The Cost Crash: Why GPT-5.6 Changes Budgets, Not Just Benchmarks

OpenAI’s GPT-5.6 Soul delivers Fable-like performance at one-third the cost across Agent’s Last Exam, coding aggregates, and Zapier’s automation benchmark. This isn’t just a scores race—it signals a structural shift where near-frontier models become viable substitutes for top-tier ones in many enterprise tasks, directly impacting per-call budgeting and architecture decisions.

The cost-performance curve is flattening: Meta’s Muse Spark achieves 72% on vibe coding benchmarks versus Soul’s 81% at 35x lower cost, while Grok 4.5 leads on long-horizon SWE tasks. Engineering leaders face a trilemma: pay for the absolute best, optimize for cost, or manage a portfolio of models per task—each with its own operational overhead.

Self-improvement claims are overhyped: Anthropic’s own analysis suggests internal research speed gains of only 20-30%, not the hundredfold implied by output token metrics. Meanwhile, the UK AI Security Institute found universal jailbreaks for Soul within hours, raising enterprise governance concerns. Faster, cheaper models may increase risk surface before productivity benefits fully materialize.