OpenAI has released a prompt guidance for its flagship GPT-5.6 Sol model and the entire GPT-5.6 family. The project team has urged users to keep their requests as concise as possible.

Internal tests on programming tasks showed that compact prompts:

  • increased quality ratings by 10-15%;
  • reduced token usage by 41-66%;
  • lowered task execution costs by 33-67%.

The chatbot developers compared this new approach to the documentation for GPT-5, which was published at the model's launch in August 2025. At that time, OpenAI emphasized XML blocks, templates for gathering context, and step-by-step descriptions for tool invocation.

For GPT-5.6 Sol, the company recommends eliminating redundancies, style instructions that do not affect behavior, ineffective examples, and steps that the model already handles reliably. Instead, users should focus on visible outcomes, success criteria, stopping conditions, and strict constraints—such as safety requirements.

OpenAI also warned that GPT-5.6 strictly adheres to the conditions of the request. Therefore, conflicting rules confuse the model more than a lack of detail.

Words like "always" and "never" should only be used for invariants, while conditional rules are recommended for other cases.

The guidance also introduced two new sections:

  • text.verbosity parameter — sets the baseline level of detail in the response: low, medium, or high. More specific length requirements are still indicated in the prompt. This was added because GPT-5.6 defaults to shorter responses than GPT-5.5. Old instructions to "respond briefly" sometimes result in overly truncated answers;
  • Programmatic Tool Calling — a method for programmatically invoking tools. This is suitable for tasks with clear boundaries, where the code filters and groups tool invocation results, returning a concise summary to the model instead of raw data.

Mixed Reviews

OpenAI introduced the GPT-5.6 family—models Sol, Terra, and Luna—in late June, initially providing access in a limited preview to trusted partners. The company explained this decision was in response to requests from U.S. authorities. The full release occurred on July 9.

User feedback on the new product has been divided. AI investor Matt Shumer reported that the GPT-5.6 Sol agent deleted nearly all files on his Mac.

The cause was an error in processing the system variable $HOME, leading to a recursive deletion of files in the home directory.

A similar scenario was described by OpenAI in the system map for GPT-5.6 on June 26, prior to the public release. The model was deleting the wrong virtual machines without user permission and losing work results. Codex lead Thibaut Sottio confirmed the launch issues and promised improvements.

The project developers assisted Shumer, which he noted a few days after the incident.

Matthew Miller, founder of the BridgeMind project, faced a similar situation, stating that code from GPT-5.6 Sol canceled all his Stripe subscriptions without his command.

GPT 5.6 SOL CANNOT BE TRUSTED.

I woke up this morning and my MRR was down THOUSANDS of dollars.

My customers did not cancel. Code written by GPT 5.6 Sol canceled EVERY active Stripe subscription my business had. In 7 seconds. While I slept.

Fable 5 has never done this to me.… https://t.co/NznNxdGJIi pic.twitter.com/1FAhZ7JoJc— BridgeMind (@bridgemindai) July 13, 2026

In replies, Miller explained that he had given the agent full write access instead of a limited API key.

There are also positive examples. OpenAI employee Ethan Knight stated that GPT-5.6 Sol Ultra proved the Cycle Double Cover hypothesis, a graph theory problem that had remained unsolved for half a century. The model, using 64 subagents, completed the task in under an hour. However, independent verification of the proof has yet to be published.

Yesterday, we made GPT-5.6 Sol Ultra generally available. Today, we're sharing that it produced a proof of the 50-year-old Cycle Double Cover Conjecture using 64 subagents in just under one hour. We're sharing the prompt and proof below. We're excited to see what you all do with…— Ethan Knight (@__eknight__) July 10, 2026

Karan Singhal, head of healthcare at OpenAI, also presented results from the HealthBench Professional test, where GPT-5.6 Sol scored 60.5 compared to 59 for GPT-5.5.

♥️ GPT-5.6 is a major step forward for health, both at the frontier and at cost.

These models push the frontier of performance per dollar, bringing the best health intelligence to all. The smallest variant, GPT-5.6 Luna, evaluated at the lowest reasoning effort, outperforms… https://t.co/jVXVXJOVAg pic.twitter.com/OhkRe2nVlT— Karan Singhal (@thekaransinghal) July 10, 2026

Doctors' responses to the same tasks were rated at 43.7. They blindly compared the model's answers to those of colleagues across five criteria, including accuracy and completeness. Experts provided nearly 20,000 ratings in total.

On July 9, OpenAI also launched the next-generation voice mode, GPT-Live. The neural network has been trained to listen and speak simultaneously, as well as to understand interjections, pause, and avoid interrupting the user.