Unconfirmed: @superalesha said on 8 November that a free DeepSeek model came close to an OpenAI hacking-focused model in Vals testing, and Automatica has not confirmed the claim.
@superalesha wrote on 8 November1 that "GPT-6 Sol, which OpenAI only offers through its restricted Daybreak program, reached about 81%" in simulated company-network tasks, while "DeepSeek V4.1 Flash with open weights reached about 73%." The post also said, "A DeepSeek run cost under $3, a Sol run about $22" and "On one of the two networks DeepSeek averaged 16.3 objectives to Sol's 15.7."
The same post described DeepSeek V4.1 Flash as available "on Hugging Face for free." Automatica did not verify that hosting claim, and the available evidence includes no primary Vals report, no OpenAI document identifying "GPT-6 Sol" or the "Daybreak" program, and no DeepSeek material setting out the comparison.
That leaves the core test conditions undocumented in the material Automatica reviewed: model versioning, prompts, tool access, run counts, scoring rules, and whether both systems were evaluated under comparable constraints. Without those details, the post establishes only that the claim was made, not that the comparison holds up.
The post also said, "Today Enclave AI published the funny part," before claiming the same DeepSeek model "went 11 for 11" on an Enclave hacking benchmark, that "5 of those wins skipped the bug each challenge was built to test," and that after Enclave "closed every shortcut" the score "dropped to 4 of 11." Automatica earlier reported the same Enclave figures on 8 November in its coverage of that audit, based on Enclave's published findings as summarized there, but that audit is a separate benchmark from the Vals comparison and does not independently confirm the 73% versus 81% result Automatica Press2.
What would confirm the rumor is a Vals benchmark report or equivalent primary material showing the methodology, task set, run counts, tool permissions, and scoring, or a statement from OpenAI, DeepSeek, or Vals backing the comparison. What would undercut it would be primary results showing different scores, different model identities, or non-comparable test setups.
In the evidence Automatica reviewed, OpenAI, DeepSeek and Vals had not publicly provided those materials or commented on the comparison. With only one qualifying social post and no retrieved benchmark documentation, the claim remains unconfirmed.