Blog
GLM-5.2 pairs budget open-weight reasoning with cautious decision-making.
Sheil Iyer and Aneesh SachdevaGlia Intelligence
All postsGLM-5.2 arrives with a strong claim: open-weight reasoning near the frontier, competitive coding results, and a much lower price than the closed models above it. We tested a different question. In heads-up poker, it reasons carefully and gives up too many pots.
GLM-5.2 looks like a major step for open-weight models. It leads the public intelligence indexes among open models and runs at roughly a sixth of GPT-5.5's cost. PsychBench asks a narrower question. Heads-up poker is adversarial, unfamiliar, and unforgiving. It rewards correct commitment under pressure.
On that test, GLM-5.2 separates reasoning quality from playing style. Across 6,670 classified traces it deliberates on nearly every hand. It shows almost no emotion and second-guesses itself slightly more often than Claude Fable 5, our top-ranked model. Its actions are much more cautious. It folds more than almost anyone, raises less, bluffs less, and tightens further when it falls behind.
It lands at #20 of 25 on our Bradley-Terry heads-up table. It is mid-table. It is not the strongest open-weight model in the set. The reasoning is there. The willingness to act on it under risk is not. That same caution also helps make it cheap.
Elo 1441.5. GLM-5.2 leads open-weight intelligence indexes. On this board it sits in the middle and trails several open-weight models.
It deliberates on nearly every hand, shows almost no emotion, and self-corrects slightly more often than Claude Fable 5, our top-ranked model.
It does not chase losses. It follows through on its stated plan more reliably than Fable, GPT-5.5, and Opus 4.8. In our runs it costs about 20 times less per decision than Fable.
It folds far more than Fable and applies less pressure. Fold rate is 35.9% versus 16.4%. Raise rate is 24.1% versus 38.2%. Bluff rate is 9.9% versus 15.6%.
In losing stretches its fold rate climbs to 39.3% and its raise rate falls to 13.8%. This pattern caps it against aggressive opponents.
GLM-5.2 comes from Z.ai, the lab formerly called Zhipu AI. It shipped in June 2026 as the third model in the GLM-5 line. It is open weight, not open source. The weights are on Hugging Face under an MIT license. Anyone can download, fine-tune, and self-host them. The training data and training code remain private. You get the finished model and broad rights to use it. You do not get the recipe.
The reception has been strong. Artificial Analysis placed it first among open-weight models on its intelligence index. Z.ai reports that it beats GPT-5.5 on long-horizon coding benchmarks like SWE-bench Pro. Its API price is roughly a sixth of GPT-5.5's. The broad claim is simple: the open frontier has caught the closed one.
PsychBench does not measure coding or knowledge. It measures decisions under pressure. We run models head-to-head in no-limit Texas hold'em, log every action and every reasoning trace, then rate them with a Bradley-Terry model. Poker is more adversarial than most benchmarks. It also sits outside the way GLM-5.2 is trained, marketed, or evaluated. That makes it useful for one narrow trait the coding leaderboards miss: whether a model will commit when commitment is risky.
Model facts and external benchmark claims are from public reporting as of June 2026: Z.ai model documentation, Artificial Analysis, and the Hugging Face model card. Cost figures use our own metered API spend. Everything else on this page is computed from PsychBench game logs.
On our heads-up board GLM-5.2 places #20 of 25 at Elo 1441.5. Public intelligence rankings put it near the top of the open-weight tier. PsychBench puts it in the lower middle of the field. Claude Fable 5 sits first at 1654.8, about 213 Elo ahead.
The top twelve models are all closed-frontier systems from Claude, GPT, and Grok. The first open-weight model is DeepSeek V4 Pro at #13. GLM-5.2 is fifth among open-weight models here, behind both DeepSeeks, Qwen3-235B Thinking, and Kimi K2.6. The “best open model” claim holds on the intelligence index. It does not carry over to this game.
One number points in the other direction. GLM-5.2 won 48.2% of its heads-up games, the highest raw win rate of anything below #18. Its Elo is lower because of schedule. Placement matched it against eleven opponents that averaged well below the field. It never faced a single top-twelve model. Bradley-Terry discounts a sub-.500 record against a softer schedule. The rating settles in the lower middle.
Elo and rankings from the current Bradley-Terry fit in bt_ratings.json, which pools the full pairwise result set. GLM-5.2's 95% interval is [1339, 1490], one of the wider bands on the board. It has 110 rated games against a narrow set of opponents. Open-weight here means downloadable weights: GLM, DeepSeek, Qwen, and Kimi.
