DeepSeek V4 Pro (0813) vs GLM-5.2

DeepSeek V4 Pro (0813) vs GLM-5.2

DeepSeek V4 Pro has one clean win an independent board confirms: LiveBench (77.4 vs 73.2, both run on LiveBench's own harness), a real gap. LiveCodeBench (87.53 vs 69.5, both vals.ai runs) used to be the second; since 2026-09-29 this site grades that benchmark saturated, so the 18-point spread is filed as tainted — GLM-5.2, a superseded model, sits far below a ceiling most current models have reached, which is not the same evidence as two current models being separated. Long-horizon coding and Agents' Last Exam (62.7, 25.7) are still the vendor's own numbers measured against GLM's independently-run scores (44, 20.4) — call those unverified, not real, until someone reruns DeepSeek's side. Multi-tool chores no longer waits on that: Toolathlon's own board run puts DeepSeek V4 Pro 0813 at 74.4 (replacing the vendor's 74.1, entry dated 2026-08-30) against GLM-5.2's 59.9 on the same Toolathlon harness — a 14.5-point real gap, and this matchup's second independently-confirmed win for DeepSeek. (An earlier version of this page called the Agents' Last Exam number a confirmed real gap; that was wrong for the same reason it was wrong before — corrected 2026-08-19, after 2026-08-17's separate metric-mix-up fix.) Terminal-Bench 2.1 doesn't compare cleanly either: this site's rows are Artificial Analysis's 78.7 for DeepSeek and vals.ai's 67.79 for GLM-5.2, different harnesses, and on vals.ai's own board, which ran both on one harness, GLM-5.2 leads DeepSeek V4 Pro 0813 by 67.79 to 54.68, a 13.1-point gap past the 10.6 band — off this site's tracked rows, but the one same-harness reading that favors GLM. Reasoning is a dead heat: HLE without tools is 41.0 vs 41.1 in Artificial Analysis's own runs. ARC-AGI-2 doesn't add a comparable line either — DeepSeek's tracked score is its Max tier (61.3) while GLM-5.2's is a flat, untiered 22.8, no matched tier to compare. On price the two now sit within about 10% of each other at list rates ($1.32/$3.96 vs $1.40/$4.40), so the discount is no longer a reason to pick either — and most of DeepSeek's apparent benchmark leads here are still unconfirmed vendor numbers. GLM-5.2 isn't actually differentiated by license: DeepSeek V4 Pro ships under the identical MIT terms. (An earlier version of this page called MIT-licensed weights GLM's exclusive edge — that was wrong; corrected 2026-08-20.) What does set GLM-5.2 apart is provenance: 11 of its 12 tracked scores are independently run, versus 8 of 11 for DeepSeek V4 Pro.

DeepSeek V4 Pro (0813) vs GLM-5.2: benchmark by benchmark

Across 14 rows comparing DeepSeek V4 Pro (0813) and GLM-5.2: 2 rows carry a confirmed real gap, 1 row lands inside the noise band, and 11 rows carry other caveats — unverified, tainted, or run on mismatched tool setups. Read each row’s Signal label before ranking DeepSeek V4 Pro (0813) against GLM-5.2.

Who should pick DeepSeek V4 Pro (0813), and who should pick GLM-5.2

  • Most workloads on a budget → DeepSeek V4 Pro (0813) (claims a lead on everything this site tracks — Toolathlon-Verified (74.4 vs 59.9, both Toolathlon-run) is now an independently-confirmed real gap, while the DeepSWE and Agents' Last Exam leads are still unconfirmed vendor numbers, and vals.ai's same-harness Terminal-Bench 2.1 runs favor GLM-5.2; on list price the two are within about 10%, so this pick rests on the benchmarks, not the bill)
  • Whoever wants the more independently-checked scorecard → GLM-5.2 (11 of GLM-5.2's 12 tracked scores are independent runs, versus 8 of 11 for DeepSeek V4 Pro — not a license edge (both are MIT open weights), but a provenance one)
  • Composite everyday benchmark → DeepSeek V4 Pro (0813) (LiveBench 77.4 vs 73.2 and Toolathlon-Verified 74.4 vs 59.9, both pairs run on one harness each — the two confirmed real gaps left in this matchup now that LiveCodeBench (87.53 vs 69.5) is graded saturated and no longer called)

DeepSeek V4 Pro (0813) vs GLM-5.2 pricing

Input / 1M tokens$1.32$1.40
Output / 1M tokens$3.96$4.40

DeepSeek V4 Pro (0813) vs GLM-5.2 FAQ

Is DeepSeek V4 Pro (0813) better than GLM-5.2?

