How to Prompt GPT-5.5, Claude Opus 4.8, and Claude Fable 5 (2026 Executive Guide)

For GPT-5.5, describe the destination, not the route. Instead of "first summarize, then identify risks, then recommend," say "a one-page briefing for a CFO who needs a clear risk summary and a recommended action." Drop unnecessary absolutes like always and never, and save them for real rules such as safety or compliance. For judgment calls, give a decision rule instead: "If the source doesn't support a claim, flag the gap rather than filling it in." Add a stopping point ("keep this under 300 words") so a capable model doesn't keep going past where you needed it to stop.

For Claude Opus 4.8, say exactly what you mean. State the full scope if you want a fix applied everywhere, not just the one place you pointed to: "apply this to every section, not just the first." Signal when to slow down: adding "this needs multi-step reasoning, so think it through before answering" to your Do's will get you a more considered answer on strategy or analysis work. Name your constraints precisely. Opus sizes its answer to how complex it judges the task, so "be concise" can land shorter than you wanted; a number like "two paragraphs per section" removes the guesswork.

For Claude Fable 5, give it the hard thing, and the reason why. Fable rewards a clear Objective more than any model before it: "I'm preparing this for our board, who need to approve more investment" tells it far more than "make this persuasive." Because Fable runs longer and acts more independently, state your Do's and Don'ts clearly: what to change, what to leave alone, and where to stop and check with you. And because it can work unsupervised for a long stretch, ask it to report what it actually did and what it couldn't verify, rather than a tidy summary that glosses over the gaps.

Across all three, set the tone yourself. All three default to efficient and direct, so for anything client-facing or team-facing, add an instruction like "use a warm, collaborative voice."

AI prompting techniques at a glance

1. GPT-5.5: best for outcome-first delegation

GPT-5.5 is OpenAI's flagship reasoning model, built for complex professional work: coding, multi-step analysis, and document-heavy tasks. It carries a context window of roughly 1 million tokens, and it's the default top model across ChatGPT's paid tiers.

How it works

Give GPT-5.5 the destination, not the steps. Instead of scripting "summarize, then flag risks, then recommend," describe the finished product: who it's for and what they need from it. Pair that with a decision rule for uncertain claims, and a hard stop such as a word count or a list of the top three items to cover.

Where it excels

GPT-5.5 is strong at choosing its own reasoning path once the destination is clear, and it handles long documents well within its context window. The tradeoff is that it can sound equally confident whether or not a claim is actually grounded in your source material, so telling it which facts must come from your documents is essential, not optional.

GPT-5.5 pricing

Consumer access through ChatGPT:

• Plus: $20/month, includes GPT-5.5 with standard usage limits
• Pro: $100 to $200/month, adds GPT-5.5 Pro and far higher usage ceilings

API access for developers:

• Standard: $5 per million input tokens, $30 per million output tokens
• Long-context requests above roughly 272,000 tokens are billed at a higher rate

2. Claude Opus 4.8: best for precisely scoped, high-stakes work

Claude Opus 4.8 is Anthropic's premium hybrid reasoning model, positioned for serious coding, agentic workflows, and enterprise document work. It carries a 1 million token context window by default and adjusts how much it "thinks" based on task complexity.

How it works

Opus follows instructions literally, which is a gift once you know how to use it. If you want a correction applied everywhere a similar error appears, say so explicitly: "apply this to every section, not just the one I flagged." For strategic or analytical tasks, tell it directly to reason through the problem before answering. And when you want a specific length, give a number rather than a vague instruction like "keep it short."

Where it excels

Opus 4.8's literal-mindedness makes it more predictable than models that infer intent, which matters for high-stakes or compliance-sensitive work where you don't want the model quietly expanding its own scope. The same trait is its main limitation: it will not generalize a fix on your behalf, so vague requests get narrower results than you might expect.

Claude Opus 4.8 pricing

Consumer access through Claude.ai:

• Pro: $20/month
• Max: $100 or $200/month, for heavier daily usage

API access for developers:

• Standard: $5 per million input tokens, $25 per million output tokens
• Fast mode (research preview): $10 per million input tokens, $50 per million output tokens, for roughly 2.5x faster output

3. Claude Fable 5: best for long, ambiguous, high-stakes projects

Claude Fable 5 is Anthropic's most capable generally available model, built for multi-hour and multi-day autonomous work: large migrations, complex builds, and open-ended projects that used to require a senior contractor.

