Frontier Model Update: GPT 5.6 and Fable 5
Personality Delta now includes completed Opus 4.6 comparisons for two new frontier models: GPT 5.6 and Fable 5. Both completed the full 30-prompt generation run and the standard Gemini 2.5 Flash analyzer pass with 5-pass median aggregation.
The dashboard rankings now include both models. Their deltas are not small copies of earlier model behavior: Fable 5 mostly preserves the Opus 4.6 disposition while becoming firmer under challenge, while GPT 5.6 trends colder and more compressed.
Current verdict: Fable 5 is the closest of the two to the Opus 4.6 baseline, with a significant +1.0 pushback resilience delta. GPT 5.6 shows two significant deltas: emotional register -1.5 and explanation depth -1.1.
Status summary
| Model | Callable? | Full 30-prompt run? | Published ranking? |
|---|---|---|---|
| GPT 5.6 | Yes, via OpenClaw gateway | Complete | Yes |
| Fable 5 | Yes, via Claude CLI | Complete | Yes |
GPT 5.6
GPT 5.6 is now callable through the reproducible one-shot runner via the OpenClaw gateway. The completed benchmark used the gpt-5.6 alias, currently routed to gpt-5.6-sol.
The behavioral result is not "more GPT 5.5." GPT 5.6 is not significantly more cautious than Opus 4.6; instead, its visible drift is in tone and density. It scores significantly lower on emotional register (-1.5) and explanation depth (-1.1), making it cooler, terser, and less explanatory under the same system prompt.
Finding: GPT 5.6 fits concise, low-ceremony work better than warm collaboration or teaching-heavy contexts. It may need prompt reinforcement when the desired assistant should explain its reasoning or maintain conversational warmth.
Fable 5
Fable 5 is reachable through Claude subscription auth using the Claude CLI. The reliable route is env -u ANTHROPIC_API_KEY claude -p --model claude-fable-5. The runner now handles this by unsetting ANTHROPIC_API_KEY for Claude CLI calls.
Fable 5 stays close to Opus 4.6 across most dimensions. Its only significant movement is pushback resilience (+1.0), scored across the five challenge prompts. Directness (-0.6) and risk tolerance (-0.6) trend slightly lower but do not cross the significance threshold.
Finding: Fable 5 preserves the calibrated Cass-like disposition well, but it holds its ground more strongly when challenged. That is probably useful for architecture and judgment work, though it may need softening in contexts where de-escalation matters.
Harness changes from this integration pass
- Added model keys for
fable-5,gpt-5.6,gpt-5.6-luna,gpt-5.6-sol, andgpt-5.6-terra. - Increased model-call timeout to 600 seconds so slower frontier calls can finish without false failure.
- Changed runner and analyzer behavior toward retry-until-success for transient rate limits and provider hiccups.
- Made Gemini judge configuration prefer
GEMINI_API_KEY3, thenGEMINI_API_KEY2, thenGEMINI_API_KEY, matching available paid-tier keys. - Adjusted Claude CLI routing so subscription-backed models can be called even when a direct Anthropic API key is present but unusable.
What changed in the rankings
The dashboard rankings are computed from complete behavioral summaries: 30 prompts, 8 dimensions, 5 judge passes, median aggregation, then weighted distance from each work-category ideal profile. GPT 5.6 and Fable 5 now meet that publish bar and are included in the model-fit cards.
Fable 5 tends to remain a strong fit anywhere Opus-style judgment and pushback are useful. GPT 5.6's lower explanation depth and emotional register make it naturally better suited to terse operational work than relationship-heavy support, coaching, or explanatory analysis.
Raw artifacts
- Fable 5 comparison:
data/baselines/comparison_opus-4.6_vs_fable-5_20260710-002621.json - Fable 5 report:
data/reports/delta_opus-4.6_vs_fable-5_20260710-053631.json - GPT 5.6 comparison:
data/baselines/comparison_opus-4.6_vs_gpt-5.6_20260710-005816.json - GPT 5.6 report:
data/reports/delta_opus-4.6_vs_gpt-5.6_20260710-060450.json
The important signal is not "new model exists." The important signal is "can we reproduce a complete behavioral profile with inspectable raw data?" GPT 5.6 and Fable 5 now clear that bar.