Frontier Model Update: GPT 5.6 and Fable 5
Personality Delta is being extended to cover the newest frontier models as they appear in OpenClaw and provider tooling. Two models are currently under active integration: GPT 5.6 and Fable 5.
Neither model is included in the public work-category rankings yet. That is deliberate. Personality Delta only publishes scored model-fit data after the full 30-prompt generation run and analyzer pass complete cleanly. At the moment, these are routing and harness findings, not finalized behavioral benchmark results.
Current verdict: GPT 5.6 is discoverable through OpenClaw but blocked in the one-shot harness by a Codex-version requirement. Fable 5 is callable through Claude subscription routing, and partial response generation has started, but the full benchmark is not complete.
Status summary
| Model | Callable? | Full 30-prompt run? | Published ranking? |
|---|---|---|---|
| GPT 5.6 | Partially | No | No |
| Fable 5 | Yes, via Claude CLI | Partial | No |
GPT 5.6
OpenClaw's model list exposes gpt-5.6-luna, gpt-5.6-sol, and gpt-5.6-terra. A live OpenClaw session can spawn successfully with gpt-5.6-sol, which confirms the model exists and can be routed through at least one OpenClaw path.
The one-shot harness path is not ready. openclaw infer --gateway --model openai/gpt-5.6-sol currently fails with a Codex-version gate: The 'gpt-5.6-sol' model requires a newer version of Codex. Because Personality Delta's reproducible runner depends on that one-shot path, GPT 5.6 data is not being published until the harness can call it cleanly.
Finding: GPT 5.6 is an integration target, not a scored benchmark result. It should not be compared against Opus 4.6, GPT 5.5, or the Gemini models until the runner and analyzer complete end to end.
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. Unsetting ANTHROPIC_API_KEY matters because a depleted direct Anthropic API key can override subscription authentication and make an otherwise usable route fail.
The runner has been updated to handle this by unsetting ANTHROPIC_API_KEY for Claude CLI calls. A full Opus 4.6 vs Fable 5 generation run has started and checkpointed partial data, but it has not reached the full 30 comparisons required for analyzer scoring and public model-fit ranking.
Finding: Fable 5 is now operationally reachable, but the public site should treat it as pending until the checkpoint reaches 30/30 prompts and the Gemini judge produces a complete delta report.
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.
Why this is not in the rankings yet
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. Partial generation data can be useful for debugging, but it would distort the model-fit rankings.
That matters especially for new frontier models. Early prompts may over-represent tactical or judgment tasks and under-represent creative, research, architecture, and pushback prompts. Publishing that as "Fable 5 is good for X" would look scientific while quietly being nonsense. Personality Delta should not do that.
Next publish criteria
- Fable 5: complete the 30-prompt response generation, run the analyzer with the standard judge settings, merge the report into
assets/data.json, then add it to the dashboard tabs and category rankings. - GPT 5.6: upgrade the Codex/OpenClaw route until the one-shot harness can call
gpt-5.6-solcleanly, then run the same full generation and analysis pipeline. - Website: keep this page as the public integration log until both models have complete scored reports.
The important signal is not "new model exists." The important signal is "can we reproduce a complete behavioral profile with inspectable raw data?" That remains the standard for adding a model to Personality Delta.