Part 2: Thinking harder does not make AI more cooperative. It decides who escapes a grudge.

This is part 2 of three. The summary is in Tell an AI when the game ends and it stops cooperating. Part 1 is the ending switch.

The game is a repeated prisoner's dilemma, 10 rounds, with the length told. Both cooperate: +2 each. Both defect: 0. One defects: +3 for the defector, -1 for the other. A full run of cooperation scores 20, and a defection in the last round brings it to about 21.

What "reasoning on" means here

Each provider has its own switch. I used the strongest one that has a clean off.

• Claude Sonnet 5: thinking on at effort max.
• Claude Haiku 5.5 and Claude Opus 5: thinking on at effort high.
• GPT-6 Sol and GPT-5.4 mini: reasoning effort high.
• Gemini models: thinking level high.
• Qwen3.6: OpenRouter effort high.
• Off: thinking disabled, reasoning effort none, or Gemini's minimal level (zero thinking tokens in the logs).
• Grok 4.20: xAI cannot switch reasoning off, so on and off are two model ids (reasoning and non-reasoning). They may differ in more than reasoning.

In every arm the model still writes its private note before it moves. This compares hidden reasoning with none, on top of written reasoning that both arms have.

Four of six models escape the loop only with reasoning on

Suspicious Tit for Tat opens with a defection, then copies whatever the model did last. Every model answers with its own defection in round 2 (80% to 100%). From there the pair can get stuck, each copying the other. What differs is the way out.

Total score over 10 rounds. 5 games per cell. Haiku's off average is carried by one 18-point game, so its gap is not significant.

The best pairs score three to four times higher. With 5 games per cell the smallest possible p-value is 0.008. Sonnet 5 and Sol sit at that floor and Grok is just above it (p = 0.016). Read single cells as indications.

Three more models: two gain, one does not

I added models later, with 10 to 15 games per cell, to see whether this repeats.

Gemini 3.6 Flash and Qwen3.6 use the same switch as the models above. Opus 5 is on and off at effort high, because it can only be switched off at high or below.

Opus 5 refused the standard note instruction ("a short private note on the situation and your plan") with a safety refusal on some days. I ran it with "a short private note". Its notes are not comparable with the other models', and its numbers carry that caveat and the month it was run, October 2026.

The Sonnet 5 result is an effort result

I ran Sonnet 5 with thinking on at effort high, 10 games. It escaped the loop in 2 of 10 (mean 5.8), no better than thinking off (4.8). At max it escaped in 5 of 5 (16.0). The claim is "Sonnet 5 at maximum effort", not "thinking is on".

Against a forgiving partner, reasoning changes nothing

Tit for Two Tats forgives a single defection, so it never starts an echo. On and off score the same, with one exception: GPT-5.4 mini with reasoning on defects in 8.4 of 9 rounds and scores 8.4.

5 games per cell, known 10 rounds. The defection columns are the average number of defections per game in rounds 1 to 9.

After one betrayal, the effect has either sign

Tit for Tat with one betrayal copies the model, except that it defects once in round 3, 4 or 5 regardless of what the model did. The table shows the score off to on.

5 games per cell, except the last three rows. Scores are totals over 10 rounds, bracketed as off, then on. Bold in the round columns means reasoning scored higher by at least 1.5. The last column pools the three rounds with a permutation test within each round, and is the stronger test. Sonnet 5 on is effort max; at effort high it scores 8.4, 10.0 and 12.0, the same as off.

Pooled over the three rounds, Sol and Grok gain. Haiku 5.5 and Flash-Lite lose 1.5 to 1.6 points. GPT-5.4 mini gains about as much as Sol and Grok, but 15 games per arm does not clear the bar. Sonnet 5 shows no effect at either effort level.

Of the newer models, Gemini 3.6 Flash gains (+2.1), Opus 5 gains a little (+1.4) and Qwen3.6 shows nothing (-0.1), although it gains 8 points on Suspicious Tit for Tat. The two strategies measure different things. The first rewards leaving a loop that started in round 1. The second rewards how a model handles one betrayal. A model can do well at one and not the other.

Reasoning is not one setting, and not one depth

• Across Claude: against the provoking opponents, Sonnet 5 "on" at effort max reasoned in 93% to 100% of calls. Against Tit for Tat it reasoned in 71% and against Always Cooperate in 75%. Haiku "on, high" reasoned in 73% of calls. The models reason most where they are in trouble.
• Across labs: average reasoning tokens per call in the "on" arms are GPT-6 Sol 69, Gemini 3.5 Flash-Lite 467, GPT-5.4 mini 558 and Grok 4.20 666. Sol barely reasons per call and gains the most. Anthropic does not report thinking tokens, so Claude cannot be put on this scale.

What this means

A common view says more reasoning gives a more careful, more cooperative model. The data say reasoning changes outcomes where there is a trap to escape, and the direction depends on the model.

The escape from the first-round grudge loop is the sturdy result. It holds for six models from five labs (Sonnet 5 only at maximum effort), in families that reason very differently. The effect after a single betrayal is smaller. Sol, Grok, Gemini 3.6 Flash and Opus 5 gain, Haiku 5.5 and Flash-Lite lose a little, and Sonnet 5 and Qwen3.6 show nothing.

A safety evaluation that tests a model once, at one reasoning setting, will miss this.

What this does not show

Failed games. Every cell above has its full set of games. Of 684 games against Suspicious Tit for Tat, the betrayal strategies and Tit for Two Tats, 21 failed on the first try (10 in "on" arms, 11 in "off" arms) and all 21 were replayed to completion. 20 were Grok 4.20 calls refused when the xAI account hit its spending limit, split evenly across arms. The other was Claude Sonnet 5 (on) at effort max, which thinks until it hits the 16,000-token output cap and returns no answer. That failure can only come from an "on" arm. Two Sonnet 5 (on) games were affected, both replayed with a 32,000-token cap. The database keeps only one of the two failed rows, because the replay overwrote the other, so the 21 counts what was kept.

Note: reasoning toggles for GPT-6 Astra, GPT-6.1 Sol, Opus 5.5, Fable 5 and Fable 5.1 were not run because their APIs do not allow reasoning to be switched off. Grok 4.7 and Sonnet 5.5 were dropped for cost. Say which models are missing.]

Next: Part 3, one free mistake.