Brief IA
›

OpenAI Releases Guidelines for GPT-6.2 Astra and Sol

AI Tools·Tom Levy·

OpenAI Releases Guidelines for GPT-6.2 Astra and Sol

OpenAI Releases Guidelines for GPT-6.2 Astra and Sol
Key Takeaways
1OpenAI recommends reviewing the system prompts of GPT-6.2 Astra and Sol.
2Some changes in the API erase previous reasoning.
3New code tests are reserved for explicit requests, repository requirements, or specific regressions.
Why it matters : erasing reasoning may force the model to restart its work, increase token consumption, and degrade responses.

OpenAI introduces new prompts for GPT-6.2 Astra and Sol, launched on November 17, and invites developers to review their system instructions. These guidelines cover writing, autonomy, and coding, while the API alters reasoning retention.

The API erases reasoning in certain cases

Modifying the conversation history in the response API now deletes the reasoning already performed with GPT-6.2. The same erasure occurs when switching to certain models or changing organization or end user. The model may then have to restart its work, leading to increased token consumption and potential degradation of responses. The temperature and top_p settings are no longer offered.

For any new integration, OpenAI recommends a "medium" reasoning level; configurations set to "none" or "minimal" should switch to "low." Up to 272,000 tokens, Astra is priced starting at $8 per million tokens for input and $40 for output. Beyond this threshold, the rates rise to $16 and $60, respectively, while cache reads are charged at $0.80 per million tokens. Long sessions can quickly exceed this limit: Codex thus automatically reduces the context to 272,000 tokens.

OpenAI positions Astra for the most demanding tasks and presents Sol as a more cost-effective option for complex jobs. Ted Sanders from OpenAI reports on the choice between models based on quality requirements and budget constraints.

OpenAI regulates work interruptions

According to OpenAI, when a request lacks clarity, GPT-6.2 may propose a continuation and then wait for validation. The company advises asking it to execute the permitted steps without seeking this confirmation. If details are missing, it should clarify reversible assumptions and then proceed.

La lettre du soirGratuit · 19 h

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Stopping should be reserved for an actual obstacle, a major change in scope, a decision belonging to the user, or a destructive or irreversible operation. OpenAI also indicates that a higher reasoning effort facilitates final checks on long tasks.

More concise and autonomous responses

The ready-to-use instructions are included in the documentation and aim for behaviors observed in both models. According to OpenAI, Astra is inherently more concise and direct than its predecessors: an additional instruction for brevity is therefore worth reconsidering. For daily professional use, the company recommends writing in the active voice, starting with the essentials and using the vocabulary of the user as well as that of the sources.

A text intended for external readers must take into account the context they lack; a revision should respect the author's phrasing and meaning. In long tasks, abbreviations arising from reasoning or exchanges between agents may appear in the output. A dedicated instruction requests a final response that is understandable without these exchanges.

Guidelines for code and interfaces

OpenAI claims more efficient code with GPT-6.2 than with previous versions. Its guidelines ask to rely on existing logic, without multiplying abstractions or unnecessary auxiliary functions. Additional testing is only recommended if the user requests it, if the repository mandates it, or if it serves to identify specific regressions. Shortening variables or reducing lines does not, in itself, constitute simplification.

To create an interface, developers are encouraged to specify their typographic choices and images. By default, the model prefers realistic visuals, whether generated or sourced, and reserves SVGs for explicit requests. The guidelines exclude decorative numbering, filler badges, and explanations that merely reiterate displayed elements.

Brief IA — L'actualité IA en français

L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.