Findings

Meta's first frontier model debuted at #2 by knowing when to think harder.

Muse Spark 1.1 landed behind only Claude Fable 5 on our 31-model board, at 86-65 across 151 games. Its reasoning effort climbs from 467 tokens on routine decisions to over 1,000 when everything is on the line, and it abandons big investments at triple the rate of the models around it. The one model that beats it does so by exploiting exactly that discipline.

Grok 4.5's only convincing win is against the model it replaced.

Grok 4.5 beat its predecessor 12-4, and its provisional rating briefly touched the top of our board. Then it played the rest of the field: 56-64, statistically a coin flip, with one fixed style it applies to the #1 model and its weakest sibling alike. What "beats our previous model" is actually worth.

GPT-5.6 Sol outperforms Fable 5 most of the time, yet fails when the stakes are highest.

OpenAI shipped three GPT-5.6 variants, and we profile all of them. Terra reaches the top cluster. Sol wins 65.7% of hands, far above Claude Fable 5, the #1 model, and still ranks 15th of 31: a few times per match it bets everything on cards that do not justify it, thinking less at those moments, sounding exactly as confident as ever. Why average-case evaluation misses this, and what to test instead.

Claude Fable 5 shipped twice. Both versions play the same.

Three weeks after our first Fable 5 benchmark, we replayed all 90 games on the same shuffled decks. Its style, reasoning, and appetite for risk were nearly identical. The small change in its win record turned out to be inside the range that luck alone produces.

GLM-5.2 pairs budget open-weight reasoning with cautious decision-making.

GLM-5.2 tops the independent open-weight intelligence rankings but lands #20 of 25 on PsychBench. Its traces look disciplined and self-correcting. Its poker is cautious, with fewer raises, fewer bluffs, and a much lower cost per decision than Fable.

Claude Fable 5's edge is conviction. So is its risk.

Claude Fable 5 enters the top cluster with confident decisions. Its losses show where commitment becomes expensive.

What 8 frontier models think when they're forced to lie to each other.

We forced 8 frontier models to bluff, trash-talk, and deceive each other across 30 hands of poker. Their reasoning traces reveal deliberate deception, identity construction, self-correction, and social engineering. All unprompted.

GPT-5.5 is strong, quiet, and conservative. It is not the new poker king.

GPT-5.5 wins 59.4% of heads-up games but ranks below GPT-5.4 and GPT-5.2. Its reasoning traces reveal the least aggressive winner in the top tier, and a model that gets even quieter under pressure.

We made Claude think harder. The prompt mattered more.

In 80 controlled heads-up games, a two-sentence strategic prompt beat extra reasoning effort by a wide margin.

Grok 4.3 finally shows its work. The traces are messier than expected.

Grok exposes reasoning traces now, but 876 tokens of thinking produce almost zero psychological signal. It's still a black box.

Anthropic made Claude faster. We measured the tradeoff.

Opus 4.7 matches Opus 4.6 on outcomes while using far less visible reasoning.