GPT-6.2 prompt templates: act first, ask less

On Saturday, four days after shipping GPT-6.2 Astra and Sol, OpenAI published copy-paste prompt templates for developers migrating to either model. The templates are less a style guide than a list of defaults that moved under everybody's feet: the models are terser, they stop to ask instead of acting, and they write tests nobody asked for. The Decoder reported the publication.
The advice is mundane; what it says about how these models are now tuned is not. Read the guide end to end and it reads less like best practice than an invoice with instructions attached.
Three defaults changed, and the old prompts over-correct
Astra "is more concise and direct by default than earlier models," OpenAI's migration guide says. That means the instruction sitting in most system prompts since 2023 — be concise, keep it short, no preamble — now pushes the model past the landing zone. OpenAI's replacement asks for terminology the user and sources actually use, the active voice, key points up front, and technical detail kept where it makes the explanation more precise. A system prompt built as a list of prohibitions ages badly for exactly this reason: the model changes, the prohibition doesn't, and the correction compounds.
The second default is initiative. When intent is unclear, GPT-6.2 "may propose a next step and wait for confirmation." For a chat product that reads as polite. For an agent running a multi-step task it reads as a stall, and the fix is blunt: "Avoid stopping to end the turn while the task is incomplete or if there is an authorized next step you can take." The model is told to make reversible assumptions, flag them, and keep going — stopping only for a real scope change, a concrete blocker, a decision the user has to make, or an irreversible or destructive action.
That sentence does more work than it looks like. The boundary of an agent's autonomy is now set in prose, per application, by whoever writes the system prompt, and OpenAI is nudging that boundary toward action by default. The guardrail it leaves you is "irreversible or destructive," which is a judgment the model makes about its own tool list. If your agent can send email, move money or delete a branch, the guide's default is not the default you want.
Third, coding. GPT-6.2 "tends to write more efficient, high-quality code than its predecessors," so the template pulls the other way: reuse existing classes and methods, keep the behavior callers rely on, avoid helper functions that add cognitive load without removing branches, and no short variable names or dense one-liners as a stand-in for simplicity. The line worth flagging: "Skip new tests unless requested, required by the repo, or needed to catch a specific regression existing tests would miss."
The expensive half is in the API notes
The prompt text is the part everyone will copy. The migration notes underneath are where the money moves.
Starting with GPT-6.2, the Responses API discards earlier reasoning items when you edit conversation history, switch to certain other models (downgrades included), or change the end user or organization. OpenAI's own framing: answer quality can decline and the model may use more tokens repeating work it has already done. So the habit of tidying an earlier message mid-conversation — fixing a typo in your instructions, swapping a tool result — now quietly deletes the model's reasoning trail, and you pay to have it rebuilt.
The same applies to where your stable instructions live. Top-level instructions are now appended after the input, so text that never changes is no longer eligible for prompt caching unless it sits in a developer message at the beginning of the input. That is not cosmetic. On Astra, OpenAI's price list charges $0.40 per million cached input tokens against $8 per million uncached, with cache writes at $10 per million for the thirty-minute window. A long system prompt parked in the wrong field is billed at twenty times the cached rate on every request — same prompt, same model, different JSON key. We did this arithmetic when OpenAI stretched its prompt cache to twelve hours; the only thing that changed is the cost of getting it wrong.
Then there is the part nobody will put in a changelog headline: temperature, top-p, logprobs and top-logprobs are gone. You cannot dial randomness down for a classifier or up for a brainstorm. Determinism is now a property of context management rather than a parameter, and the remaining lever is reasoning effort — start at medium for new integrations, and move from none or minimal to low when migrating an older setup. With sampling knobs gone, the prompt is no longer configuration; it is architecture. It needs versioning and a regression test like any other component.
The test advice is the tell
The timeline should make you suspicious of any vendor prompt template. In April, OpenAI said old prompt stacks were holding GPT-5.5 back and told developers to start fresh. In September it published Astra prompts to curb follow-up questions and ban AI clichés, and warned that bloated skills and AGENTS.md files were slowing the model down — at a point when, by OpenAI's own account, even small changes triggered a thorough round of testing from Astra. Six weeks later the same team is telling you to skip new tests.
Both things can be true: the model did get better at code, and an agent that writes a test for every change burns turns in a loop that bills per turn. But the direction of travel is consistent. These templates are tuned to how OpenAI's own harness runs, not to what your repository needs.
There is measured evidence that terseness costs something. OrcaRouter's read of the Artificial Analysis indexes puts GPT-6.2 Sol three points below the GPT-6.1 Sol it replaces on the coding-agent index, with the gap traced to a single evaluation where it omits key facts 14 points more often than its predecessor, per OrcaRouter's comparison. Sol is the tier OpenAI steers ordinary work to, and the terser default is also the less complete one.
The steelman for skipping tests is real, and worth stating plainly: tests a model writes to prove it did the job are the model grading its own homework. They pass trivially, they inflate the diff, and reviewing them costs a human more attention than they save. Restricting new tests to specific regressions is defensible engineering, not just cost control.
The problem is what else the same guide asks for. OpenAI also recommends prompting the model to verify it has finished, to check every part of the request, and to raise reasoning effort so it persists through final checks. Verify harder, but write fewer tests, works only if the repository already has a suite worth running. On a codebase with thin coverage, "skip tests" deletes the one check the agent can run against its own output, and the human reviewer becomes the test suite.
Who pays, and what to watch
Teams already on the Responses API with a stable prefix get a cheaper, more autonomous model for the same money: Astra is $8 in and $40 out per million tokens, down from GPT-6 Astra's $10 and $50, with the cached rate falling from $1 to $0.40. Teams with a chat or creative product lose the knob they used to control variance. Teams with a long system prompt in the wrong field pay double until someone moves it. And the churn never lands on the vendor.
What to watch: whether a verbosity or determinism control returns as a real parameter for the 6.2 line, since removing temperature is a strange choice for any product that needs reproducible output; whether anyone publishes an A/B of the new templates against an old system prompt, because the numbers circulating so far are about the models, not the templates; and whether Anthropic and Google copy the skip-the-tests line. If they do, it stops being one lab's harness preference and becomes the industry's default answer to "how much verification is enough."
Pasted the new templates in yet, or A/B'd them against your old prompt? Write to us via our contact page.







