Scaling Verifiable Test-Time Compute in Alchemax-o3 Pro

Abstract & Key Findings

For the past five years, frontier AI progress was driven predominantly by pre-training compute scaling. With Alchemax-o3 Pro, we demonstrate that scaling test-time search verified against deterministic execution environments yields dramatic gains on problems that previously stumped the largest dense models.

Rather than relying on ungrounded chain-of-thought heuristics, o3 Pro decomposes complex specifications into verifiable sub-goals. When synthesising distributed systems code or formal mathematical proofs, the model interacts with an internal verification kernel to prune invalid branches before generating its final output.

Across SWE-bench Verified (89.2%), AIME 2026 (96.7%), and GPQA Diamond (88.4%), we observe a clean log-linear improvement in accuracy as test-time compute budgets scale from 1x to 64x.