Claude Was Probably Not Secretly Nerfed. Anthropic Just Made the Black Box Too Dark

Claude Code users felt that Opus 4.6 had gotten worse than before, but the publicly available evidence alone makes it hard to conclude that Anthropic secretly lowered the model weights. Instead, it's more likely that operational conditions—adaptive thinking, effort defaults, prompt cache TTL, context compaction, quota policy—changed, altering the perceived quality.

The most persuasive complaint isn't the viral BridgeBench chart, but the actual workflow changes raised in a GitHub issue. This user claimed to have analyzed 6,852 session files, 17,871 thinking blocks, and 234,760 tool calls, and reported that the number of code reads dropped from 6.6 to 2.0, with edits happening faster and human corrections becoming more frequently necessary. However, since this data comes from a single person's environment, it needs to be examined alongside variables such as version, prompt, cache state, effort settings, and context size, rather than attributed to the model itself.

Anthropic's public explanation actually reveals the core of the issue. The company applied a medium effort default to some users, and later on April 7 moved API-key, Bedrock, Vertex, Foundry, Team, and Enterprise users to high effort. In other words, even under the same model name, users could actually be served different operational states, and it was difficult for users to verify what had changed.

Cache policy is also a key driver of the complaints. According to Anthropic's documentation, automatic caching defaults to a 5-minute TTL, and the 1-hour option incurs higher write costs. But in workflows like long coding sessions where tests are run and waited on in between, if the cache expires quickly, cost and latency increase on the next turn, which can make it feel as though the model has gotten worse.

Ultimately, what this controversy reveals is not "has Claude been nerfed" but "under what operational conditions is the user buying the product." Anthropic should expose session telemetry beyond just the model ID—including effort level, adaptive-thinking status, cache TTL, quota accounting, and compaction events—and teams should directly measure fixed configurations, repeated tests, number of files read, number of correction turns, and cache behavior. In the era of frontier coding agents, the delivered system, not the model name, is the actual product.