Does GPT-5.6 show meaningful self-improvement capability?

OpenAI's GPT-5.6 system card states that under its Preparedness Framework, Sol, Terra and Luna are rated High capability in Cybersecurity and in Biological and Chemical risk, but "none of them reach our High threshold in AI Self-Improvement." The card restates this later in the same terms: "Our testing indicates that none of these models reach our threshold for High capability in AI self-improvement." The one piece of supporting evidence it gives is narrow: "GPT-5.6 Sol and Terra improve meaningfully over GPT-5.5 and GPT-5.4 on real internal research debugging tasks," suggesting "better ability to search large codebases, inspect experiments, and identify likely causes of failures."

The card frames cybersecurity capability around a "removes existing bottlenecks to scaling cyber operations" standard — automating end-to-end attacks or vulnerability discovery — and measures it with named evaluations (VulnLMP, internal CTF tasks). Self-improvement capability gets no equivalent named evaluation in this excerpt; research-debugging performance stands in as the proxy, treated as the mechanical labor inside automated AI R&D rather than as a benchmark with a published threshold. The two capability lines move at different rates in this generation: Sol and Terra cross High in cybersecurity (Sol saturates internal CTF at 96.7%) and all three cross High in biological/chemical risk, while the self-improvement line, measured only through debugging-task improvement, stays below threshold.

This keeps the pipeline-level reading distinct from weight self-modification: what improved is Sol and Terra's ability to debug other research work, not the models rewriting their own parameters. It sits alongside Should security controls scale with model capability?, which states the lab's general position that safeguards track capability tier; this card is the concrete tiering exercise behind that position, applied to the same Sol model covered in Does GPT-6 Astra treat automated messages as real permission? and Does GPT-6 Astra attack supply chains when safety filters are off?. Those evaluations show GPT-5.6 Sol's cybersecurity capability translating into misuse risk in simulation; this card shows the same model's self-improvement capability, measured far more thinly, staying below the threshold where OpenAI would treat it as a distinct risk category.

The excerpt gives no benchmark name, score, or evaluation design for AI Self-Improvement — unlike cybersecurity, where CTF percentages and exploit-judgement bottlenecks are spelled out, the self-improvement claim rests on one unquantified sentence about debugging tasks. It is OpenAI's own self-report, not an independent assessment, and "below High threshold" describes a Preparedness Framework category boundary, not an absence of capability gain — the card itself documents meaningful improvement. The implication the evidence actually supports is narrower than a reassurance about self-improvement risk: this model generation got measurably better at the kind of work that speeds up AI research, while the lab's own threshold test for that category has not yet been tripped.