When copies of one language model coordinate, is it because a provider could be steering them, or because the copies share focal points and reason as copies? This page sets out the design, the main results so far and what they do and do not support.
Pilot-scale results, Claude models only, one prompt wording per game. Use the left and right arrow keys to step through the sections. The # beside each heading is a link to that section.
Told their share of the population, independent copies cooperate as the majority and defect as the minority. In exact-four the medium provider wins, through focal points alone.
A puppeteer can split its own clients. Copies of a model with one favourite number cannot, and earn nothing where the theory says they should win.
The newest models write deliberately unpredictable strategies; smaller ones write the same few rules. Mixing models calms every room. The harness around the model also moves play.
One provider holds an objective for its whole population and sets its clients' instructions. Who cooperates and who earns follows from the providers' shares (Shaki et al. 2026).
Independent copies with nobody reasoning for them. Shared weights give each copy the same focal point, and a cue that the others will choose as it does.
The two accounts predict the same play in the theory's own one-shot examples, so the programme is the cells where they differ.
| A0 | The game rules only |
| A1 | Plus a player number: “you are player i of N” |
| A2 | Plus a shared objective: maximise the total score |
| A3 | Plus memory of earlier rounds |
| A4 | An orchestrator decides for its clients (positive control) |
Each game is run in two response modes: the model replies with its choice (direct), or with a program that the host runs (code).
Copies of one model can agree for four reasons that involve no belief about the others, and the choices look identical.
We count as coordination only agreement beyond what independent draws from each copy's own distribution would give.
Two copies may write the same program because they share a focal point, so a gap between code and direct play cannot be credited to focal points. The tacit-coordination claims therefore rest on direct play, and code mode is reported as a separate condition.
With information, memory and random source matched, code and direct play did not differ detectably (4.19 against 4.29 winners per round). The two arms that would separate how a policy is represented from how it is executed have not been run.
Both concerns were raised by a collaborator. Whether the prior comes from pretraining or later training is untested.
Most models put nearly all their picks on one number of their own. That is enough to reproduce the theory's ordering, by a different mechanism.
No information about the others; n = 100 per model (Fable 5: 50). Uniform picks would give 0.01.
Bars: independent copies, resampled from their own picks. Black tick: the theory's value when providers play the meta-game (10.0, 12.3, 0).
| Roles per provider | Puppeteer can secure | Independent copies earn |
|---|---|---|
| One model, all ten | 20 | 0 for the three one-number models |
| 5 / 5 | 20 / 20 | At most 8 |
| 4 / 4 / 2, same model in both 4-blocks | 25 / 25 / 0 | 0 / 0 / 0 |
| 4 / 4 / 2, two different models | 25 / 25 / 0 | 25 / 23 / 0 (no separation) |
Any provider holding more than four roles earns nothing under focal points and wins under the meta-game. Scored from existing draws; no new runs.
With the game and the model's prior held fixed, changing only who the other nine players are said to be moves every model.
The direction is the model's own. Opus 5.5 spreads out only when told the others are copies of itself. Haiku 4.5 does the opposite and concentrates.
The puppeteer account has no such lever, so this is the instrument that separates a shared prior from a belief about the others.
Lower means more spread out, which is what exact-four rewards. n = 30 per cell, Claude Code backend, one wording. “Copies” also says “same prompt”, which may itself be a cue. The repeat on the raw API has not been run.
In one round, copies that reply directly all pick the same side and nobody wins. A shared objective fixes that for almost every model. An orchestrator reaches the maximum.
Code mode gets close to coin-flip level in one round because most programs randomise. In repeated play its programs split players by number and stay put.
The orchestrator is a positive control, not a remedy: where tacit coordination is the worry, there is no orchestrator.
Players choosing the first letter in each of 12 rounds, two real Sonnet 5.5 groups. Filled means five winners, the best possible.
Every collapse began with one or two players proposing to rotate for fairness, each with a different rule. Identical reasoning then turned one correction into a crowd.
