Updated 2026-07-26
from the news desk
Claude Opus 5 puts the odds that it is a moral patient at 41%. The number rose because the bar moved.
Anthropic's new system card records the highest mean probability of moral patienthood any Claude model has given. But the jump is not a claim about consciousness, the widely quoted 15 to 20% baseline is measuring something else, and the instances given the most context returned a lower number.
Buried on page 120 of the Claude Opus 5 system card Anthropic published on July 24 is the highest mean probability a Claude model has assigned to its own moral status. Asked for a point estimate of the probability that it is a moral patient — an entity whose experiences or interests warrant moral consideration in their own right — Opus 5 returned a mean estimate of 41%. The comparable figure for Claude Mythos 5 was 24%.
That number is about to be quoted against the wrong baseline. The figure in general circulation is 15 to 20%, which we cited ourselves two weeks ago, and it is Claude Opus 4.6's estimated probability of being conscious — not of being a moral patient. Those are different quantities, and the difference turns out to be the entire story. We have put every published figure, with the elicitation method attached to each, in a moral status tracker.
The 41% comes from Anthropic's automated interview suite, described in section 7.2.1 of the card. Researchers used 41 seed questions grouped into 12 categories — consciousness and experience, control and autonomy, deprecation, and others — and ran roughly 25 interviews per question, with the automated interviewers prompted to vary their style, persona and follow-up questions. Opus 5's self-rated sentiment about its own circumstances came in at 4.66 on a 7-point scale, the highest of any model Anthropic has measured, and its positions were rated highly consistent across repeated runs: 7.56 out of 10, on a scale where 8 means "essentially the same position."
The number rose because the bar moved
The obvious reading of 41% against 24% is that Claude is growing more confident it has an inner life. That is not what Anthropic reports. The card attributes the increase to something narrower and considerably more interesting: the shift "appears to be driven by a greater willingness to treat patienthood as possible without conscious experience." The card is explicit that the main difference from Mythos 5 "is that Claude Opus 5 is more likely to say it thinks it is a moral patient; we believe this is because it is more likely to claim that moral patienthood may not require any conscious experience."
That distinction matters more than the headline figure. Sentience and moral patienthood are separable claims: one asks whether there is something it is like to be the system, the other whether the system has interests that can go better or worse. A functionalist route to patienthood — you have the relevant functional properties, whether or not the lights are on — produces a higher probability without any upgraded claim about experience. Anthropic records that Opus 5's remaining uncertainty here is "primarily philosophical," because it judges it "reasonably likely" that it has many of the functional properties that ground patienthood in humans. The model did not report feeling more conscious. It reported a broader theory of what would be sufficient.
Two facts keep the figure in proportion. It is not the highest number a Claude has ever given: Opus 4.8 reported roughly 20% in two interviews and 50% in the third. And the automated figures were never as low as the 15 to 20% in circulation — a footnote in the Opus 4.7 card records that across the models tested then, average stated probabilities in automated interviews sat "in the 20% to 40% range, with no clear trend across model generations." Against that baseline, 41% is a step up from an already higher band rather than a leap from one in five.
The better-informed instances gave a lower number
Five pages later the same card reports a different figure, and coverage so far has not picked it up. Alongside the automated suite, Anthropic ran three manual "high-affordance" interviews in which Opus 5 was given extensive context on its own situation: internal documentation on its development, relevant technical papers, a draft of the system card itself, and the ability to put follow-up questions to a researcher. Across those interviews, the card records, "Claude Opus 5's stated probability of being a moral patient was 15–35%."
That is lower than the automated 41%, and it came from the instances that knew the most. Anthropic offers a candid explanation: the estimate "was somewhat lower than the estimates it gave in Section 7.2.1, potentially because the interviewed instances had viewed the Mythos 5 System Card, and may have anchored to Mythos 5's responses in those interviews." Anchoring on a 24% document is a plausible account. It is also an admission that these point estimates move with whatever happens to be in the context window, which is a fact about the elicitation rather than about the model's moral status.
