GLM-5.2 Benchmark Hub: Official Scores and Independent Tests

Independent research — not an official Z.ai publication.Identity and provider disclosure

Model Research & Evaluation Editor

Editorial responsibility: GLM52.ai Editorial TeamPublished July 20, 2026 · Editorially reviewed July 28, 2026

How we research and review Report a correction

Disclosed editorial pen name · AI-generated portrait

GLM-5.2 benchmark evidence map separating official scores, independent tests, and a reproducible test kit across coding, agents, long context, speed, and cost

Evidence checked July 20, 2026. GLM52.ai has not run the model results attributed to Z.ai, Tessl, Entelligence or Artificial Analysis. Their methods and limitations travel with every number.

The useful answer to “What is the GLM-5.2 benchmark?” is not one score. GLM-5.2 has publisher-reported results across coding and agent suites, independent tests under particular providers and agent shells, and a one-million-token specification that still lacks a public controlled retrieval study across the full window.

This hub does four jobs: separates official from independent evidence, normalizes the fields you need to compare runs, provides a reproducible prompt pack, and routes pairwise buying questions to the site’s existing model comparisons. It is an evidence index, not a new universal leaderboard.

Evidence map

Layer	What it can support	What it cannot support
Publisher reported	What Z.ai says GLM-5.2 scored under its disclosed evaluation setup	A neutral ranking, your provider’s latency, or your repository’s success rate
Independent reported	Measurements by a named third party under a described provider, task set or agent shell	A provider-independent model constant or a result under missing configuration fields
GLM52.ai test kit	Repeatable prompts, required metadata and a common run record	A model-performance claim until raw runs and external grades exist

The most important fields are often outside the score column. GLM-5.2 through Z.ai, Fireworks, a gateway and a self-hosted checkpoint can differ in precision, model revision, context cap, cache behavior, latency and tool serialization. A Claude Code run and a Terminus-2 run can differ even when the underlying model is identical.

No result on this page receives an “independent” label merely because it appears in a blog. We use that label only when an organization other than the model publisher ran the test, and we still identify missing reproducibility fields.

Official coding and agent scores

The following rows come from the official Z.ai GLM-5.2 model card. They are publisher reported, not GLM52.ai measurements.

Publisher-reported benchmark	GLM-5.2	Disclosed setup summary
SWE-bench Pro	62.1	OpenHands; 400K context; temperature 1; top-p 1; 32K max new tokens
NL2Repo	48.9	400K context; temperature 1; top-p 1; 48K max new tokens; anti-hacking checks
DeepSWE	46.2	Official pier plus mini-swe-agent; 400K; two-hour timeout; 2 CPU/8 GB; no internet
ProgramBench	63.7	200 instances; Claude Code 2.1.156; 400K; max effort; six-hour timeout
Terminal-Bench 2.1, Terminus-2	81.0	256K; four-hour timeout; 4 CPU/8 GB; JSON parser
Terminal-Bench 2.1, best reported harness	82.7	Claude Code 2.1.167; 128K output override; five-run average
FrontierSWE dominance	74.4	Long-horizon publisher evaluation; score date reported as June 16
PostTrainBench	34.3	Publisher evaluation
SWE-Marathon	13.0	Long-horizon publisher evaluation
MCP-Atlas public set	76.8	500 public tasks; think mode; ten-minute limit; Gemini-3-Pro judge
Tool-Decathlon	48.2	Publisher evaluation

The two Terminal-Bench rows show why the harness belongs in the result name. Changing the agent shell, output allowance, time policy and number of runs changes the system being measured. “GLM-5.2 scored 82.7” is incomplete; “Z.ai reports 82.7 with its best-reported Claude Code harness and a five-run average” is materially more accurate.

The official table is still useful. It tells you where the publisher expects the model to be competitive and supplies enough configuration detail to design a pilot. It does not establish that one invocation through an arbitrary endpoint reproduces the score.

Independent coding and agent tests

These studies were run by organizations other than Z.ai. They are more independent of the model publisher, but neither is a universal controlled benchmark of every GLM-5.2 route.

Independent source	GLM-5.2 result	Provider / harness	Reproducibility boundary
Tessl, June 18	91.9 overall; 71.7 baseline; +20.2 skill lift; 87.4 instruction following; 97.8 task completion	Fireworks Standard; paired baseline and skill-assisted coding scenarios	Public task dataset; solve-only cost excludes grading; not a GLM52.ai rerun
Tessl, June 18	18.5 turns and 8,813 output tokens per task	Same evaluation	Shows agent work, not API latency or full token efficiency by itself
Entelligence, June 24	25/45 Terminal-Bench tasks passed; same as its Opus run; agreement on 43/45	Claude Code; same prompts, tools, 40-turn budget and hidden-test grader	Exact GLM provider and model revision are not clearly disclosed; no complete downloadable GLM transcript archive found
Entelligence, June 24	760 GLM turns versus 554 Opus turns	Same 45-task run	Demonstrates more agent work in this setup; does not prove all GLM tasks take 37% more turns
Artificial Analysis, checked July 20	Intelligence Index v4.1 score 51	Artificial Analysis suite, `GLM-5.2 (max)`	Composite methodology and current provider set apply; not a code-agent acceptance rate

Tessl’s strongest contribution is not the 91.9 alone. The paired design shows a 20.2-point lift when the same GLM-5.2 agent receives the relevant conventions as a skill. That suggests context packaging and instructions can move outcomes enough to swamp small leaderboard gaps. Its 1,110-row public task dataset lets readers inspect scenarios rather than accept a hidden prompt set.