How it works

Fable rewards a strong Objective more than any prior model. Tell it why you need the output, not just what the output should look like: "I'm preparing this for the board, who need to approve more investment" changes how it approaches the work. Because it can act independently for long stretches, spell out its Do's and Don'ts clearly: what it may change, what it must leave alone, and where it should pause and check in. Ask it to report what it actually did and what it couldn't verify, rather than a polished summary.

Where it excels

Fable is designed for the messiest, most open-ended work: turning scattered notes into a finished project plan, or carrying a build through to completion with minimal hand-holding. Its main limitation is cost and verbosity. At roughly double the token price of Opus 4.8, and without an explicit stopping point, it will happily keep working (and spending) longer than most tasks require.

Claude Fable 5 pricing

Consumer access through Claude.ai:

• Included within the usage allowance on Pro ($20/month) and Max ($100 or $200/month) plans, subject to weekly limits

API access for developers:

• Standard: $10 per million input tokens, $50 per million output tokens
• Prompt caching cuts cached input cost by roughly 90%

What's the actual ROI of prompting this way?

The payoff isn't a faster first draft. It's fewer rounds of correction. A sharp Objective, matched to the failure mode of the specific model in front of you, is what turns a frontier model from a fast typist into an actual collaborator on the outcome you're after.

How does Lead with AI decide which prompting techniques to recommend?

Lead with AI works directly with senior leaders and teams at organizations including Microsoft, Apple, McKinsey, BCG, and Harvard, training them on how to get real leverage from frontier AI models. Recommendations here are revised at least monthly based on what executives in Lead with AI's programs actually use and prefer, alongside proprietary data from the AI Leader Advanced course.

Can I use GPT-5.5, Claude Opus 4.8, or Claude Fable 5 for free?

Free tiers exist for all three, but none include the flagship model discussed here at full strength. ChatGPT's free tier runs GPT-5.3 Instant, not GPT-5.5. Claude's free tier gives limited access to Claude's models with usage caps and does not include Opus 4.8 or Fable 5 at meaningful volume. Both companies reserve their strongest models for paid plans starting around $20/month.

Should I always describe the destination instead of the steps?

It depends on the model and the task. For GPT-5.5 and Claude Fable 5, describing the outcome and letting the model choose its route is usually the stronger approach. For Claude Opus 4.8, the opposite instinct applies for anything requiring precision: state the scope explicitly, because Opus won't infer that a fix should apply beyond what you named.

Can AI replace a leader's own judgment on hard decisions?

No, and none of these models are designed to. All three depend on a human supplying the Objective, the definition of what "good" looks like for this specific situation. That judgment call is exactly what a frontier model cannot make on its own; it can only execute against the destination you give it.

Which model is best if my team is already inside the Microsoft or Google ecosystem?

Ecosystem fit matters more than raw benchmark scores for most day-to-day work. Teams standardized on Microsoft tools will find GPT-5.5 more tightly integrated through Copilot; teams on Google Workspace will lean toward Gemini. Claude's models are the strongest standalone choice when the priority is precision (Opus 4.8) or long autonomous work (Fable 5) rather than deep integration with a specific productivity suite.

What other AI models are worth knowing about?

The models in this article are the ones executives in Lead with AI's programs consistently tell us they get the most value from. But the market is broader, and a few alternatives are worth knowing about if your needs are more specialized.

• Gemini 3.1 Pro (Google): Strong on long-context and multimodal reasoning, with a 1 to 2 million token context window. Starting at $19.99/month for Google AI Pro.
• Grok 4.3 (xAI): Notable for real-time X data access and aggressive API pricing. Consumer access bundled into X Premium+ starting around $16/month.
• Claude Sonnet 5 (Anthropic): A faster, lower-cost alternative to Opus 4.8 for everyday tasks that don't need the top-end reasoning tier. Included on Claude's free and paid plans.

• Lead with the Objective, not the process. The clearer you are about the destination and the reason behind it, the less you need to script the steps, especially with GPT-5.5 and Claude Fable 5.
• State scope explicitly with Claude Opus 4.8. It won't extend a fix beyond what you named, so say "everywhere this appears" if that's what you mean.
• Give Claude Fable 5 boundaries, not just tasks. Because it works independently for hours, your Do's and Don'ts are what keep it inside the lines.
• Set the tone yourself. All three models default to efficient and direct; add a line about voice for anything client- or team-facing.
• Match the model to the failure mode, not just the benchmark score. A sharper prompt beats a "smarter" model almost every time.

Want to build this into a repeatable skill across your team? AI Leader Advanced covers the full CODO framework and how to apply it across frontier models. See also our guide to AI fluency.