OpenAI tells developers to rewrite system prompts for GPT-6.2 Astra and Sol as defaults shift
OpenAI says developers moving to GPT-6.2 Astra or Sol should review their system prompts. The company published copy-ready templates in the models' documentation, and the Responses API now discards earlier reasoning content if conversation history is edited.
OpenAI says developers moving to GPT-6.2 Astra or Sol should review their system prompts, and has published copy-ready templates in the models' documentation to curb behaviors it observed in both releases. The guidance, reported by The Decoder, also covers API changes: starting with GPT-6.2, the Responses API discards earlier reasoning content under certain conditions.
OpenAI describes Astra as built for the most demanding tasks, while Sol handles complex work at lower cost. The company shipped both models on November 17. Through the API, Astra costs $8 per million input tokens and $40 per million output tokens. According to OpenAI employee Ted Sanders, which model fits depends mainly on budget and quality needs.
Writing: shorter by default
According to OpenAI, Astra writes more briefly and directly than earlier models by default. Developers switching over should reconsider instructions such as be concise. For everyday colleague-style use, the recommended prompt tells the model to adopt user and source terminology, use active voice and put key points first. For external readers, it should account for their lack of context. When editing, it should preserve the author's meaning and phrasing.
On long tasks, OpenAI says the model sometimes carries shorthand from its reasoning or from other agents' messages into the final response. A dedicated prompt instructs it to treat the final answer as the user's first look at the work, make it fully self-contained, and avoid copying or closely paraphrasing its reasoning.
Initiative: assume, flag, continue
OpenAI says GPT-6.2 sometimes suggests a next step and waits for confirmation when user intent is unclear. The recommended prompt tells the model to carry out authorized next steps itself. It should pause only for:
- a major change in task scope
- a real blocker
- a decision the user must make
- an irreversible or destructive action
When details are unclear, the model should make reversible assumptions, flag them and keep working. OpenAI adds that higher reasoning effort helps the model follow through on final checks during long tasks.
Coding: less abstraction, fewer tests
OpenAI says GPT-6.2 writes more efficient code than its predecessors. Its template tells the model to:
- reuse existing classes, methods and APIs in the user's repo where correctness is preserved
- add new tests only when requested, required by the repo, or needed to catch a specific regression that existing tests would miss
- avoid fewer lines, shorter variable names or dense one-liners as substitutes for simplicity
- avoid unnecessary helper functions and class abstractions
For frontends, developers should specify imagery and typography directly. By default, the template favors realistic images from image generation or search, and uses SVGs only on request. It excludes decorative numbering, filler badges and helper text that repeats what the interface already shows.
API changes
Starting with GPT-6.2, the Responses API discards earlier reasoning content if developers edit conversation history, switch to certain other models, or change the end user or organization. OpenAI says this can hurt response quality and increase token usage as the model repeats work.
For new integrations, OpenAI recommends the medium reasoning level. Existing setups using none or minimal should move to low. The parameters temperature and top_p are no longer supported.
Background
The guidance follows earlier advice from OpenAI. In April, the company advised against carrying over old prompt stacks for GPT-5.5 and urged short, outcome-focused prompts. In September, it published GPT-6 Astra prompts to curb frequent follow-up questions and ban AI clichés. OpenAI developer Eric Provencher warned then that bloated skills and AGENTS.md files slow the model down. At that time, even small changes triggered thorough testing from Astra, whereas GPT-6.2 is meant to write new tests only on request.
What this means
Prompts tuned for earlier models may now work against the defaults. An instruction like be concise can stack on top of a model that already writes tersely, and test-writing instructions may conflict with the new request-only behavior. Teams migrating should audit system prompts instead of porting them unchanged.
The Responses API change carries a direct cost. Applications that edit conversation history, such as trimming or rewriting earlier turns, will lose prior reasoning content, which OpenAI says can raise token usage and lower quality. Removal of temperature and top_p means any pipeline that sets them needs code changes, and anyone running none or minimal reasoning needs to adopt low per OpenAI's guidance.
Related Articles
OpenAI Python SDK v3.53.0 adds GPT-6.2 Astra, Astra Law and Sol identifiers
OpenAI Python SDK v3.53.0, dated 2026-11-17, adds model identifiers for GPT-6.2 Astra, GPT-6.2 Astra Law and GPT-6.2 Sol. It also adds 2-hour and 8-hour Responses background windows.
OpenAI releases GPT-Live-1.1 in the API, with standby mode billing idle WebRTC time at $0.02 per minute
OpenAI has released GPT-Live-1.1 in the API. The full-duplex voice model improves multilingual accents and transcription, and a new standby mode bills idle WebRTC time at $0.02 per minute instead of $0.05.
GPT-6.2 Astra scores 41.2% on XDOF's Ember-0 robot benchmark, up from 19.6% for GPT-6 Astra
Robotics infrastructure company XDOF reports that GPT-6.2 Astra reached a 41.2% average success rate across the 16 tasks of its Ember-0 benchmark, versus 19.6% for GPT-6 Astra. It is the first model to complete rope knot-tying and T-shirt unfolding on the benchmark. Average prediction latency fell from 15 seconds to 10.1 seconds.
OpenAI says GPT-6.2 Astra safeguards cut reasoning-extraction attack success to 0.2%
OpenAI announced on November 20, 2026 stronger defenses against adversarial distillation. It reports that adaptive attacks extracting protected reasoning from an earlier request succeed 0.2% of the time on GPT-6.2 Astra with updated safeguards, versus 54.0% on GPT-6 Astra with original safeguards.
Comments
Loading...