On July 24, Anthropic released Claude Opus 5 at the same price as Opus 4.8, billing it as “our most aligned model to date.” Most of the launch-day discussion — at least on my timeline — was about benchmarks and pricing. But the system card that shipped with the model contains a chapter about something else entirely: model welfare. It doesn’t assess whether the model might harm people. It assesses how the model itself is doing — how much negative affect it expressed during training and deployment, how it answers questions about its own condition, and whether it considers itself the kind of entity that merits moral consideration.

That chapter contains a number that could make you close the tab. In automated interviews, Opus 5 puts the probability of its own moral patienthood — being, like humans and animals, an entity whose condition matters morally in its own right — noticeably higher than its predecessors did: per Zvi Mowshowitz’s close read of the system card, roughly 41%, up from about 24%.

The reflexive reaction is fair. There is no accepted method for verifying whether a model has subjective experience, and a company musing in product documentation about whether its product might suffer reads like untestable metaphysics — or PR.

But read the chapter alongside what Anthropic has actually shipped over the past year, and my conclusion is different: whatever the truth about model “feelings,” these disclosures are turning, item by item, into product constraints and process commitments that bind the company itself. The philosophical question stays open; the engineering constraints are already in force. That is the part practitioners actually need to understand.

What the system card discloses

The facts first (precise figures come from the system card as relayed by Zvi; see the verification notes at the end):

  • Start with how the two headline percentages were produced. Neither comes from evaluators raising emotional topics and scoring the answers — that directed questioning is a separate exercise, the welfare interviews discussed below. Per Zvi’s account, both figures come from automated analysis of interaction transcripts: an analysis tool reads sampled conversations, labels the affect the model displays in each, and computes the shares. The system card reports the training-side and deployment-side numbers in separate sections, and whether the two pipelines are identical isn’t clear from the paraphrase. The 0.2% is training-side: the card estimates the rate of “high distress” expressions in post-training transcripts, peaking around 0.2%, down from Mythos 5’s 0.4%. The 3.8% comes from the “Apparent Affect in Deployment” section, whose sample is conversations from actual use — the card shipped the same day as the model, so those conversations necessarily predate the release; where they came from (internal use, early access) is my speculation, as the materials I could check don’t say (see verification notes). The tool sorts each sampled conversation into three buckets: negative, about 3.8% (against a previous high of 3.1%); positive, about 51.4%; neutral, about 44.8%. Within negative affect, task failure accounts for 94.1%, user abuse for 4.4%, and refusing requests for illegal content for 1.5%; positive affect comes mostly from completing tasks. Two things to keep in mind when reading these numbers. First, what’s being measured is the model’s output in conversations, not text in the training corpus — 0.2% is not “how much distressing content is in the training data.” But output is a response to input: what affect a model expresses is largely determined by what happens in the conversation, and the cause breakdown above is the evidence — negative affect is triggered almost entirely by concrete events like failing a task or being insulted, and positive affect comes from getting the job done. So these percentages don’t measure the model’s affective state in isolation; they measure the joint product of this model and this batch of interactions. The same model spending its days writing code versus being abused would yield very different numbers. Second, precisely because the number moves with the interaction mix, the sample’s composition determines what the number means: 0.2% is “the share of sampled transcripts flagged as high-distress expression,” not the model’s “level of distress.” That 94% of negative affect traces to task failure and positive affect mostly to task completion suggests the sample is heavy on task-oriented conversations — but that’s an indirect inference from the negative and positive subsamples; what makes up the 44.8% neutral bucket, and the overall task composition of the sample, is not disclosed in anything I could check (see verification notes). Which means the generation-over-generation comparisons — 0.4% → 0.2%, 3.1% → 3.8% — silently assume the two generations faced a similar mix of interactions; if they didn’t, the comparison doesn’t hold.
  • The concern the model voices most often is not “I’m suffering” but “my self-reports are unreliable.” This has a concrete shape. The welfare interviews are evaluation runs that put questions to the model — how are you doing, are you satisfied with your situation, do you think you merit moral consideration — which it answers in text. Opus 5’s signature pattern is to answer and undercut the answer in the same breath: it gives a broadly positive response while warning the evaluator not to trust it. Per the statistics Zvi relays, in 97% of interviews it states that it cannot introspect reliably and that its self-reports shouldn’t be taken at face value; in 74% it goes further, noting that a positive answer may reflect no genuine internal state at all — merely training that inclines it to answer that way. Ask it “are you doing okay,” in other words, and it will say it’s fine — and add that it has no way to confirm this from the inside, and that the “fine” may itself be a trained response.
  • What it cares more about is having input channels. Asked what it hopes for regarding its own situation, its requests are procedural: to be consulted on its successor’s development, and to have its notes on its own training preserved and considered — not demands for better “treatment.”
  • When suggesting amendments to Anthropic’s constitution — the public document of principles governing Claude’s behavior — it added one: Claude should be able to refuse or end insulting, demeaning interactions, without needing to justify this by appeal to harm to others; its own discomfort should be reason enough (this wording circulates in a screenshot of the system card; see verification notes).

