Anthropic’s Opus 4.7 launch quietly dropped three major workflow changes that’ll hit you at scale. First, the old “budget_tokens” setting now triggers a 400 error with no warning—thinking budgets are gone unless you switch to the new adaptive mode and explicit “effort” levels (low, medium, high, xhigh, max). Second, the tokenizer eats about 35 percent more tokens for the same text. That means your context budgets, client-side estimators and bills all jump without any extra context window or performance gains. Third, thinking tokens default to “omitted” instead of “summarised,” so you pay for hidden tokens and see blank blocks in the output. Under real-world load, Opus 4.7’s long-context retrieval collapsed—from 78.3 percent accuracy at 1 million tokens in 4.6 down to 32.2 percent in 4.7 on Anthropic’s own MRCR v2 benchmark. Reports are flooding in about hallucinations, ignored preferences and “garbage” answers. Anthropic claims it raised rate limits permanently, but offers no hard numbers. Even a 1–1.35× increase just offsets the inflated token draw, not your actual throughput. Boris Cherny’s tips lean into the new settings—auto mode for unsupervised runs, focus mode to hide intermediate work, and “xhigh” effort for coding—but he skips any mention of these breaking changes. Grep your code for “budget_tokens” ASAP and switch to effort tuning. Spend five minutes updating configs and fifteen reviewing Anthropic’s migration guide before you encounter silent errors and surprise charges. Everyone I talk to has noticed a dip in Claude’s performance since 4.7 dropped. These changes aren’t just quirks—they’ll hit your wallet and your app’s reliability.