Opus 5.5 returns HTTP 400 on ALPHA FORGE's forced tool spec

The episode is a case study in how quickly a model upgrade can turn into an outage. ALPHA FORGE runs two AI workers: a Claude-based judge that decides trades and a Gemini-based shadow that only solves the same problems for comparison. When the developer asked Claude Code to pull every model name supported by each provider's official API, the list filled with brand-new releases, and selecting Opus 5.5 seemed like a straightforward improvement. The forced tool specification dates back to an earlier installment, when the design required three prices to be stated every time; Opus 5.5 and Fable 5.1 reject that constraint outright and return HTTP 400, while Opus 5 and Sonnet 5 still comply. Because the app's logging recorded only the status code and not the response body, the Saturday trial was the only thing standing between the setup and a Monday morning pick of zero results accompanied by unexplained errors.

The developer chose the AI's second option rather than its recommended first one, deliberately avoiding a weekend change to the core judgment logic in favor of a configuration swap that worked immediately. That swap made the trade-off visible: running on Opus 5 took 174 seconds per pick and raised input tokens 12% and output tokens 44%, with estimated cost about 6.4 times higher than Sonnet 5. The developer judged that the accuracy gain had not been measured against that gap, and reverted both the settings and the model-list feature with a revert commit.

Separately, the comparison surfaced that the Gemini shadow had not succeeded since Sept 13. It had been returning HTTP 402 with the message "Your prepayment credits are depleted." Because the shadow does not affect production judgments, the screen kept working normally and nobody noticed for 13 days. That one could not be fixed in code; the developer topped up on Google's side and the response returned to 200 the same day.