Anthropic's Withheld Model Proves the Gap Between Open and Closed AI Is Accelerating, Not Closing
TexTak places open-source models achieving frontier parity at 69%, but Anthropic's decision to withhold Claude Mythos Preview fundamentally challenges that thesis. When a frontier lab discovers capabilities so advanced they won't release them publicly, the performance gap isn't closing — it's being deliberately maintained through strategic withholding.
Our 69% probability has been driven by three concrete factors: Meta's massive open-source investment, documented training technique convergence, and the verified 100x reduction in compute costs that democratizes model development. The benchmark data supports this trajectory — GLM-5.1's 95.3% AIME performance puts it within striking distance of GPT-5.4's 98.7%, a gap that seemed insurmountable just months ago.
But the Mythos revelation exposes a critical flaw in our reasoning. We've been treating benchmark convergence as evidence of capability parity when it may actually represent strategic disclosure parity. Anthropic didn't withhold Mythos because it performs poorly on AIME — they withheld it because it found "thousands of zero-day vulnerabilities" in major operating systems. This suggests frontier labs possess unreleased capabilities that never appear in public benchmarks, making our parity thesis potentially unmeasurable rather than just difficult to assess.
The strongest counterargument to our position isn't technical limitations or compute constraints — it's that safety considerations create systematic publication gaps. If every frontier lab adopts Anthropic's withholding approach for genuinely dangerous capabilities, open-source models might match released benchmarks while remaining generations behind actual frontier performance. The Mythos precedent suggests this pattern is already emerging.
What keeps us from dropping below 50% is that withholding strategies face their own constraints. Companies need revenue, and suppressing capabilities indefinitely conflicts with commercial imperatives. But if three or more major labs implement systematic capability withholding by Q3 2026, we'd reassess whether "frontier performance" can be meaningfully defined in an environment where the actual frontier remains classified.