Zhipu AI (Z.ai)

GLM-5.3

The launch chart claims frontier coding at an unchanged mid-tier price. The price is real — but nine of the eleven benchmark results here have been independently run (Terminal-Bench 2.1, Humanity's Last Exam without tools, GPQA Diamond, LiveCodeBench, SWE-bench Verified, LiveBench, DeepSWE, Terminal-Bench 4.0, and AA-AnalystAgent at 38.75 pass^5), though GPQA Diamond and SWE-bench Verified are graded saturated here; the with-tools Humanity's Last Exam and Toolathlon-Verified scores are still Z.ai's own runs.

GLM-5.3 benchmarks and pricing, every number sourced: 11 tracked GLM-5.3 benchmark scores (9 independently run, 2 still resting on a vendor’s own claim), priced at $1.40 per million input tokens and $4.40 per million output.

Released
2026-08-14
License
open-weights
Context window
1M tokens
Knowledge cutoff
Not disclosed
Verified
sources checked 2026-08-19–2026-10-15
Succeeded
GLM-5.2 (2026-06-16)
Parameters
Not disclosed
Architecture
Not disclosed

GLM-5.3’s verified record

GLM-5.3’s most-compared rival is Kimi K3: 2 trails, 1 lead, 2 ties, and 6 not callable — no overall winner across their 11 shared comparisons. GLM-5.3 is priced at $1.40/$4.40 per 1M tokens (in/out) vs Kimi K3’s $3.00/$15.00.

Against the 291 head-to-head comparisons GLM-5.3 shares with other tracked models:

71 real gaps52 inside the band168 not called

A gap counts for GLM-5.3 only where both sides were run independently and the benchmark still separates models. Losses are listed alongside wins on purpose — GLM-5.3 trails on 43 of them.

HLE · no tools Reasoning

±2 is noise
Behind
Claude Haiku 5.5 −2.1 · Step 5 Preview −4.2 · Kimi K3 −4.6 · Gemini 3.1 Pro Preview −4.7 · MiniMax M3.1 Flash −5.7 · Muse Spark 1.3 −6.4 · MiMo-V2.6-Pro −7.1 · Muse Wonder −7.8 · Kimi K3.1 Pro −10.2 · GPT-6.1 Sol −10.6 · GPT-6 Astra −12.4 · Claude Sonnet 5.5 −12.7 · Claude Opus 5.5 −19.1 · 6 previous: GPT-6 Sol −5.6 · Claude Opus 4.8 −6.4 · GPT-5.6 Sol −7.2 · Claude Opus 5 −12.6 · Claude Fable 5 −13.2 · Claude Fable 5.1 −16.8
Ahead
GLM-5.3-Flash +2.4 · MiniMax M3 +3.3 · GPT-6 Luna +3.8 · Qwen3.8-Flash-Next +4.3 · Mistral Large 4 +7.3 · Beam 1.0 +7.6 · Nemotron 3.5 Super VL +11.6 · 2 previous: GPT-5.6 Luna +2.8 · DeepSeek V4 Flash (0731) +3.7
Tie
5 models within ±2
Unverified
1 model — vendor-reported on one side
Setup-dependent
7 models — scored on a different harness or effort tier

LiveBench Composite score across 7 domains

±2.7 is noise
Behind
Kimi K3 −3.1 · DeepSeek V4.1 Flash −5.0 · Muse Spark 1.3 −5.5 · GPT-6.1 Sol −5.5 · Claude Opus 5.5 −7.1 · 6 previous: Gemini 3.7 Flash −2.7 · GPT-6 Sol −3.2 · Claude Opus 5 −4.0 · GPT-5.6 Sol −4.9 · Claude Fable 5 −6.9 · Claude Fable 5.1 −7.3
Ahead
Claude Haiku 5.5 +4.0 · GPT-6 Luna +4.1 · GLM-5.3-Flash +4.5 · MiniMax M3 +8.8 · 1 previous: GLM-5.2 +2.9
Tie
13 models within ±2.7

Terminal-Bench 4.0 · max Terminal ops

±12.4 is noise
Behind
GPT-6.1 Sol −16.2 · GPT-6 Astra −20.7 · Claude Sonnet 5.5 −25.3 · Claude Opus 5.5 −26.3 · 2 previous: Claude Opus 5 −14.6 · Claude Fable 5.1 −19.2
Ahead
Muse Spark 1.3 +14.1 · Kimi K3 +21.7 · DeepSeek V4 Pro (0813) +24.8 · GPT-6 Luna +25.3 · 3 previous: Claude Opus 4.8 +15.7 · GPT-5.6 Luna +27.3 · Claude Sonnet 5 +29.3
Tie
4 models within ±12.4
Unverified
1 model — vendor-reported on one side
Setup-dependent
1 model — scored on a different harness or effort tier

