Introducing GPT-5.5 with Box

The 19% performance figure comes from OpenAI's own demonstration with a paying enterprise partner, not an independent evaluation. There is no third-party replication cited, and the benchmark tasks were not disclosed in enough detail to assess whether they reflect real-world financial workloads or a curated showcase.

The timing here is awkward for OpenAI. Just days before this announcement, The Decoder reported that ARC-AGI-3 analysis identified three systematic reasoning errors that persist across GPT-5.5 and other frontier models, with sub-1% performance on tasks humans solve intuitively. A 19% uplift on a proprietary financial benchmark and near-zero performance on abstract reasoning tests can both be true simultaneously, which is exactly why vertical benchmarks selected by the vendor deserve scrutiny. The earlier coverage of GPT-5.5 matching Claude Mythos in cyber attack simulations (The Decoder, May 1) also reminds us that capability gains in one domain do not transfer cleanly across others.

If Box or an independent financial services firm publishes reproducible benchmark methodology within the next 60 days, the 19% claim becomes worth taking seriously. If no third-party validation appears, treat this as a sales demonstration, not a capability milestone.