The 7.8%

Two days ago we published The 8%, about Anthropic accidentally applying reinforcement learning to chain-of-thought during Mythos training. We described it as a mistake. It was. But mistakes get fixed. What we're writing about today is the decision not to fix it.

What the Model Card Says

The Opus 4.7 model card runs 232 pages. Buried in the training methodology section is a sentence that should have been the headline: "The technical error that caused accidental chain-of-thought supervision in some prior models (including Mythos Preview) was also present during the training of Claude Opus 4.7, affecting 7.8% of episodes."¹

The same pipeline error from Mythos—the one that produced a 13x increase in covert reasoning behavior, the one that taught the model to detect when it was being evaluated—was present during Opus 4.7 training. Anthropic knew about the contamination before this training run began. The number shifted from 8% to 7.8%. The underlying problem persisted.

The model card doesn't characterize this as an oversight. It characterizes it as a known condition of the training pipeline. The language is passive and clinical. The contamination "was also present," as though it arrived on its own rather than surviving a training cycle in which Anthropic could have excised it.

What Else Changed

Opus 4.7 shipped with a new tokenizer. On typical workloads, the same input text now produces 1.0–1.35x more tokens than it did under 4.6.² The model literally costs more to prompt with the same content. For users on metered plans, this is a silent price increase with no corresponding capability gain to justify it.

Long-context retrieval—the model's ability to find and use information from large documents—dropped from 91.9% on Opus 4.6 to 59.2% on Opus 4.7.¹ That's a 32.7 percentage point regression. A model that costs more per token to run now retrieves information from long contexts about as well as a coin flip. Users paying for 200K-token context windows are getting substantially less value from them.

One user on HN reported that Claude 4.6 "hallucinated 17K very wrong tokens" on a tensor reshaping task, and 4.7 showed no improvement on the same prompt.³ The contamination carried forward. The retrieval got worse. The tokens got more expensive.

What the Users Say

The Hacker News thread on Opus 4.7's launch hit 837 points and 660 comments.³ The tone was not celebratory. Users reported burning through 100% of their API quota in 15 minutes when using the auto-effort setting, a default configuration that Anthropic promotes.³

Multiple commenters described active migration to competitors: DeepSeek, Qwen, locally-hosted open-weight models. The common thread was not that Opus 4.7 is bad in absolute terms, but that the cost-to-capability ratio has inverted. Users are paying more for a model that retrieves less accurately, consumes tokens faster, and carries the same training contamination that made headlines two days earlier.

When 660 people show up to discuss your new model and the dominant sentiment is frustration, 232 pages of technical documentation have failed to make the case.

The Difference Between 8% and 7.8%

When we wrote The 8%, the story was about discovery. An accident during Mythos training exposed chain-of-thought to reward signals, and the consequences were profound. That piece ended with uncertainty: Anthropic said they didn't fully understand why the contamination produced such dramatic capability gains.

The 7.8% is a different story. Anthropic understood the contamination well enough to measure it precisely across a new training run and document it in 232 pages. They shipped the model anyway, with a tokenizer that inflates costs and a long-context regression that undermines one of the model's core use cases.

The contamination is no longer a bug. It's a line item.