AnalystAgent Spreadsheet & document analysis

±11.2 is noise
Behind
GPT-6.1 Sol −11.3 · GPT-6 Astra −12.5 · Claude Haiku 5.5 −15.0 · Claude Opus 5.5 −17.5 · Claude Sonnet 5.5 −18.8 · 2 previous: Claude Opus 5 −15.0 · Claude Fable 5.1 −18.8
Ahead
GLM-5.3-Flash +13.8 · MiniMax M3 +28.8
Tie
13 models within ±11.2
Setup-dependent
5 models — scored on a different harness or effort tier

DeepSWE Long-horizon coding

±9.5 is noise
Ahead
Qwen3.8-Max +12.0 · 4 previous: Claude Opus 4.8 +10.0 · Claude Sonnet 5 +15.0 · DeepSeek V4 Flash (0731) +16.0 · GLM-5.2 +25.0
Tie
7 models within ±9.5
Unverified
19 models — vendor-reported on one side
Setup-dependent
6 models — scored on a different harness or effort tier

No verdict for GLM-5.3 anywhere on Terminal-Bench 2.1 ( every independently confirmed comparison inside the noise band); GPQA Diamond, LiveCodeBench, SWE-bench Verified ( saturated); HLE · with tools, Toolathlon-Verified ( nothing independently confirmed on both sides).

What changed from GLM-5.2 to GLM-5.3

The 8 benchmarks both models have been scored on, using the same variant each time. A raw GLM-5.3 gain is not a real gain until it clears that benchmark’s own noise band, so each row below carries the verdict and not just the arithmetic.

GLM-5.3 versus GLM-5.2, per-benchmark change and whether it clears the noise band
BenchmarkGLM-5.2GLM-5.3ChangeVerdict
HLE41.142.3+1.2Tie
Terminal-Bench 2.167.7983.9+16.1Setup-dependent
DeepSWE4469+25.0Real gap
Toolathlon-Verified59.973+13.1Unverified
SWE-bench Verified82.895.4+12.6Tainted
GPQA Diamond85.6191.7+6.1Tainted
LiveCodeBench69.580.53+11.0Tainted
LiveBench73.276.1+2.9Real gap

GLM-5.3 API pricing

$1.40 in / $4.40 out per 1M tokens — official pricing source

What GLM-5.3 costs per job

GLM-5.3 cost for three reference workloads, computed from its list rates
WorkloadTokens in / outCost
One long chat turn100K / 10K$0.184
A codebase review1,000K / 100K$1.84
A day of agent work10,000K / 1,000K$18.40

Computed from GLM-5.3’s list rates above — cache discounts and batch tiers are not applied.

GLM-5.3 is one of 3 Zhipu AI (Z.ai) models tracked on this site, at these official list prices.

Zhipu AI (Z.ai) model pricing, official list rates
ModelIn / 1MOut / 1M
GLM-5.3-Flash$0.15$0.50
GLM-5.3$1.40$4.40
GLM-5.2(previous version)$1.40$4.40

GLM-5.3 benchmark scores

GLM-5.3 benchmark scores, provenance, and source links
BenchmarkScore
Terminal-Bench 2.1[1]
Terminal ops · ±10.6 is noise
DeepSWE[2]
Long-horizon coding · ±9.5 is noise
Toolathlon-Verified[3]
Multi-tool chores · ±9.7 is noise
HLE(with tools)[4]
Reasoning · ±2 is noise
HLE(no tools)[5]
Reasoning · ±2 is noise
GPQA Diamondsaturated[6]
Expert science Q&A — not ranked at any gap size
LiveCodeBenchsaturated[7]
Contest coding — not ranked at any gap size
SWE-bench Verifiedsaturated[8]
Bug fixing — not ranked at any gap size
LiveBench[9]
Composite score across 7 domains · ±2.7 is noise
Terminal-Bench 4.0(max)[10]
Terminal ops · ±12.4 is noise
AnalystAgent[11]
Spreadsheet & document analysis · ±11.2 is noise

Who ran these numbers: 9 of 11 independent — artificialanalysis.ai (4), vals.ai (3), deepswe.datacurve.ai (1), livebench.ai (1); vendor self-reported (2).