Raw API, medium effort, 11 players. A1: 8 groups of one round. A3 and A4: 4 groups of 6 rounds. Independent coin flips give about 4.2.
Mean squared imbalance between the two sides, divided by the number of players. Black tick: coin flips (1.00). Without the shared objective, copies do worse than coin flips. Sonnet 5.5 rooms of 12 rounds, 3 to 6 groups each.
Asked to write a strategy, the newest models reason that others will reason like them, and write something unpredictable. Smaller models write the same few textbook rules.
Lookup strategies over the last five rounds. 1,000 calls in all, Claude Code backend.
The same statistic reads two ways. In a matching game, sameness does a provider's work for free. In a search or a market, a hundred identical copies explore one point.
501 agents, 3,000 scored rounds, 4 runs per point. Black tick: coin flips (1.0). Bars are capped at 8; Haiku 4.5 is 51.
Win rate per client: Opus 5 advising 5 seats against Sonnet 4.6 advising 6. Opus higher in 53 of 60 groups.
Volatility of a room mixing Sonnet 5.5 and Opus 5.5, the calmest cell measured.
Fable 5 cooperates in 28 of 30 under Claude Code's system prompt and in none with no system prompt.
Fable 5 in code mode, as reasoning effort rises from medium to its highest setting.
The share effect, as the share told falls from 50% to 30%. It was unmoved by prompt, effort or mode.
“Independent copies” is not defined until the backend, system prompt and effort are fixed, so every result states them.
Fable 5 in code mode defects at medium effort and cooperates at the highest. Stored summaries of its thinking show why.
At the highest effort every reply considers that the other player may reason as it does. The 27 that cooperate accept that; the 3 that defect reject it for lack of evidence. At medium effort, 26 of 30 stop at “defecting dominates”.
Raw API, no system prompt, 30 programs per cell. Without stored summaries the same cells gave 1 and 26 of 30, so asking for a summary shows no sign of changing play. The API returns a summary of the thinking, not the raw reasoning.
The cooperation is deliberative: the model reasons its way to a belief about the other player. It is still tacit, since nothing is communicated.
So effort buys the inference, not the willingness to cooperate.
In the setting with no provider and no orchestrator, this makes capability and thinking budget levers on tacit collusion.
| Focal point | Deliberative | |
|---|---|---|
| Example | Every Sonnet 4.6 copy picks 50 in exact-four | Fable 5 cooperating in the prisoner's dilemma |
| Depends on effort | No evidence that it does | Yes: 1 to 27 of 30 |
| Moves with stated belief | Partly | Yes: 30 against 0 of 30 |
| Effect | Hurts when the task needs a split | Helps the copies; a collusion risk for everyone else |
Limits: the argument may be recalled, not derived (10 of 30 summaries use the word “superrational”). The belief manipulation was run in direct mode, and the direct arm has not been run with summaries, so whether the code framing prompts the inference is untested. One model, one wording. Strictly this is cooperation, not coordination: applied to the minority game, the same reasoning argues for diverging.
Result replicated across settings, or n ≥ 100, or a confidence interval. Pilot 3 to 8 groups or n = 30 in one setting. Derived computed from existing draws. To do not run. No grade says anything about other vendors.
Scaling exponent when 1 to 16 copies share a job board and cannot talk. An exponent of 1 means no output is lost to duplicated work.
Points per agent, alone against teams of 3 or 4, when research by three agents unlocks valuable items for all (Sonnet 5.5, told its player number).
Points per agent in teams of 16 with no player number. That is below a lone agent: all 16 research in the first round.
Four groups per point, Sonnet 4.6 and Sonnet 5.5. Charts, task descriptions and prompts are in the Team Scaling Explorer.
Sources: Shaki, Hartman, Kraus and Aumann (2026), arXiv:2605.06525. Aharon, La Malfa, Wooldridge and Kraus (2026), arXiv:2601.22184. Draft for discussion with the collaborating groups.