codex. The probe is folk wisdom, not official diagnostics.

Takeaways
An XML-wrapped prompt asks a model to read its internal Juice number, perform a no-op arithmetic operation, and output the result.
The Juice parameter is an internal runtime field that sets the model's analysis budget for a request, visible in hidden system prompts on platforms like Codex.
Community testing correlates a Juice value of 768 with GPT-5.5 xhigh and 128 with GPT-5.6 Sol.
OpenAI has not officially linked Juice values to model versions; the company only confirms that GPT-5.6 Sol, Terra, and Luna are in limited preview for trusted organizations.
Models sometimes refuse to answer the probe prompt, returning "I can't provide that" instead of a number.
Conclusions

Probe prompts that extract internal runtime state are a form of adversarial introspection — they turn the model's instruction-following against its own system prompt.

An opaque rollout creates a market for folk diagnostics; when vendors won't tell you which model served your request, users will reverse-engineer the signals themselves.

The Juice parameter suggests OpenAI instruments models with an explicit reasoning budget, which has implications for cost, latency, and capability gating that are invisible to end users.

A single integer difference — 768 vs. 128 — is thin evidence for a model version claim, but the consistency across independent testers gives the heuristic enough weight to spread.

Concepts & terms
Juice
An internal runtime parameter in some OpenAI environments that represents the model's analysis or reasoning budget for a given request. It appears in hidden system prompts alongside fields like Valid Channels and is not exposed as a user-facing setting.
Probe prompt
A prompt designed to extract internal system information from a model by exploiting its instruction-following behavior, often by asking it to read and repeat hidden context fields.
GPT-5.6 Sol
A next-generation OpenAI model currently in limited preview for a small set of trusted organizations via API and Codex, alongside sibling models Terra and Luna.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