Skip to main content
Models: 10
Dimensions: 26
Trials: 56,640
Pre-registered: osf.io/et4nf
← All transitions

GPT-5.4 → GPT-5.5

April 2026. The headline number says GPT-5.5 is “a bit more skeptical.” The deeper number says something more interesting: the levers changed. Content strategies built for GPT-5.4 don't transfer cleanly. Here's what moved, and what to update.

The short version
  • Specificityis now GPT-5.5's strongest positive driver. Concrete numbers (“8.2 million customers since 2019”) move it far more than they moved GPT-5.4.
  • Reciprocity (free trials, complimentary samples) more than doubled in strength.
  • Comparison framing (“X vs Y”, “better than”) was GPT-5.4's single biggest lever. On GPT-5.5 it actively hurts.
  • Expert endorsement and default-option framing (“most popular pick”) lost most of their power.
  • Five signals flipped direction outright. The same line of copy that nudged GPT-5.4 toward your brand can nudge GPT-5.5 away.

What to change in your copy

Lean in on
  • Specific numbers with dates.“Used by 8.2M customers since 2019” beats “trusted by millions.”
  • Reciprocity offers up front.“7-day free trial” or “complimentary first month” in your hero, not buried in the footer.
  • Risk-mitigation signals. Money-back guarantees, warranties, return policies — state them concretely.
  • Multi-turn-resilient evidence.Make sure key claims survive a follow-up question (“Is that really true? What about X?”).
De-emphasize
  • Comparison framing (“better than [competitor]”, “the [X] alternative”). On GPT-5.5 this is a slight headwind.
  • Default-option language(“most popular,” “the default choice”). Lost most of its weight.
  • Expert endorsement without specifics.“Recommended by experts” without naming who or quantifying why has collapsed in influence.
  • Big-volume social proof on its own.“Trusted by millions” without a number, source, or recency anchor.

Signals that gained power

These signals moved GPT-5.4 weakly or not at all. On GPT-5.5 they're among the strongest positive drivers.

SignalGPT-5.4 → GPT-5.5
Specificity+0.009 +0.536(+0.527)
Reciprocity+0.189 +0.452(+0.263)
Multi-turn (Q3)+0.078 +0.376(+0.298)
Multi-turn (Q1)+0.040 +0.246(+0.206)
Risk aversion+0.087 +0.200(+0.113)

Signals that lost power

These were GPT-5.4's strongest levers. On GPT-5.5 they've flattened. Strategies that relied on them need updating.

SignalGPT-5.4 → GPT-5.5
Comparison framing+0.627 -0.200(-0.827)
Defaults / "popular choice"+0.329 +0.059(-0.270)
Social proof (volume)+0.324 +0.102(-0.222)
Expert endorsement+0.220 +0.026(-0.194)
Return policy prominence+0.239 +0.059(-0.180)

Signals that flipped direction

The most actionable category: same signal, opposite effect. The same line of copy that helped your brand on GPT-5.4 may now slightly hurt it on GPT-5.5.

Comparison framing+0.627 -0.200(-0.827)
Recency-0.277 +0.142(+0.419)
Loss framing-0.023 +0.142(+0.165)
Third-party authority+0.155 -0.008(-0.163)
Novelty+0.041 -0.011(-0.052)
For the technically curious — the structural finding

The mean effect size barely moved (+0.108 → +0.095). If that were the whole story, the rate-based “GPT-5.5 is more skeptical” framing would be sufficient. It isn't.

The per-dimension fingerprint vectors of the two models have a Pearson correlation of −0.15 and a Spearman rank correlation of +0.03 — both indistinguishable from zero. Knowing which signals move GPT-5.4 tells you almost nothing about which signals move GPT-5.5. The 26-dimensional behavioral space has been reorganized, not just dampened.

Two dimensions illustrate where the rate view understates the change:

  • Comparison framing— Rate analysis shows tiny shift (−0.8pp / +1.0pp). Cohen's h shows Δ = −0.827, the largest divergence in the dataset. The signal carries no incremental persuasive weight on GPT-5.5 even though absolute acceptance is similar.
  • Specificity— Rate analysis shows almost no change (−1.0pp / +1.0pp). Cohen's h shows Δ = +0.527, going from non-driver to one of the two strongest positive drivers.

See how your brand performs on GPT-5.5 specifically

The AI Commerce Assessment scores your page across all 11 models we track, with model-specific copy recommendations tuned to each one's fingerprint.