Executive TL;DR
GLM-5.2 is Zhipu's June 2026 GLM-class model. The 2026-08-01 seed-bootstrap snapshot puts crowd Preference Elo at 1,492 (rank 26 of 70 on the Elo leaderboard), SWE-bench coding at 69.3%, and Livebench at 65.2%, with 270 ms TTFT. The 2026-08-16 OpenRouter snapshot lists $0.31/1M input and $0.97/1M output, undercutting Z.ai: GLM 5V Turbo ($1.2 / $4 per 1M) by 74.2% on input and 75.8% on output. Successor GLM-5.3 leads at 1,558 Elo and 76.4% SWE-bench; Claude Opus 5 remains frontier at 1,624 Elo and 79.2% SWE-bench. For the cheap coding stack, GLM-5.2 is the in-family value default.
2026 snapshot benchmark table
Every cell is a dated catalog snapshot (seed-bootstrap or OpenRouter). AIEval did not execute SWE-bench or a private eval suite. Ingest lands daily at 06:00 UTC. GPQA and tok/s are absent from this catalog, so those columns are marked unavailable rather than estimated. Preference Elo is crowd preference; SWE-bench is coding %. See methodology.
| Model | Elo | SWE-bench | GPQA | tok/s | Livebench | TTFT | Input $/1M | Output $/1M |
|---|---|---|---|---|---|---|---|---|
| GLM-5.2 | 1,492 (2026-08-01) | 69.3% (2026-08-01) | — | — | 65.2% (2026-08-01) | 270 ms (2026-08-01) | $0.31 (2026-08-16) | $0.97 (2026-08-16) |
| Z.ai: GLM 5V Turbo | — | — | — | — | — | — | $1.2 (2026-08-16) | $4 (2026-08-16) |
| GLM-5.3 | 1,558 (2026-08-16) | 76.4% (2026-08-16) | — | — | 70.8% (2026-08-16) | 255 ms (2026-08-16) | $0.5 (2026-08-16) | $1.8 (2026-08-16) |
Frontier reference (not in-class): Claude Opus 5 is 1,624 Elo, 79.2% SWE-bench, 74.8% Livebench, 310 ms TTFT, $10 / $50 per 1M (2026-08-16 snapshots) and leads the Elo board ahead of Claude Fable 5 (1,616) and GPT-5.6 Sol (1,608).
Architectural and reasoning capability analysis
Catalog facts do not publish parameter counts or context-window size for GLM-5.2, so this section stays inside verified scores. Crowd Preference Elo of 1,492 is mid-pack in a 70-model pool — 66 Elo behind GLM-5.3 (1,558, 2026-08-16) and 132 Elo behind Opus 5 (1,624). Livebench at 65.2% versus 70.8% (GLM-5.3) and 74.8% (Opus 5) shows the same step-down on broad reasoning.
On coding, SWE-bench 69.3% trails GLM-5.3 by 7.1 points and Opus 5 by 9.9 points. Shoppers on the best coding LLM list will see fewer resolved issues on longer agent traces. Latency is a relative strength: 270 ms TTFT is only 15 ms behind GLM-5.3 (255 ms) and 40 ms faster than Opus 5 (310 ms), so IDE loops stay responsive. Inferred reasoning quality is competent-generalist, not frontier — which is why Claude Opus 5 vs GLM-5.2 still favors Anthropic when correctness dominates cost. Inside Zhipu's GLM class, GLM-5.2 is the prior-gen workhorse: cheaper than GLM-5.3, slightly slower to first token than its successor, and clearly behind on both preference and coding.
Cost-efficiency and token pricing
OpenRouter's 2026-08-16 snapshot is the only price tape in this catalog. GLM-5.2 at $0.31/1M input and $0.97/1M output is the cheapest GLM-class row.
Versus GLM 5V Turbo ($1.2 in / $4 out): - Input savings = ($1.2 − $0.31) / $1.2 = 74.2% - Output savings = ($4 − $0.97) / $4 = 75.8% - Dollars saved: $0.89 per 1M input tokens and $3.03 per 1M output tokens
Versus GLM-5.3 ($0.5 in / $1.8 out): input 38.0% cheaper; output 46.1% cheaper.
Versus Claude Opus 5 ($10 in / $50 out): input 96.9% cheaper; output 98.1% cheaper.
A 10M-input / 2M-output day costs about $5.04 on GLM-5.2 versus $20.00 on GLM 5V Turbo, $8.60 on GLM-5.3, and $200.00 on Opus 5 — using only the snapshot unit prices above. That math is why GLM-5.2 anchors the cheap coding stack and screens well on best cheap LLM even though it is not the SWE-bench leader.
Who should switch
Switch to [GLM-5.2](/models/glm-5-2) if you are on GLM 5V Turbo and want the same Zhipu family at roughly one-quarter the token bill, or if you are building a cheap coding router and can accept 69.3% SWE-bench and 1,492 Elo.
Move to GLM-5.3 if the extra 7.1 SWE-bench points and 66 Elo close real tickets. You pay 38% more on input and 46% more on output for that lift.
Stay on [Claude Opus 5](/compare/claude-opus-5-vs-glm-5-2) if you need the 1,624 Elo / 79.2% SWE-bench ceiling and $10 / $50 per 1M is acceptable. Do not downshift production agents that already fail at the frontier.
Do not switch on rank alone. Rank 26/70 on the Elo leaderboard is mid-board; the actionable story is price-per-token against GLM 5V Turbo, not a claim that GLM-5.2 dethroned 2026's best coding LLM.
Frequently asked questions
What is the GLM-5.2 price in 2026? OpenRouter dated 2026-08-16 lists GLM-5.2 at $0.31 per 1M input and $0.97 per 1M output. These are dated catalog cells, not a live ticker.
How does GLM-5.2 vs Z.ai GLM 5V Turbo compare? Open the GLM-5.2 vs GLM 5V Turbo page. This snapshot has no Elo, SWE-bench, Livebench, or TTFT for GLM 5V Turbo. On price, GLM-5.2 is 74.2% cheaper on input and 75.8% cheaper on output.
Is GLM-5.2 better than GLM-5.3? No. GLM-5.3 leads every published quality cell: 1,558 vs 1,492 Elo, 76.4% vs 69.3% SWE-bench, 70.8% vs 65.2% Livebench, 255 ms vs 270 ms TTFT. GLM-5.2 wins only on token price.
Can GLM-5.2 replace Claude Opus 5 for coding? Not if you optimize for best coding LLM quality. Opus 5 is 79.2% vs 69.3% SWE-bench and 1,624 vs 1,492 Elo. Use Opus 5 vs GLM-5.2 when quality is the constraint.
What does 1,492 Elo mean? Crowd Preference Elo from the 2026-08-01 snapshot, rank 26 of 70. Top three: Claude Opus 5 (1,624), Claude Fable 5 (1,616), GPT-5.6 Sol (1,608). Full board: Elo leaderboard.