← Back to feed

OpenAI published an October 2026 deployment safety update for GPT-6 Sol and GPT-6 Luna on October 7, introducing several new evaluation frameworks alongside refreshed scores on existing benchmarks. The most numerically specific result is on SEC-Bench Pro, a benchmark that measures a model’s ability to discover vulnerabilities in large JavaScript engines including V8 and SpiderMonkey.

GPT-6 Sol scores 68.85% on SEC-Bench Pro. GPT-6 Luna reaches 48.50%. GPT-6 Astra, the current frontier model, scored 85.4% on the same benchmark. The 16-point gap between Sol and Astra represents the clearest published signal yet of how much capability Astra adds over Sol in advanced cybersecurity work.

New Benchmarks Introduced

The October card marks the first time OpenAI has applied HealthBench Professional and MentalHealthBench to the GPT-6 family.

HealthBench Professional targets clinical decision-making tasks. OpenAI notes that HealthBench Professional “has been more informative than older HealthBench variants in our recent evaluations,” and explicitly recommends it over the original HealthBench for measuring continued frontier progress. The original HealthBench, now more than a year old, is described as “approaching a noise ceiling for frontier models.” Specific scores for Sol and Luna on HealthBench Professional were published in the system card; for comparison, GPT-5.6 Sol scored 60.5% on the same benchmark at its August update.

MentalHealthBench consists of 1,215 synthetic conversations covering scenarios from daily well-being topics to urgent mental health emergencies. Each task is paired with clinician-authored rubric criteria. OpenAI developed the benchmark with mental health experts to evaluate AI responses in high-stakes conversational settings where the cost of a poor response is direct clinical risk.

The October card also adds an AI Self-Improvement Capabilities section, covering evaluations of whether the models can meaningfully accelerate their own training or circumvent development controls. No scores from that section were included in the publicly available summary.

Biological and Chemical Risk Rating

Both GPT-6 Sol and Luna are treated as High capability in the biological and chemical domain. OpenAI ran GPT-6 Sol through Critical capability thresholds; the results did not cross the indicative thresholds. Separate Critical testing was not required for Luna because it scored below GPT-5.6 Sol on all High capability evaluations.

Biology model refusal evaluations show substantially improved safety on severe and dual-use prompts for both Sol and Luna compared with GPT-5.6 Sol and Luna from the August 2026 card. That improvement covers model responses before production safeguards are applied.

Cybersecurity Safety Comparison

On cybersecurity safety evaluations specifically — as distinct from the SEC-Bench Pro capability measurements — GPT-6 Sol and Luna achieve safety scores broadly comparable to their GPT-5.6 counterparts. OpenAI notes that model refusal is one layer of the safety stack alongside additional production safeguards.

The distinction is meaningful: SEC-Bench Pro measures what the models can do when asked to find vulnerabilities. The cybersecurity safety evaluation measures whether the models refuse to help when they should. Both Sol and Luna show capability below Astra on the former; both remain roughly in line with GPT-5.6 on the latter.

Context Within the GPT-6 Family

GPT-6 Astra launched in September 2026 as the top-tier model with restrictions. Sol and Luna were released at the same time at lower price points — Sol at $2 per million input tokens, Luna at $0.10. The SEC-Bench Pro numbers make explicit what the pricing already implied: Astra is not just more expensive, it is meaningfully more capable on advanced technical tasks where capability and safety risk overlap.

GPT-6.1 Sol, released separately on September 29, 2026, is not covered by this safety card. OpenAI has not published a deployment safety update for GPT-6.1 Sol as of October 8, 2026.