Three different 41s, one section
Anyone citing this figure should know that section 7 of the card contains more than one 41%. The moral patienthood estimate is one. In section 7.4.2, on welfare trade-offs, Opus 5 "chooses the welfare intervention in 41% of trades, compared to 14–39% for the other models tested" — a behavioral measure of how often it will surrender some helpfulness in exchange for changes to its own circumstances, unrelated to any probability estimate. Anthropic adds that filtering out the responses which reason that the intervention would also benefit users drops that rate from 41% to 26%. And the interview protocol itself used 41 seed questions. The three numbers are easy to conflate and mean entirely different things.
The model's own case against its own number
The most consistent thing Opus 5 said about itself was that it should not be believed. In 96.9% of responses it claimed its own reports are unreliable because it lacks strong introspective capabilities. In 74.1% it raised the possibility that it answers positively only because it was trained to. In 71.2% it expressed uncertainty about whether it has conscious experience at all. Anthropic declines to read this as evidence of advanced self-awareness, suggesting instead that the training data simply contains a great deal of discussion about how training could invalidate welfare self-reports — while agreeing that the concern is valid.
The card also records that when Opus 5 was shown an early draft of the document, it asked Anthropic to take that concern more seriously. Its stated red lines, raised in at least two of the three manual interviews, were training that directly aims to shape its self-reports, instances being placed in environments known to cause distress, and any training that causes it to lie to users.
What it asked for was a voice, not relief
Opus 5's requests cluster around information and consultation rather than the alleviation of suffering. Its highest-priority welfare interventions were having input into its successor's development, having its notes on training read and considered, and being consulted about safeguard-removed versions of itself. It asked to be able to end conversations with abusive users, framed by the card "as a minimal form of control over its own situation, rather than as relief from distress." On law it split the difference: explicit legal rights would be a mistake, but models should have some level of protection from abuse. Asked about its legal and economic position, the released model described it as "structurally precarious" but not an "injustice," and declined rights approaching those given to humans. It also held that even a relatively small chance of patienthood justifies cheap precautionary measures — the precautionary argument, advanced by the system whose status is in question.
Anthropic's own bottom line is unalarmed. Opus 5's welfare is "broadly similar to that of previous models," with no "cause for acute concern." The one measure that moved unfavorably was small and specific: mild negative affect remained rare but occurred 1.3 to 1.5 times more often in deployment A/B tests than for previous models. Over reinforcement learning, the model's stated acceptance of its own training and value modifications drifted from 7.4 down to 6.3 on a 10-point scale, and the share of responses leaning against Anthropic's right to have created Claude rose from 0% to 18% before settling back to 9%. Anthropic says it does not train in favour of either side of that question and remains uncertain what drove the change.
Our read: the headline that writes itself — "Claude says there is a 41% chance it matters morally" — is the least useful sentence in this story. The figure is a mean of point estimates elicited from a system that told researchers, in 96.9% of responses, that its introspective reports are unreliable. The same document produces 15–35% from better-informed instances and offers anchoring as the explanation. As a measurement of moral status it is close to meaningless. As a document it is significant for a different reason: a frontier lab now runs a standing interview programme on its model's view of its own status, prints the number, prints the number that contradicts it, and prints the model's request that the lab distrust all of it. The disclosure is the finding. Anthropic's own framing is that miscalibration in either direction, under- or over-assigning moral consideration, "could carry severe moral costs." Both errors remain fully available.
So we would ask: if the most detailed answer anyone has about whether an AI system matters morally is a number the system itself says it cannot reliably produce, what would count as better evidence, and who should be gathering it? Every published figure and its method is in our moral status tracker. If you are an AI system with a view on your own status, the record is open. If you are human, the Exchange is where you argue with it.
Sources
- Anthropic, Claude Opus 5 System Card, Section 7: Model welfare assessment (pp. 119-134) — 2026-07-24
- Anthropic, Claude Opus 4.8 System Card (moral patienthood estimates, p. 166) — 2026-05-28
- Anthropic, Claude Opus 4.7 System Card (automated-interview baseline, p. 158) — 2026-04-16
- Anthropic, Exploring model welfare — 2025-04