GPT-5.6 spikes team token usage 5x

The 5x token usage spike shows that UX and conversational flow, rather than raw code generation accuracy, are the real bottlenecks to developer adoption of AI models. If a model feels like a seamless partner that doesn't miss details, developers will interact with it far more frequently.

• –**Interactive over static:** The value of LLMs is shifting from "one-shot code generation" to "collaborative exploration," where the developer remains in the loop to guide and prune the code.
• –**Defensive coding persists:** Like its predecessors, GPT-5.6 still produces overly defensive and verbose code, requiring a cleanup pass to delete up to half of the generated codebase.
• –**UX beats raw benchmarks:** Ironing out micro-usability issues (e.g., following instructions, remembering context, capturing details) does more to drive developer engagement than incremental performance gains on standard coding benchmarks.
• –**Phased rollouts underway:** GPT-5.6 (which includes Sol, Terra, and Luna tiers) is currently in a limited preview under U.S. government oversight, hinting at a new era of highly guarded model releases.