Prompt Differences by Model: GPT-6, Claude, Gemini, Grok (Reference Table)

TL;DR: Prompt differences by model now come from vendor defaults more than syntax. As of September 16, 2026, GPT-6 Astra asks more questions, Claude Opus 5 writes longer and self-verifies, Gemini 3.8 Flash wants short prompts with instructions last, and Grok 4.6 always reasons. None of the four want you tuning temperature.

Why do prompts behave differently from model to model?

Prompts behave differently because each vendor trains toward different defaults, then documents workarounds for them. The fundamentals are shared: a clear task, relevant context, constraints and an output format work everywhere. What diverges is everything around them, such as how long the answer runs, how much the model checks its own work, whether it asks before acting, and which request parameters it will even accept.

The useful part is that vendors now publish these quirks. OpenAI, Anthropic and Google each have a current-model prompting guide with specific behaviours and fix-it prompts, and xAI documents its reasoning parameters in detail. This page puts what they say side by side, with the date we read it.

It's a companion to three narrower references on this site: the model-specific formatting cheat sheet for markup and message structure, the temperature settings reference table for sampling, and the video duration parameters by model table for video tools.

The reference table: prompt differences by model

Every cell below comes from the vendor's own documentation, read on September 16, 2026. Where a vendor publishes nothing on a row, the cell says so.

Two cells deserve a note. Anthropic's context figure is from its models overview, which lists 1M tokens and 128K max output for Opus 5. And Google's model page gives 65,536 output tokens while its latest-model guide rounds that to "64k max output tokens"; both describe the same limit.

How should you prompt GPT-6 Astra?

Prompt GPT-6 Astra with explicit permission to act, a stated writing style, and no sampling parameters. OpenAI's latest-model guide says Astra is "more likely to ask for clarification where earlier models would make assumptions." That is good in a chat product and irritating in an automated one.

OpenAI's fixes are published prompts. The shortest version of its autonomy instruction reads:

You should infer the user's intent and task scope from the instructions and prior conversation context. Your job is to bias towards action and carry the user's intended task to completion.

On style, OpenAI says Astra "tends to use lists, tables and Markdown to make responses scannable." If you want prose, say so. OpenAI's own sample begins "Default to using clear, concise paragraphs, each developing one main idea."

Three more Astra details from the same guide:

• Audit your instruction files. OpenAI says Astra "can be more sensitive to instructions contained in skills and other files, such as AGENTS.md," and strongly recommends auditing them. Our explainer on the AGENTS.md standard covers what belongs in one.
• Remove sampling parameters. The migration list says to remove temperature, top_p and top_logprobs.
• Use Responses for tools. Astra supports Chat Completions, but "tool calling requires Responses."

How should you prompt Claude Opus 5?

Prompt Claude Opus 5 with an explicit length instruction and fewer verification instructions than you used before. Anthropic's Opus 5 prompting page says its "default user-facing responses run longer than prior Opus models'", and that lowering effort doesn't reliably shorten them. Its suggested fix is a plain request for brevity.

The bigger change is what to remove. Anthropic says Opus 5 "verifies its own work without being told to", and that explicit verification steps cause over-verification. If your Claude prompts still say "double-check your answer", delete the line and compare costs.

Other Opus 5 specifics from Anthropic's pages:

• Scope can drift outward. Anthropic says Opus 5 "can also expand the scope of a task," and publishes a prompt that begins "Deliver what was asked, at the scope intended."
• It delegates readily. Opus 5 "delegates to subagents more readily than prior models", the opposite of Astra's documented tendency.
• Prefill is gone. Anthropic's best-practices page says prefilled assistant messages on the last turn "return a 400 error" on Claude 4.6 models and later.
• Temperature is locked. Anthropic's Messages API reference says models released after Opus 4.6 accept only 1.0.

Anthropic also suggests starting at the default effort, high, and using low and medium wherever quality holds. For the thinking side of this, see prompting Claude's extended thinking, and for templates adapted to Claude, Claude prompt templates vs ChatGPT.

How should you prompt Gemini 3.8 Flash?

Prompt Gemini 3.8 Flash briefly, put the question after the data, and leave sampling alone. Google's Gemini 3.x guidance says Gemini 3.x "responds best to direct, clear instructions" and warns that verbose prompt engineering "may cause the model to over-analyze."

Google is the most insistent vendor about sampling. For all Gemini 3.x models it says temperature, top_p and top_k are "no longer recommended", and it tells you to "Remove these parameters from all requests." For determinism, it suggests explicit rules in a system instruction instead.

Reasoning is set with thinking_level. On 3.8 Flash, Google's latest-model page lists medium as the default and the model page states that minimal "is not supported and returns an error." That is a change from earlier 3.x Flash models, so a config copied from an older project can fail on the first call.