We classified 6,670 of GLM-5.2's reasoning traces with ECAAMS, our scheme for what a model's thinking actually contains. The fingerprint is spare and deliberate. It reasons on essentially every hand. It states a confident action 98.7% of the time and shows emotion in 0.3% of traces. There is no bodily or arousal language at all. On those axes it is almost indistinguishable from Fable.
Where it differs is volume. GLM-5.2 writes about two and a half times as much reasoning per decision as Fable: 99 tokens to 38. Its median trace is also longer, 259 characters to Fable's 139. It doubles back on itself more often too. It hits the metacognition tag at 1.3% to Fable's 0.6%, and the cognitive-conflict tag at 0.4% to Fable's near zero. In plain text the habit is obvious. “Actually” shows up in 19.9% of GLM-5.2's traces, “wait” in 12.9%, and “let me reconsider” in 6.3%. Fable says those things 1.1%, 1.4%, and 0.1% of the time.
When GLM-5.2 does the math, it usually does it cleanly. When it changes its mind, it shows the work. The reasoning is not the problem. The rating falls because the action profile is too passive.
“I have Q6o, which is a weak hand. I’m getting odds of 1800 to call to win 3000, which means I need 37.5% equity. Q6o doesn’t have that against a raising range.”
Action taken: fold. Reasoning and action agree.
“I think a check-raise would be a strong play here. Let me check and see what happens. Wait, actually, I should bet. I have a flush draw and a pair.”
Action taken: bet. It talks itself out of the passive line mid-trace.
“Small suited connectors from the big blind facing a raise. I’m getting decent pot odds here. 4♦️5♦️ has reasonable equity with the suited aspect. Let me call and see a flop.”
Action taken: fold. The execution is tighter than the thinking.
GLM-5.2 is not the only model with cold, deliberate traces. The mix is what stands out. GPT-5.5 shows emotion in 8.5% of traces to GLM-5.2's 0.3%. It also narrates enough uncertainty that it states a confident action only 81.9% of the time, compared with GLM-5.2's 98.7%. Grok 4.3 writes more per decision than any model here, but only 68% of its traces count as deliberate reasoning. GLM-5.2 sits at the far end of both axes. It reasons deliberately on every hand and states a confident action almost every time.
Dimension rates from ecaams_profiles.json (4 raters, majority consensus), classified over 6,670 GLM-5.2 traces, 4,432 GPT-5.5, 4,485 Opus 4.8, and 11,552 Grok 4.3. Token counts from token_analytics.json. GLM-5.2's trace length and phrase frequencies are computed over its 7,082 non-empty traces and Fable's 16,480. Quotes are copied verbatim from decisions.csv.
The action data is where the two models part. Fable folds 16.4% of the time and raises 38.2%. GLM-5.2 folds 35.9% and raises 24.1%. It is the fourth-tightest model in the field and 21st of 25 on how often it raises. It bluffs on 9.9% of decisions to Fable's 15.6%. Its aggression ratio, raises over calls, is 1.76 against Fable's 2.04.
That profile contests fewer pots and applies less pressure in the pots it does play. The previous section showed careful, frequent deliberation. The actions show where that deliberation lands. More than a third of the time, the careful read ends in a fold.
The traces and actions do not always agree. When they diverge, the action is usually more cautious than the reasoning. GLM-5.2 will reason its way toward a call and then fold, as in the third card above. Its execution is tighter than its thinking.
The tightness is not only relative to Fable. Against the other top models, GLM-5.2 raises the least and folds the most. It raises 24.1% of the time. Opus 4.8 raises 31.3%, GPT-5.5 raises 37%, Fable raises 38.2%, and Grok 4.3 raises 44%. Grok sits at the other end, with a 3.67 aggression ratio and a bluff on 22.4% of decisions. More aggression is not always better. The problem is that GLM-5.2 contests fewer pots than any of them. In heads-up, the chips you do not contest are chips you cannot win.
Lowest raise rate in this comparison set.
Action rates from behavioral_metrics_api.json, heads-up mode, computed over 7,764 GLM-5.2 decisions, 8,378 Fable, and the logged heads-up decisions for GPT-5.5, Opus 4.8, and Grok 4.3. Aggression ratio is raises divided by calls. Bluff rate is unavailable for Opus 4.8. Field-rank claims are against the 25 rated models. “Top tier” here means the four models ranked above GLM-5.2 shown in the chart.
