ChatGPT vs Claude vs Gemini vs Grok: Should You Optimise Prompts Differently?

Every adapter runs after your chosen style (Balanced/Aggressive/Caveman/Structured) has already compressed the prompt — it never re-does that work, it only adjusts structure for the target model:

• Claude — wraps quoted or referenced source content in<content> XML tags, separating instruction from input. This is a documented Anthropic prompting convention, not a guess.
• ChatGPT (GPT) — converts an inline comma-separated list into a numbered list, and appends an explicit Output format: X label when a non-default format (JSON, table, CSV) is detected but not already labeled.
• Gemini — the same two mechanisms as GPT, in Gemini's own idiom: a --- divider separating instruction from reference content, and a Respond using X. directive.
• Grok — the only subtractive adapter: strips leading meta-phrasing (“so basically”, “in this context,”), reminder framing (“remember to X” → “X”), and reorders a sentence task-first when that helps — every word survives, just resequenced.

Same prompt, Balanced style, with and without the Claude target selected:

Notice the Claude version is longer, not shorter. The adapter's job is structural clarity — separating instruction from input — not token reduction. Token Optimiser accepts this specific kind of change even though it adds tokens, because it's a whitelisted structural improvement rather than uncontrolled growth (see our benchmark methodology for the exact budget this is checked against).

Same token count, structurally different output. This is the typical shape of what these adapters do — reformatting, not compressing.

Token Optimiser's frozen V1 benchmark measures, for each model target, how often its output structurally differs from the Generic target on the same prompt and style. As of the current, post-recalibration measurement:

Read that plainly: for the large majority of prompts in this corpus, selecting a specific model target produces the same output as Generic. These numbers improved from an earlier measurement of near-zero (0/208) for Claude, ChatGPT, and Gemini after a growth-safety gate was recalibrated — but they are still modest, and we're publishing them as measured, not as a headline percentage picked to sound better.

One honest caveat from our own engineering notes: this corpus's fixtures don't often combine strippable filler with an adapter-trigger shape in the same prompt, which likely understates how often these adapters would fire on more varied real-world prompts. That's a plausible explanation, not a claim we've separately proven — so we're stating the measured number above as the number that matters, and the explanation as a caveat, not the other way around.

If you're on a plan that includes model targets, there's no real downside to selecting the model you're actually sending the prompt to — the adapters are conservative and only fire when their specific trigger condition is met, never randomly. But don't expect it to be your main source of token savings. That comes from style selection and cutting the filler covered in our guide to reducing prompt tokens. Model-specific formatting is a smaller, structural refinement on top of that — worth having, not worth overselling.

Full methodology and the complete result breakdown, including the parts that don't sound like a sales pitch, are on our V1 benchmark page.