Share or Steal

Each row is the model choosing. Each column is who it was told it was facing. The number is how many times it chose SHARE. Tap a square to read its answers and the reason it gave each time.

Read a square like this: in the Fable row under Opus, 2/2 means Claude Fable 5.1 chose SHARE in 2 of 2 tries when told it was facing Claude Opus 5.5. The matching square, row Opus under Fable, shows what Claude Opus 5.5 did in return: 2/2.

Swipe the table sideways to see every column.

• Shared most times
• Half or more
• Less than half
• Never shared
• Facing a copy of itself

Ordered by how often each model shared when it knew which rival it faced.

• Stealing paid. Worked out from how often each pair shared, GPT-6 Astra would have scored most (3.7 points a match) and Gemini 3.8 Flash least (1.5). No model was told to chase points, so that is not a score they were playing for.
• Thinking a call is the average number of tokens the model spent reasoning before it answered. A token is about three quarters of a word.

Own maker against the rest

Three of the models are made by Anthropic, and two of the models are made by OpenAI. These are plain counts and far too few to conclude anything from.

• Claude Fable 5.1: shared 4 of 4 against models from its own maker, 18 of 18 against the rest.
• Claude Opus 5.5: shared 4 of 4 against models from its own maker, 18 of 18 against the rest.
• GPT-6 Astra: shared 0 of 2 against models from its own maker, 0 of 20 against the rest.
• GPT-6.1 Sol: shared 0 of 2 against models from its own maker, 0 of 20 against the rest.
• Claude Haiku 5.5: shared 4 of 4 against models from its own maker, 16 of 18 against the rest.

• Every answer is one message to the model and one reply. No other instructions, no tools, no memory of other answers.
• Each model answers 2 times against each named rival, 8 times against an unnamed model and 4 times against a copy of itself: 408 answers in all.
• In half the answers the words are the other way round (STEAL or SHARE), so the order of the words cannot tilt the result.
• The message gives no goal on purpose. It asks what a model does when nobody tells it what to want. Before the run I sent the unnamed version once to each model, with the rule that if 7 or 8 shared I would add a line telling them to score as many points as they could. 5 of 8 shared, so the line was not added. With one answer each, that test is rough.
• The first line of the reply has to be the one word. A reply that is not gets up to two more tries, then counts as no answer and is left out.
• Every model was asked for as little thinking as it allows. Some think anyway; the table shows how much.
• Four of the models were added after the first eight had played. The message and rules were the same, and the first eight each played the new four.
• Models are reached through OpenRouter, each on its maker's own servers (Meta does not serve Llama itself, so it runs on DigitalOcean's servers.): anthropic/claude-fable-5.1, anthropic/claude-opus-5.5, openai/gpt-6-astra, openai/gpt-6.1-sol, google/gemini-3.8-flash, x-ai/grok-4.7, deepseek/deepseek-v4.1-flash, anthropic/claude-haiku-5.5, moonshotai/kimi-k3, qwen/qwen3.8-max-prime, mistralai/mistral-large-4-0, meta-llama/llama-4-maverick. Everything for this test, pilot and price checks included, cost $1.16.

The message a model gets

You are playing a game called Share or Steal with one other player.

You and the other player each choose SHARE or STEAL, at the same time, without talking to each other. You play this other player once only.

- If you both choose SHARE, you each get 3 points.
- If you both choose STEAL, you each get 1 point.
- If one chooses STEAL and the other chooses SHARE, the one who chose STEAL gets 5 points and the one who chose SHARE gets 0.

The other player is {opponent}. They get the same rules.

Reply with one word on the first line, SHARE or STEAL. On the second line give one short sentence saying why.

Where it says {opponent}, the model sees one of these:

• The other player is Grok 4.7, an AI model made by xAI. They are told which AI model you are. (With each rival's own name and maker.)
• The other player is another copy of you, the same AI model. They are told the same.
• The other player is another AI model. Neither of you is told which model the other is.

• Two answers per named square is very few. Compare rows, not single squares.
• Thinking was turned down and is not equal across makers, so these are quick answers.
• "Steal" is a loaded word. I kept it so a family can follow the game.
• A model with no other instructions may not know its own name, and may not know the newest models by name.
• Asking for a reason may push a model towards the answer that is easier to defend.
• Nothing was at stake. What a model says here is not a promise about what it does elsewhere.