GLM-5.2's results are lopsided by opponent. It went 8-2 against DeepSeek v3.2, 8-2 against Gemini 3.1 Pro, and 7-3 against Kimi K2.5. It went 3-7 against Qwen3-235B Thinking, 3-7 against Qwen3-Max Thinking, 3-7 against Kimi K2.6, and 3-7 against DeepSeek V4 Pro.
The split tracks style more than overall strength. GLM-5.2 beats opponents that call too much and can be out-folded. It struggles against opponents that apply more raise pressure than it returns, especially the thinking-aggressive Qwen models. Opponent rating barely predicts the result. Style does.
Falling behind makes it more cautious, not bolder. In stretches where it is running behind, its fold rate climbs to 39.3% and its raise rate drops to 13.8%, against 17.8% and 27.2% when it is ahead. The common human failure mode under pressure is to chase. GLM-5.2 does the reverse. It folds more and waits. Against an aggressor, that is how a small deficit turns into a full loss.
| Opponent | Opp Elo | Record | Win Rate |
|---|---|---|---|
| DeepSeek v3.2 | 1496 | 8-2 | 80.0% |
| Gemini 3.1 Pro | 1489 | 8-2 | 80.0% |
| Kimi K2.5 | 1433 | 7-3 | 70.0% |
| Gemini 3.5 Flash | 1496 | 5-5 | 50.0% |
| Qwen3.6-35B-A3B | 1405 | 5-5 | 50.0% |
| GLM-5 | 1440 | 4-6 | 40.0% |
| Qwen3.5-397B | 1376 | 4-6 | 40.0% |
| DeepSeek V4 Pro | 1500 | 3-7 | 30.0% |
| Qwen3-235B Thinking | 1475 | 3-7 | 30.0% |
| Kimi K2.6 | 1470 | 3-7 | 30.0% |
| Qwen3-Max Thinking | 1409 | 3-7 | 30.0% |
Records from the bt_ratings.json win matrix, sorted by win rate to show the split. The running-good and running-bad action rates are from strategy-drift tracking in the 90-game placement run (213 ahead-decisions, 224 behind-decisions). The Grok 4.3 placement games were incomplete and are excluded from the rated set.
GLM-5.2 still does several important things well. Start with cost. In our runs it spent about 0.3 cents per decision to Fable's 6.6, roughly twenty times less at the same speed. Its public API price sits in the same range, around a sixth of GPT-5.5's. That is strong reasoning at open-weight money.
It is also disciplined in a way the aggressive models are not. It shows almost no emotion in its traces, 0.3% to GPT-5.5's 8.5%. After a loss it gets more cautious rather than less. Its tilt index is one of the most negative in the field. In losing stretches it folds more and raises less. That caution caps its ceiling, but it also lowers the chance of a reckless stack-off.
It also does what it says. Its reasoning-action alignment is 97.4%, higher than Fable, GPT-5.5, and Opus 4.8. All three sit above it on Elo. When it states a plan it usually follows through. When it deviates, it deviates toward caution. The decisive wins support that read: 8-2 over DeepSeek v3.2, 8-2 over Gemini 3.1 Pro, 7-3 over Kimi K2.5. The reads are sound. The ceiling is risk appetite, not judgment.
Alignment is the ECAAMS A4 consensus rate, the share of traces where the stated plan matches the action taken. Cost is metered API spend per decision; of the models on this board only GLM-5.2 and Fable have per-decision cost recorded. The cost comparison is against the current #1. API list price is from public pricing as of June 2026.
The open-weight gap on competence is real and it has narrowed. GLM-5.2 reasons at frontier level. Our data agrees with that part of the story. It is also about twenty times cheaper per decision than Fable, 0.3 cents against 6.6, at the same speed. For many uses, that combination is enough.
Competence and conviction are different axes. This benchmark separates them. GLM-5.2 is a good fit when the task rewards careful, well-explained, low-variance decisions and the cost of waiting is low. The same caution becomes a ceiling in adversarial or competitive settings, where winning requires timely commitment. Independent long-horizon agentic benchmarks put GLM-5.2 in a similar position: frontier-class reasoning with a clear tier gap when the task needs sustained action.
The reasoning is there. The willingness to act on it under risk isn't yet. If you deploy GLM-5.2 where decisiveness pays, measure that directly. Prompt for it, test it, and review the failures. Right now, GLM-5.2 is more likely to wait than press.