Google also says 3.8 Flash "can use more tokens on longer running and complex tasks, by design", because it takes smaller reasoning steps and verifies as it goes. If that's more than you need, lower the thinking level. For tool overuse, Google suggests a system instruction like this:

You have a limited action budget of <n> tool calls. Use them efficiently.

Our comparison of Gemini vs ChatGPT prompting differences goes deeper on system instructions and structured output.

How should you prompt Grok 4.6?

Prompt Grok 4.6 as a reasoning model that never switches reasoning off, and enable search when the task needs current information. xAI's reasoning page says reasoning_effort "defaults to "high"" and that "Reasoning cannot be disabled."

That has two practical consequences:

• Some parameters are refused. xAI says presencePenalty, frequencyPenalty and stop "cannot be used with reasoning models" and that requests including them return an error. Strip them before porting an older prompt config.
• Latency is part of the deal. If you need speed, lower effort to low, which xAI describes as suited to latency-sensitive agentic use and simple tool calling.

Grok also has no live data by default. xAI's models page says Grok "has no knowledge of current events or data beyond what was present in its training data" unless you enable its Web Search or X Search tools. A prompt asking "what happened today" won't work without them.

xAI publishes little general prompting advice. The clearest list we found sits in its Microsoft Foundry integration page, under "Prompting Grok effectively": "Encourage step-by-step reasoning when needed", along with specifying the output format and using clear tool schemas. That step-by-step advice runs against what Google and Anthropic now say for their reasoning models, so test it rather than assume it. The Grok 4.6 page also strongly recommends setting a prompt_cache_key so repeated requests hit the cache.

For Grok's consumer app modes, see Grok's modes explained.

Which reasoning setting should you start with on each model?

Start at each vendor's documented default, then move down for speed or up for hard problems. The catch is that the defaults, and even the names, are different on all four models, so "medium" doesn't mean the same thing everywhere.

Here is a starting map, based on each vendor's own description of its settings. Where a cell is our judgement rather than a vendor recommendation, test it:

A few notes on where those cells come from. OpenAI's reasoning guide describes medium as the "Default configuration for most workloads" in its general effort table, but says defaults are model-dependent, and the GPT-6 Astra model page doesn't state one. It also advises using xhigh only "when your evals show a clear benefit that justifies the extra latency and cost."

Anthropic's Opus 5 page says to start at the default high and use low and medium liberally wherever quality holds, stepping up to xhigh "for demanding coding and agentic work". Google's latest-model page recommends its default medium for complex code and agentic use cases. xAI describes low as suited to "Latency-sensitive agentic use and simple tool calling."

Effort is also a cost lever, not just a quality one. Higher settings generally spend more reasoning tokens, so a prompt that works at a lower setting is cheaper to run at scale.

Where do vendors give opposite advice?

On four points, the vendors' own guidance points in different directions, and a single "universal prompt" will be wrong for at least one of them:

The placement question is closer than it looks. Anthropic and Google both say long material goes first and the question last. OpenAI's prompt engineering guide lists context as the last section of a developer message and says it "is usually best positioned near the end of your prompt". All three put the data before the thing you want done with it.

A portable prompt with per-model adapters

Keep one base prompt and a short adapter block per model. The base carries the task; the adapter corrects for that model's documented defaults.

<task>
{{what you want done, in one or two sentences}}
</task>

<context>
{{documents, data, code or background, placed before the request}}
</context>

<requirements>
- Audience: {{who reads the result}}
- Format: {{exact output shape}}
- Constraints: {{must / must not}}
</requirements>

<request>
Based on the context above, {{restate the task as a direct instruction}}.
</request>

{{adapter for the target model}}

Adapters, each written to counter a default the vendor documents:

# GPT-6 Astra
Proceed without asking unless an answer would materially change the result.
Use short paragraphs; use lists only for genuinely parallel items.

# Claude Opus 5
Keep the response focused and brief.
Deliver what was asked, at the scope intended; do not add unrequested steps.

# Gemini 3.8 Flash
(Keep the base prompt short. Put no sampling parameters in the request.)
Answer directly. If you use tools, use no more than {{n}} calls.

# Grok 4.6
(Enable Web Search or X Search if the task needs current information.)
State the output format explicitly. Show your reasoning steps only if I ask.

The adapter lines are ours, written from each vendor's documented behaviour, not copied from their guides. Treat them as starting points and test on your own tasks.

Where does Prompt Architects fit?

This is the job our product does. Prompt Architects rewrites a plain request into a structured prompt, and its multi-model optimization tailors the result for ChatGPT, Gemini or Claude, or picks automatically. The adapters above are the kind of thing it adds, so you don't have to remember which model wants what.

We don't host or run models, and we don't replace reading a vendor's migration notes when you move production traffic. For the day-to-day case, writing one prompt and using it in several chat tools, a saved prompt with per-model variants is faster than rewriting from memory. The free plan includes 5 prompt enhancements per day, according to our FAQ page.