DeepSeek V4 Pro has one clean win an independent board confirms: LiveBench (77.4 vs 73.2, both run on LiveBench's own harness), a real gap. LiveCodeBench (87.53 vs 69.5, both vals.ai runs) used to be the second; since 2026-09-29 this site grades that benchmark saturated, so the 18-point spread is filed as tainted — GLM-5.2, a superseded model, sits far below a ceiling most current models have reached, which is not the same evidence as two current models being separated. Long-horizon coding and Agents' Last Exam (62.7, 25.7) are still the vendor's own numbers measured against GLM's independently-run scores (44, 20.4) — call those unverified, not real, until someone reruns DeepSeek's side. Multi-tool chores no longer waits on that: Toolathlon's own board run puts DeepSeek V4 Pro 0813 at 74.4 (replacing the vendor's 74.1, entry dated 2026-08-30) against GLM-5.2's 59.9 on the same Toolathlon harness — a 14.5-point real gap, and this matchup's second independently-confirmed win for DeepSeek. (An earlier version of this page called the Agents' Last Exam number a confirmed real gap; that was wrong for the same reason it was wrong before — corrected 2026-08-19, after 2026-08-17's separate metric-mix-up fix.) Terminal-Bench 2.1 doesn't compare cleanly either: this site's rows are Artificial Analysis's 78.7 for DeepSeek and vals.ai's 67.79 for GLM-5.2, different harnesses, and on vals.ai's own board, which ran both on one harness, GLM-5.2 leads DeepSeek V4 Pro 0813 by 67.79 to 54.68, a 13.1-point gap past the 10.6 band — off this site's tracked rows, but the one same-harness reading that favors GLM. Reasoning is a dead heat: HLE without tools is 41.0 vs 41.1 in Artificial Analysis's own runs. ARC-AGI-2 doesn't add a comparable line either — DeepSeek's tracked score is its Max tier (61.3) while GLM-5.2's is a flat, untiered 22.8, no matched tier to compare. On price the two now sit within about 10% of each other at list rates ($1.32/$3.96 vs $1.40/$4.40), so the discount is no longer a reason to pick either — and most of DeepSeek's apparent benchmark leads here are still unconfirmed vendor numbers. GLM-5.2 isn't actually differentiated by license: DeepSeek V4 Pro ships under the identical MIT terms. (An earlier version of this page called MIT-licensed weights GLM's exclusive edge — that was wrong; corrected 2026-08-20.) What does set GLM-5.2 apart is provenance: 11 of its 12 tracked scores are independently run, versus 8 of 11 for DeepSeek V4 Pro.

Which is cheaper, DeepSeek V4 Pro (0813) or GLM-5.2?

DeepSeek V4 Pro (0813) costs less per output token ($3.96 vs $4.40 per 1M tokens, official listed rates).

Are both models open weight?

Yes — both publish MIT-licensed open weights, so licensing separates nothing here. What does: provenance and one confirmed real gap. GLM-5.2 has 11 of 12 tracked scores independently run versus 8 of 11 for DeepSeek V4 Pro, and DeepSeek V4 Pro leads LiveBench 77.4 vs 73.2 on LiveBench's own harness.

Which is better at agent work, DeepSeek V4 Pro or GLM-5.2?

Split, with a caveat. DeepSeek V4 Pro's Toolathlon-Verified 74.4 vs 59.9 is an independently-confirmed real gap, while vals.ai's single-harness Terminal-Bench 2.1 runs favor GLM-5.2. DeepSeek's DeepSWE and Agents' Last Exam leads are still the vendor's own numbers measured against GLM's independent runs — unverified, not real, until someone reruns DeepSeek's side.

What happened to the LiveCodeBench comparison?

It used to be DeepSeek V4 Pro's second confirmed win (87.53 vs 69.5, both vals.ai runs). Since 2026-09-29 this site grades LiveCodeBench saturated, so the 18-point spread is filed as tainted: GLM-5.2 is a superseded model sitting far below a ceiling most current models have reached, which is not the same evidence as two current models being separated.

Is DeepSeek V4 Pro still current?

Yes. DeepSeek's V4.1 Flash replaced the smaller V4 Flash, not the Pro tier, and this site computes no successor for V4 Pro. GLM-5.2 is superseded — GLM-5.3 is its computed successor — which is one reason the LiveCodeBench spread reads as tainted rather than as a win.