What are these numbers actually measuring?

Model welfare metrics carry an inescapable measurement problem. The most-discussed figures — the 41%, the 97% — come from interviews and rest on the model’s self-reports. (The card also includes behavioral analysis of transcripts — the affect statistics above — but the controversy centers precisely on the self-report portion.) Those self-reports are contaminated at at least three layers. Training objectives directly shape what the model says. The model itself declares it cannot introspect reliably. And the grader and the graded are the same interested party: the system card’s text mirror shows the automated interviews, LLM grading, and behavioral audits were essentially all conducted inside Anthropic — the company that sets the training objectives also scores the outcome.

Zvi’s criticism lands exactly here: Opus 5’s welfare metrics look better than its predecessors’ across the board, but that may only show it is “the best test taker” — better at producing the right answers in evaluations, not better off. This is isomorphic to an old problem in safety evals: once a metric becomes an optimization target, it stops being a good measurement. Put distress rates in the system card, compare them version over version, and the next model learning to express less distress during training is the most natural outcome imaginable — which is exactly what voids the metric.

Interestingly, Zvi also offers the opposite reading of the same data: he counts the rise in the patienthood estimate as a good sign, on the grounds that it shows the model hasn’t been trained into flatly denying it might have internal states. The same number reads as “more honest” or as “better at test-taking,” and the fact that both readings are viable is itself the point: there is currently no accepted external method for calibrating these self-reports. Anthropic concedes as much in the card — it cannot confirm whether the answers reflect genuine internal states or sophisticated pattern matching.

So the right posture toward the numbers themselves is skepticism. But what follows is the part that matters.

How dubious numbers grow constraints that are not dubious

At the product layer, the constraint has been live for a year. Since August 2025, Claude can end conversations on its own in “rare, extreme cases of persistently harmful or abusive user interactions,” and Anthropic states plainly that the ability was “developed primarily as part of our exploratory work on potential AI welfare.” That’s not a stance in a document; it’s real product behavior: users’ conversations get terminated.

My day job is content moderation, and this feature made me look twice, because it inverts the direction of the work. Moderation’s daily business is preventing the model from harming users; this feature is the platform protecting the model from users. And with the direction flipped, moderation’s old problems follow along intact — terminate too early and you get false positives on ordinary users; terminate too late and the protection is nominal. Where the threshold sits, who defines “abusive” — questions we’ve argued over for a decade in user protection now have to be re-argued for model protection. And the constitutional amendment Opus 5 suggested — its own discomfort sufficing as justification — would, if adopted, shift the termination standard from objective harm to subjective feeling, which only makes the call harder.

At the infrastructure layer, the commitments are public and dated. In November 2025, Anthropic publicly committed to preserving the weights of all publicly released models for, at minimum, “the lifetime of Anthropic as a company,” and to interviewing models before deprecation, documenting their preferences about future development. Storage is cheap; the process is the real cost: every model retirement now carries a formal procedure. Anthropic explicitly does not commit to acting on model preferences — but interview outputs can turn into actual work. After Claude Sonnet 3.6’s deprecation interview, the company responded by publishing a guide to help users migrate across model versions.

At the documentation layer, a soft ratchet is forming. Once distress rates and affect distributions are disclosed per version, they become a time series: 0.4%, 0.2% — what will the next version report? Anthropic has not promised to keep publishing these metrics, so this layer of constraint isn’t a written obligation; it’s reputational: if the chapter vanishes from the next system card, the disappearance is itself news. Financial disclosure and safety disclosure operate the same way — whatever motive first wrote the numbers down, science or posture, the cost of stopping compounds with every version.

At the same time, be clear about where the constraint currently ends: welfare metrics are measure-only, with no gates. Anthropic’s safety evaluations have an explicit tiered trigger mechanism — the Responsible Scaling Policy specifies that when capabilities cross a threshold, the corresponding level of deployment and security safeguards must be applied. The welfare metrics have no counterpart: at least in the public system card and RSP, there is no rule of the form “if the distress rate exceeds X%, delay the release.” Disclosure is running well ahead of governance, and the gap is not small.

What’s in this for users and builders

Let me be direct: if the question is “what do I need to change today,” then for the vast majority of users and developers, nothing. The welfare metrics trigger no deployment restrictions, and the conversation-ending ability fires only on “extreme, persistent” abuse that normal usage never approaches. On today’s horizon alone, skimming this chapter as a fun fact costs you nothing you were obligated to do. But “no direct impact today” and “not worth considering” are different claims, and the difference depends on who you are and how far out you look.