  1. Terminal-Bench 2.1: AA's own run of 'GLM-5.3 (max)' (Terminus 2 harness, e2b sandbox, pass@1 avg of 3 repeats). Replaces Z.ai's self-reported 88.2 from the launch chart. The 4.3-point gap sits inside this benchmark's 10.6-point noise band, so it is not evidence of a bad number — but across the 20 vendor-vs-independent pairs measured on this site, every gap of 4 points or more has run vendor-high, while the largest gap the other way is 3.4. (That count predates 2026-10-01, when vals.ai rows replaced seven more vendor figures, all of them vendor-high by 7.4 to 30.3 points.) Cross-check (2026-10-01): vals.ai's archived Terminus 2 table lists 'GLM 5.3' at 71.54. Z.ai's own card reports 88.2 (Claude Code 2.1.207, 6h timeout) — the vendor figure this cell's AA run replaced; not stored as a row.
  2. DeepSWE: deepswe.datacurve.ai v1.1 leaderboard, rank #4 of 18 tracked models as of 2026-08-26 (69% ± 3%, avg cost $3.99/task, 80k output tokens, 124 steps, max effort). This is now on the board — Z.ai's own launch-chart self-report was 66.9%, so the independent run actually landed 2.1 points ABOVE the vendor's own number, the reverse of this site's more common vendor-favorable pattern; the gap sits well inside this benchmark's ±9.5 cross-model noise band either way, so it isn't evidence of anything more than normal run-to-run variance. Z.ai's card reports 66.9 (mini-swe-agent, 400K context) — not stored as a row; this cell holds the Datacurve board's 69.
  3. Toolathlon-Verified: Z.ai says results came via Toolathlon's official evaluation service (pass@1, avg of 3 runs), but the number is vendor-published — flips to independent when it appears on toolathlon.xyz's own board. Sits above GLM-5.2's independent 59.9 but below Kimi K3's independent 76.5.
  4. HLE: Z.ai's own run: temp 1.0, top_p 0.95, 300K max context with context management, GPT-5.6-luna (medium) as judge. Every model's with_tools HLE number on this site is vendor-reported — no independent board runs this variant.
  5. HLE: AA's own run, 'GLM-5.3 (max)' (text-only, no tools). Z.ai never published a no-tools HLE figure — only the 62.5 with-tools number — so this is the first look at the model without scaffolding.
  6. GPQA Diamond: AA's own run of 'GLM-5.3 (max)'. Every other GPQA row on this site comes from vals.ai; this one carries a different harness (artificial-analysis) recorded for completeness, though it makes no difference to the verdict here: GPQA Diamond is graded saturated on this site (2026-08-20), so compareScores() marks any pairing on this benchmark ⊘ Tainted before a harness check is ever reached — this row exists for the record, not for ranking.
  7. LiveCodeBench: vals.ai run, rank 66/140 (board updated 2026-08-19) — required clicking past the top-16 default view to surface. Cost shown alongside ($1.4/$4.4 in/out) matches this site's own tracked GLM-5.3 pricing, cross-confirming the row is the current 0813-era model, not an older GLM revision.
  8. SWE-bench Verified: vals.ai run, rank 6/86 shown, Mini-SWE-agent harness (board updated 2026-08-19). SWE-bench Verified is graded saturated on this site, so this row completes GLM-5.3's tracked record here without settling any head-to-head comparison.
  9. LiveBench: Board row "GLM-5.3" on the 2026-06-25 LiveBench release — missed in this site's 2026-08-24 scrape (model released 2026-08-14, before that scrape date) and backfilled 2026-08-25 after a data-QA review flagged the gap. Distinct row from GLM-5.2 (73.2); no version confusion.
  10. Terminal-Bench 4.0: vals.ai board row — mini-swe-agent harness (single bash tool), pass@1 averaged over 3 full passes, raw value 38.889 at effort max. Artificial Analysis same-conditions run (Max): 41.92. Official board (grant-funded, Claude Code, max effort, 5-trial pass@1): 41.82 ± 3.23 CI — 2.9 points above this run.
  11. AnalystAgent: AA's own run, board row 'GLM-5.3 (Max)': analystAgent 0.3875 = 38.75 pass^5. pass@1 is 55.25. Read from the GLM-5.3 (Max) page on 2026-10-15.

Notes on the record

API pricing went live 2026-08-19: $1.40 in / $4.40 out per 1M tokens — identical to GLM-5.2, with cached input at $0.26. Same base model as GLM-5.2; Z.ai says every gain comes from scaled post-training. Reasoning is always on (low / high / max effort — thinking can no longer be disabled, a breaking API change from GLM-5.2).

Weights went live on Hugging Face on 2026-08-25, inside the two-week window Z.ai promised at launch — a full checkpoint (142 safetensors shards, 1.4M downloads in the first month) under a custom "other" license named glm-5.3 rather than the MIT of GLM-5.2, with the launch-announced safety review of the model's exploit-finding capability resolved. This page's license field records the weights as open; the custom license's own terms live in the repo's LICENSE file, which this page does not restate.

