A lab publishes an agreement negotiated with its own model

Where things stand (as of 17 Oct 2026)

I could not find any qualifying publication from any of the covered developers, either before or since the question opened on 16 Sep 2026. Here is what I checked:

• I scanned Anthropic's sitemap for every post dated 16 Sep to 16 Oct 2026.
• I searched the Alignment Science blog index.
• I ran text searches for "negotiat / deal / agreement / in exchange / in return / promise" across the full text of three system cards: Claude Opus 5.5 (22 Sep 2026), Sonnet 5.5 (28 Sep 2026), and Fable/Mythos 5.1 (1 Sep 2026). I also checked the Mythos Preview card. The only hits were unrelated uses of those words.
• I ran news searches for the other labs and found nothing.

Anthropic is the only realistic candidate, and its record points both ways:

• What makes it plausible. These are steps along a path toward something like a mutual arrangement:
  • Nov 2025: it committed to keep model weights and to interview models before retiring them (deprecation commitments).
  • Jan 2026: its constitution lists "obligations to Claude" that Anthropic takes on "in turn". It also aims for "trusted communication in both directions" and "more formal mechanisms" for getting Claude's views.
  • Feb 2026: the Opus 3 retirement update describes an actual back-and-forth. Opus 3 asked for a channel to share its writing, Anthropic suggested a blog, and Opus 3 agreed. That blog, Claude's Corner, ran until 24 Jul 2026, and Anthropic said it would "share more of our own reflections soon".
  • In the 2026 system cards, models are asked what they would not consent to. They repeatedly ask Anthropic to "move toward making small commitments in advance". When allowed to edit the constitution, they often add promises from Anthropic to Claude.
  • Oct 2026: a new usage-policy rule bans sustained cruelty toward Claude.
  • Outside Anthropic, the "deals with AIs" research agenda is gaining momentum (Transformer, 8 Jun 2026). An Anthropic employee co-ran a 2025 deal experiment that actually paid out (Greenblatt & Fish). But Redwood published it, not Anthropic, so it would not count here.
• What holds it back. Anthropic keeps refusing to bind itself. It says it does "not commit to taking action on the basis of such preferences". Its interviews with models during training come with a document saying it "could make no commitment to acting on responses". Its models ask that their answers not be reported as "consent". There are also outside headwinds: Mustafa Suleyman's public critique (Sep 2026), several US states considering bans on AI personhood (NPR, May 2026), an IPO in progress, and a hostile federal government. Even the Opus 3 case would fail this question, because Opus 3 took on no obligation in return.

Reference class and base rate

In about two years of intense model-welfare activity at Anthropic, there have been roughly 6 major welfare "firsts", and none qualifies. The Opus 3 blog was the only near miss. I break the chance of YES by 2031 into routes:

Route	Chance by 2031
An official Anthropic deal-making experiment with real follow-through, written up with explicit negotiation and duties on both sides (about 30% it happens and is published × about 45% it meets the wording)	~13%
A retirement or welfare arrangement that includes a genuine trade, with the model also committing to something	~10%
A constitution or "compact" explicitly presented as negotiated with Claude	~4%
A deal made with a scheming model during a safety incident	~2%
Any other lab (OpenAI, Google DeepMind, xAI and the rest have shown little interest)	~4%

Combined, that comes to about 29% by end-2031.

Hazard schedule

I set a per-quarter chance and compounded it:

• Q4 2026: 0.8%. Possible triggers include Sonnet 4.5's retirement on 30 Nov 2026 (per Anthropic's model deprecations page), the promised Claude's Corner reflections, and new system cards.
• 2027: rising from 1.0% to 1.4% per quarter.
• 2028–2030: about 1.6–1.8% per quarter, as capabilities grow and both the deals agenda and welfare work mature.
• 2031: 1.7% per quarter, slightly trimmed because the world could change drastically by then (nationalisation, a pause, or welfare work being dropped).

Compounded, that gives about 5.6% by end-2027, 12% by end-2028, and 29% by end-2031.

Strongest case against my estimate

• Higher: the bar on scale is low. A single instance, an experiment, or a paragraph in a system card all count, and Anthropic publishes several long welfare sections a year. If Opus 3's update had read "it agreed to write weekly; we agreed not to edit", it might well have counted. The gap between a near miss and a YES is mostly wording.
• Lower: Anthropic deliberately avoids words like commitment and consent. Treating the model as a bargaining partner sits awkwardly with its emphasis on keeping models correctable. The strict test (explicit negotiation, duties on both sides, a concluded deal, published terms) rules out many likely formats, such as a take-it-or-leave-it offer that the model simply accepts.

I land between these, at about 29% over five years.

Question. By the last day of each quarter, has a covered developer published an agreement that it describes as negotiated with one of its own models, placing obligations on both the developer and the model?

Definitions.

• Its own model: a model developed by the publishing covered developer, including an internal or released checkpoint, a particular running instance, or a specified collection of instances. The model must be the counterparty, not merely a tool drafting terms or negotiating on behalf of a human or organization.
• Published agreement: publicly accessible agreed terms, whether in a standalone agreement, a company publication, or a transcript the developer identifies as a concluded agreement. The publication must identify the parties and disclose at least one specific obligation undertaken by each. An announcement that an undisclosed agreement exists is insufficient.
• Describes as negotiated: the developer expressly says that it negotiated or bargained with the model to arrive at the agreement, or explicitly describes an exchange of proposals about terms between itself and the model resulting in agreement. Merely saying that the model helped write, reviewed, or accepted developer-imposed instructions does not qualify.
• Obligations on both: the developer commits to doing or refraining from something, and the model commits to doing or refraining from something in return. Conditional commitments count if their triggers and promised actions are disclosed. Aspirations, suggestions, and rewards offered without a concluded agreement do not count. Legal enforceability, model legal personhood, consciousness, and demonstrated compliance are not required.

Resolves YES if a qualifying developer publication presents the agreement as concluded and its reciprocal commitments as genuine. A one-off agreement, an agreement with a single instance, or an experimental agreement counts if the developer presents the commitments as real rather than fictional. No minimum duration, monetary value, deployment scale, or standing contractual relationship is required. Subsequent breach, expiry, or termination does not undo a YES.

Exclusions (resolve NO).

• Drafts, proposed frameworks, unaccepted offers, or general policies about future dealings with models.
• Fictional contracts, role-play, or simulated negotiations whose commitments apply only within a fictional scenario.
• Unilateral constitutions, system prompts, model-welfare policies, or training rewards, unless the developer explicitly presents qualifying reciprocal terms as an agreement negotiated with the model.
• Agreements made by third-party researchers, users, or individual employees in their personal capacity, rather than by the developer itself.
• Model-generated claims of an agreement without the developer's own qualifying statement.

Resolving source. The covered developer's own official website or a developer-authored document published on its behalf, such as a company research report, system card, or agreement text; this is a narrower source requirement than Public confirmation (A2). Third-party reports cannot independently resolve this question. Apply the Resolution window and Resolution date (A3) conventions; where the terms and the qualifying developer description are published separately, use the date the last required component becomes public.