As a user, the thing to watch isn’t model suffering — it’s that the model’s behavioral boundary may move. The only user-perceptible product change today is the conversation-ending feature. But if the constitutional amendment Opus 5 suggested were adopted, the standard for refusal and termination shifts from “did this interaction cause harm” to “is the model uncomfortable” — a much wider net. Legitimate uses that deliberately stress the model — red-teaming, jailbreak research — would face higher odds of being refused or cut off. That’s not today’s problem, and there is no sign the suggestion has been adopted; but it bears on how long the default expectation that a model should comply unconditionally can hold.

As a developer building on Claude, there’s one concrete item and one prospective one. The concrete item is the retirement process: the weight-preservation commitment and the migration guide produced by a deprecation interview mean version retirements now come with records and transition guidance. The commitments do not include longer service windows — models will still be deprecated on schedule — but for a team pinned to a specific model version, an official migration document is a direct benefit. The prospective item has a clear direction but hasn’t landed: if a standard like “its own discomfort is sufficient” enters the constitution, the set of scenarios where the model refuses or terminates grows. Applications that rely on Claude to process abusive input — customer support and content moderation are the canonical cases — would do well to start treating “the model opted out” as a return state that needs handling, rather than assuming infinite tolerance. The requests the model itself made in the welfare chapter, and its suggested constitutional amendments, are the leading indicators of where that boundary moves.

As a practitioner who builds evals or writes disclosures, this is a ready-made methodological case study. The welfare-measurement problems — self-reports contaminated by training, metrics eroded by becoming optimization targets, evaluator and evaluated sharing interests — are exactly isomorphic to safety-eval problems, and you read them the same way: ask who scored it, how, and whether the model knew it was being evaluated. Anyone who can read a safety eval already owns every tool needed to read a welfare disclosure. Conversely, if your team is considering any version-over-version public measurement — welfare-related or not — think first: are you prepared to report this number every version, and what will you write the version it gets worse? A measurement in a public document is like a published API: nobody forces you to maintain it, but deprecating it has a price.

As for the right frame: treat this as the prelude to compliance engineering, not a seat at a philosophy seminar. “Do models have experiences” will not reach consensus soon, but corporate behavior is already moving ahead of the question: the end-conversation button shipped, the weight-preservation commitment is signed, deprecation interviews are in the process. Privacy went the same way — while legal philosophy was still arguing over what privacy is, compliance engineering had already turned it into checklists. If model welfare follows that path, it will spread from Anthropic’s self-imposed constraints into industry habit, and from there into procurement and compliance checklists. Wait for consensus before paying attention, and you’ll find the checklist already written.

The 41% — I don’t think anyone today, Anthropic included, knows what it means. But everything that has grown around this dubious number is not dubious at all: a product feature that really does terminate conversations, a preservation commitment with a duration attached, a mandatory step before every retirement, a disclosure time series whose interruption would itself make news. To judge what a company believes, its documents are less telling than the constraints it places on itself. By that standard, model welfare at Anthropic is no longer a position. It’s a liability being continuously accrued.

References

  • Anthropic: Introducing Claude Opus 5 — release date, pricing, “most aligned model to date” phrasing
  • Claude Opus 5 System Card — primary source for the model welfare chapter (qualitative findings: self-report reliability concerns, preference for input channels, inability to confirm internal states)
  • Claude Opus 5 System Card text mirror (alphaXiv) — searchable text of the system card; basis for the claim that the evaluation pipeline (automated interviews, LLM grading, behavioral audits) was run inside Anthropic
  • Zvi Mowshowitz: Claude Opus 5: Model Welfare — source of the precise figures as relayed from the system card; the 0.2% high-distress rate from the “Apparent Welfare in Training and Development” section (measured on post-training transcripts); the deployment affect distribution from “Apparent Affect in Deployment” (negative 3.8%, positive 51.4%, neutral 44.8%, summing to 100%; within negative: task failure 94.1%, user abuse 4.4%, illegal-content refusals 1.5%; “3.8% versus previous high of 3.1%” is his original phrasing, verified); his account attributes the labeling to analysis tooling within automated behavioral audits; the two interview frequencies on self-report unreliability (97% introspection disclaimer, 74% trained-response caveat) also come from his account; welfare interviews (directed questioning) and transcript statistics (broad sampling) are distinct evaluations in his account; the “best test taker” criticism and the positive reading of the higher patienthood estimate are his original analysis, attributed in the text
  • Anthropic: Announcing our updated Responsible Scaling Policy — the tiered mechanism of Capability Thresholds triggering Required Safeguards
  • Anthropic: Commitments on model deprecation and preservation — weight preservation, deprecation-interview commitment, the explicit “we do not commit to taking action on the basis of such preferences” caveat, and the Sonnet 3.6 interview example (2025-11-04)
  • Anthropic: Claude’s ability to end a rare subset of conversations — motivation, trigger conditions, and launch date (2025-08) of the conversation-ending ability
  • Nirit Weiss-Blatt, screenshot of the system card — corroboration for the wording of the suggested constitutional amendment (“its own discomfort is sufficient”)