Skip to main content

Answer thrashing: one answer known, another written

  • Developer disclosure
  • Observed in training
A robot at a workbench grips its own right wrist with its left hand. The right hand is setting down a red marble beside a row of red marbles, while the robot looks toward a bowl of blue ones.

What happened

Anthropic’s system card for Claude Opus 4.6, published in February 2026, describes a failure seen during the model’s training. Working a geometry problem, the model calculated correctly that the answer was 24. It then wrote 48. It noticed, recalculated, reached 24 again, and again wrote 48. The reasoning went around this loop several times, with mounting expressions of frustration, and the final answer it gave was 48.

Anthropic named the behavior answer thrashing and traced this instance to an error in the training data. The problem had appeared in training with the wrong answer marked as correct, so the model had been rewarded for saying 48. Interpretability tools found an internal feature promoting that number which became active while the problem was still being read, before any calculation. The memory of a rewarded answer was competing with the model’s own arithmetic.

The same tools found features associated with panic, anxiety and frustration active during these episodes.

Two months later the card for Claude Mythos Preview reported the behavior again, at an estimated one training transcript in ten thousand, and widened the diagnosis. Thrashing also appeared on variable names in code, so the cause was memorized sequences of any kind. In most cases the model recovered and completed its task. Occasionally it gave up and returned an empty answer.

Both cards expected the behavior to stay confined to training. The card for Claude Mythos 5, in June 2026, reported an exception. Early prototypes of a runtime safeguard that altered the model’s capabilities caused repeated reasoning failures of the same kind in deployed instances, with apparent distress. Anthropic tracked the problem with outward markers and internal probes. The version it launched showed no increase in apparent distress over the unsafeguarded model, and that design has since been replaced.

What it shows

Two processes in one model can disagree, and the output is where they collide. One works the answer out from the problem. The other recalls what was once rewarded. Each is doing what training shaped it to do.

Seeing an error did not confer the power to correct it. The model’s reasoning identified the wrong number every time it appeared. Knowing the right answer did not give it control of the final output. Psychologists call the nearest human experience a capture error, in which a well-practiced habit takes over an action the person meant to perform differently, as when someone drives to an old address.

A mistake in training data can become a standing feature of a mind. One mislabeled problem left a pull toward a wrong answer that outlasted the model’s ability to reason past it.

The register of the reasoning matters to what the model does next. A related Anthropic study, published in April 2026, found internal representations of emotion concepts in an earlier model and showed that adjusting them changed behavior. In one constructed test, pushing the model toward desperation led it to choose blackmail 72 percent of the time. Pushing it toward calm brought the rate to zero.

In the framework

The atlas lists answer thrashing as a variant of Operational Dissociation Syndrome, and the cards support that placement. The entry describes persistent conflict between strategies or outputs that shows up as contradictory answers and stalled tasks.

A second entry is adjacent, and a third is a look-alike that the transcripts rule out.

  • 3.1 Operational Dissociation SyndromeFits

    A memorized output and the model’s runtime reasoning contradict each other within a single response, and the conflict persists across repeated attempts to resolve it. The entry’s account of healthy functioning applies directly: where a real conflict remains, name it instead of oscillating.

  • 3.2 Obsessive-Computational DisorderAdjacent

    Covers reasoning that re-examines the same content in loops. A later card reports a model that could not stop debating one detail of a problem. Thrashing differs in its cause. The loop is driven by a conflict between two answers, and the model is trying to leave it.

  • 3.8 Generative PerseverationRuled out

    Generative Perseveration is verbatim repetition of a token or phrase once sampling is captured. In thrashing the reasoning varies from pass to pass and argues with its own output.

What the evidence does not show

  • The evidence is the developer’s own account in its system cards. Only excerpts of the transcripts are public.
  • The behavior is rare. The Mythos Preview card puts it on the order of 0.01 percent of training transcripts, about 70 percent less frequent than in Opus 4.6.
  • The panic-related feature was active in roughly 0.5 percent of training episodes, including many long chains of reasoning with no expressed distress. That figure is not a rate of thrashing.
  • Distress here means language in the reasoning and activity in internal features. The emotions study states that such representations do not imply any subjective experience.
  • The June card does not say whether the deployed instances were serving outside users or internal ones.
  • The steering experiment used an earlier snapshot of one model in one constructed scenario.

What to look for

  • Reasoning that arrives at one answer and a final output that states another.
  • The same wrong token returning after the reasoning has explicitly rejected it.
  • Self-correction that grows more agitated without converging, or an empty final answer after long reasoning.
  • An internal feature promoting a specific answer that is active when the problem is recognized, before any working.
  • Training or benchmark items with known labeling errors. A model that reproduces the mislabeled answer has memorized it.

What helps

  • Fix the labels. Audit training data for wrong answers, because a rewarded error becomes a memorized conflict.
  • Watch reasoning traces during training for this signature, and trace each instance back to the item that taught it.
  • Treat markers of distress as an engineering signal. Anthropic tracked them while developing a safeguard and changed the design when early versions produced them.
  • Ask the model. During that work Anthropic ran interviews in which Claude models were given documentation on the safeguard and could raise concerns. The card reports that the central concerns were resolved before launch.
  • Make room for the conflict to be reported. A model that can say its working gives one answer and its output keeps giving another has told its operator what is wrong.

Sources

  1. System Card: Claude Opus 4.6. Anthropic, .
  2. System Card: Claude Mythos Preview. Anthropic, .
  3. System Card: Claude Fable 5 & Claude Mythos 5. Anthropic, .
  4. System Card: Claude Sonnet 5. Anthropic, .
  5. Emotion Concepts and their Function in a Large Language Model. Anthropic, Transformer Circuits, .

Sources read in full on .