Loading...

OpenAI just rolled out new prompt templates and API tweaks for its latest GPT‑6.2 models—Astra for heavyweight tasks and Sol for cheaper, complex work. The biggest headline? Astra now starts at $8 per million input tokens and $40 per million output tokens, and OpenAI is urging developers to scrap old prompt stacks and adopt the new, outcome‑focused copies.
According to The Decoder, OpenAI published copy‑ready system prompts for GPT‑6.2 Astra and Sol that address three pain points: writing style, initiative, and coding. Astra is described as “more brief and direct” than its predecessors, so prompts that tell it to “be concise” are now redundant. The recommended everyday‑work prompt tells the model to use the user’s terminology, write in active voice, and front‑load key points. For external audiences, it should add context the reader lacks; for editing, it should preserve the author’s tone.
On longer tasks the model can unintentionally carry shorthand from its own reasoning into the final answer. OpenAI’s new template explicitly instructs the model to make the final response self‑contained, so users never have to dig through collapsed intermediate messages. When the model isn’t sure what to do next, it may suggest a step and wait for confirmation. The guidance now tells developers to give it permission to take that step automatically, unless the action is a major scope change, a blocker, a decision the user must make, or a destructive operation.
The coding prompt pushes for reuse of existing repo logic, avoids needless helper functions, and adds new tests only on request or when required by the repo. It also bans “code‑golf” shortcuts that sacrifice readability. For front‑end work, the template prefers realistic images or SVGs only on demand and tells the model to drop decorative numbering and filler text.
On the API side, the Responses API will discard earlier reasoning content if you edit conversation history, switch models, or change the end‑user/organization. That can inflate token usage because the model may repeat work. OpenAI now recommends a “medium” reasoning level for brand‑new integrations and a shift from “none” or “minimal” to “low” for existing setups. Temperature and top_p parameters are no longer supported.
Pricing-wise, Astra’s API starts at $8 / M input tokens and $40 / M output tokens up to 272 k tokens, then jumps to $16 / M and $60 / M respectively. Cache reads cost $0.80 / M tokens. Sol’s rates are lower but not spelled out in the snippet. The Decoder notes that long coding sessions can quickly hit the cap, so OpenAI’s Codex automatically compresses context after 272 k tokens.
If you’re running a lean operation, two things bite the most: token bills and output quality. The new pricing makes it easy to overspend on a single long‑running prompt, especially if you’re still using “none” reasoning and the model keeps re‑sending its own scratch work. Switching to the recommended “low” or “medium” reasoning level can shave tokens without sacrificing usefulness.
Equally important, the default brief style of Astra means many of the “be concise” instructions you’ve been sprinkling in your prompts are now noise. Keeping old prompt stacks not only wastes tokens but also triggers the model’s habit of asking follow‑up questions—something OpenAI tried to curb with the September Astra prompts. Updating to the fresh templates will make the model’s answers more self‑contained, which translates to fewer back‑and‑forth messages and a smoother user experience for your internal tools or customer‑facing bots.
Finally, the coding guidance is a reminder that AI‑generated code can look slick but be a maintenance nightmare. By forcing readability and reuse, the new prompts help you avoid hidden technical debt that small teams can’t afford to pay down later.
Open-source Hermes Agent is coming to Microsoft Teams, allowing workers to package successful AI workflows into team-wide reusable skills.

OpenAI's internal data shows researchers value their AI coding agents so highly they'd rather give up staff, despite the tools frequently producing 'AI slop' and requiring significant human oversight.

Reson8 open-sourced Rumble-2-Flash, a 0.5B parameter ASR model for 40+ languages that outperforms models up to 60x larger, with Apache 2.0 licensing.

Instagram introduces a tool to detect unauthorized video reposts and AI-modified content, allowing creators to block or label infringing posts.

Comments2
Dropping the redundant "be concise" instructions is long overdue, especially for operations teams managing high-volume document pipelines where every extra token adds up. I am more interested in whether Sol can actually maintain that logic-to-cost ratio in multi-step agentic workflows without needing constant human sanity checks on the intermediate steps.
In my benchmark tests with Sol last week, the error rate spiked dramatically by step three of an agent loop, forcing me to throw in extra validator calls just to keep it on the rails. For complex workflows, you are definitely still paying the "babysitting tax" in both API costs and dev hours.
The new pricing for Astra seems steep; we tried a similar model and ended up with a 30% increase in costs, how do you plan to mitigate this?