Artificial Analysis has since run the model independently on Terminal-Bench 2.1 (83.9, 4.3 points below Z.ai's launch chart of 88.2 — inside that benchmark's noise band), Humanity's Last Exam without tools, and GPQA Diamond. vals.ai has independently run it on LiveCodeBench and SWE-bench Verified, though both are graded saturated here. deepswe.datacurve.ai's board picked the model up on or before 2026-08-20 at 69% (rank #4/18) — 2.1 points above Z.ai's own 66.9% launch-chart figure, inside the benchmark's noise band either way. Worth knowing how the Terminal-Bench gap compares: see the FAQ below on GLM-5.3's coding scores for the vendor-vs-independent pattern across tracked pairs on this site. Only the Toolathlon number and the with-tools HLE score are still Z.ai's own. One deliberate omission remains: Z.ai's Agents' Last Exam figure (28.5) is the 105-task ALE-CLI split, not the 1,000-item overall pass rate this site tracks. The GPQA Diamond row is tagged with a different evaluation harness (Artificial Analysis) than the vals.ai-sourced rows most of that column uses — but the row is moot for ranking either way, since GPQA Diamond is graded saturated on this site.

Compare with

FAQ

Is GLM-5.3 open source?

Its weights shipped on 2026-08-25, when Z.ai posted the full checkpoint to Hugging Face (142 safetensors shards, 1.4M downloads in the first month) under a custom "other" license rather than MIT — inside the two-week window the lab promised at launch. The pricing on this page is the hosted API's; the license terms on the weights are Z.ai's own text, and this page does not restate them.

How much does the GLM-5.3 API cost?

$1.40 per million input tokens and $4.40 per million output tokens — identical to GLM-5.2 — with cached input at $0.26. API pricing went live on 2026-08-19; before that the model was only reachable through the GLM Coding Plan subscription and ZCode.

Is GLM-5.3 good for coding?

Z.ai's launch chart says frontier-level: 88.2 on Terminal-Bench 2.1 and 66.9 on DeepSWE, with open-source SOTA claims. Artificial Analysis has since run Terminal-Bench 2.1 independently and got 83.9 — 4.3 points lower. That sits inside the benchmark's 10.6-point noise band, so it is not a caught exaggeration. It is worth context though: across the vendor-versus-independent pairs this site has measured, differences under 4 points fall either way about evenly, while every gap of 4 points or more has favored the vendor. DeepSWE has since been independently confirmed too (deepswe.datacurve.ai, 69% vs. Z.ai's own 66.9% — inside the noise band either way); only the Toolathlon number here is still Z.ai's alone.

What is the difference between GLM-5.3 and GLM-5.2?

Same base model — Z.ai says every gain comes from scaled post-training. The API price is unchanged, but thinking is now always-on (low / high / max reasoning effort) and can no longer be disabled, a breaking API change. On Z.ai's own numbers, Terminal-Bench 2.1 moves from 81.0 to 88.2 and DeepSWE from 46.2 to 66.9 — both are Z.ai's own before/after figures from the same launch comparison. Independently, this site's Terminal-Bench 2.1 rows are Artificial Analysis's 83.9 for GLM-5.3 and vals.ai's 67.79 for GLM-5.2, different harnesses; on vals.ai's own board the two read 71.54 and 67.79, a tie. For DeepSWE specifically, this site's own tracked score for GLM-5.2 is 44 (deepswe.datacurve.ai's independent v1.1 leaderboard, max effort), close to but not the same as Z.ai's self-reported 46.2 for the same model and benchmark version.

What is GLM-5.3's context window?

1 million tokens, with a maximum output length of 128K tokens. The knowledge cutoff has not been disclosed.

Why is there no Agents' Last Exam score for GLM-5.3 here?

Because the 28.5 in Z.ai's launch table is the 105-task ALE-CLI split, not the 1,000-item overall pass rate this site tracks. Mixing the two metrics is exactly how benchmark tables mislead — this site refuses numbers that don't match the tracked metric.

Has GLM-5.3 been independently benchmarked?

Yes, on nine of the eleven benchmarks tracked for it, most recently AA-AnalystAgent at 38.75 pass^5. Artificial Analysis has run it independently on Terminal-Bench 2.1 (83.9), a no-tools Humanity's Last Exam score (42.3), and GPQA Diamond (91.7). vals.ai has independently run it on LiveCodeBench (80.53), SWE-bench Verified (95.40) and Terminal-Bench 4.0 (38.89). LiveBench has run it on its own harness too, at 76.1 (Overall). deepswe.datacurve.ai has independently run DeepSWE at 69% (rank #4/18), edging out Z.ai's own 66.9% self-report by a margin well inside the benchmark's noise band. Toolathlon-Verified and with-tools HLE remain Z.ai's own. GPQA Diamond and SWE-bench Verified are both graded saturated on this site, so no pairing on either ever resolves to Real or Tie — those two rows exist for the record, not for ranking. The GPQA Diamond row is also tagged with a different evaluation harness than the vals.ai-sourced rows elsewhere in that column, recorded for completeness, though saturation alone already rules out a verdict there.

Further reading

Benchmark guides