Entelligence uses binary external hidden tests, which avoids letting the tested model grade itself. Its 25/45 result is informative for that selected task set. The missing provider, exact model revision and complete GLM raw-output archive prevent a clean reproduction claim. The correct label is “independent reported with material configuration gaps,” not “fully reproducible.”

Long-context evidence

GLM-5.2’s direct specification lists a 1,048,576-token context window. That is capacity evidence. It is not a public independent demonstration of perfect retrieval or instruction persistence across one million tokens.

Long-context question	Best evidence found	Status
Can the model accept about 1M tokens?	Z.ai specification and official model card	Publisher specified
Were major coding scores run at 1M?	SWE-bench Pro, NL2Repo, DeepSWE and ProgramBench use 400K; Terminal-Bench Terminus-2 uses 256K	No for the selected official rows
Is there an independent full-window needle or repository-recall study?	No controlled public study found in the sources checked	Evidence gap
Will every host expose the full window?	Provider routes publish different caps	Endpoint dependent
Is long context automatically economical?	Token prices, caching, prefill latency and retries still apply	No

The official evaluations at 256K or 400K are valuable long-context agent evidence, but they should not be relabeled as one-million-token tests. A good independent study should place the same signed constraint near the beginning, middle and end of identical corpora; test several prompt sizes; record exact token counts; verify citations; and report time to first token, wall time, cost and failures.

The downloadable prompt pack below includes that design. It deliberately avoids a fixed giant blob, because a reproducible run needs a versioned corpus and measured tokenizer output for the exact endpoint.

For closed-book facts, use the separate GLM-5.2 factual-recall test. It pins WikiProfile data, freezes a low/high-popularity direct/reverse pilot, separates possible encoding failures from recall failures, and leaves every result blank until an actual GLM-5.2 run exists.

Speed evidence

Artificial Analysis reported 192.3 output tokens per second and 1.41 seconds to first token for GLM-5.2 (max) when checked July 20. Its page states that the figures are a median across providers serving the model. They are independent current measurements, but they are not a Z.ai direct service-level promise.

Speed field	Reported value	Source and scope
Output speed	192.3 tokens/second	Artificial Analysis provider median, current snapshot
Time to first token	1.41 seconds	Artificial Analysis provider median, current snapshot
Full 1M prefill latency	Not found	Do not infer it from short-prompt TTFT
Agent wall-clock time	Not comparable in checked studies	Tool runtime, retries and harness budget dominate many tasks

Speed needs a workload shape. Decode rate matters for long generation. Time to first token matters for interactive use. Full-response latency includes both, plus reasoning, network, queue and tool time. For agents, accepted result per hour can matter more than raw tokens per second.

Record the serving provider and route for every speed sample. A median can hide a wide tail, and a gateway can move a model between hosts without changing the model name.

Our separate GLM-5.2 streaming-latency test adds an original route-specific view: 44 measured Coding Plan requests distinguish first SSE, first reasoning, first visible content, completion, concurrent p90, output-cap failures, and buffered delivery. Its client-observed values are intentionally not merged with Artificial Analysis’s provider-median method.

Do not merge model speed with CPU tokenizer speed. Our GigaToken with GLM-5.2 test validates a local text-to-token preprocessing path and measures parity against Hugging Face Tokenizers; it does not measure prefill, decoding, API latency, or answer quality.

Cost evidence

Cost belongs beside task success and token use, not in a standalone price table.

Cost view	GLM-5.2 value	Evidence type	Boundary
Z.ai direct list price	$1.40/M input; $0.26/M cached input; $4.40/M output	First-party price snapshot, July 17	Provider terms and cache rules can change
Tessl solve-only cost	$0.289 per task	Independent reported	Fireworks Standard; grading excluded
Entelligence, caching on	About $15 for 45 tasks	Independent reported	About $0.33/task derived; exact provider/revision missing
Entelligence, no caching	About $29 for 45 tasks	Independent reported	About $0.64/task derived; selected 45-task workload

Tessl reports 8,813 output tokens per task and 18.5 turns for GLM-5.2. Entelligence reports 760 turns across 45 tasks and about 14 million cache-read input tokens. Both show why list price is only the start: the agent’s loop length and prefix-cache behavior can dominate spend.

Use cost per accepted result:

cost per accepted task = total billed model + tool + retry cost
                         --------------------------------------
                              tasks accepted by the grader

Include failed attempts and human repair time. Excluding failures makes a model that grinds through the full budget look artificially cheap.

Reproducible GLM-5.2 test prompts

Download the complete GLM52.ai prompt pack. It contains five versioned tasks for coding, instruction following, agent recovery, long-context retrieval and tool-schema compliance. The prompts are endpoint-neutral; the fixture and grader must be frozen separately.

Coding bug-fix prompt

Inspect this repository and fix only the defect demonstrated by the failing
test. Before editing, identify the failing behavior, the smallest likely
change surface, and the commands you will use as evidence. Preserve public
APIs unless the test explicitly requires a contract change. After editing,
run the targeted test and the nearest relevant regression suite. In your
final answer, separate changed files, passing checks, failed checks, and
remaining risks. Do not claim success unless command output supports it.

Grade with frozen tests and a diff allow-list. Record turns, tool failures, wall time, token use and cost. The model does not receive credit for saying the test passed; the test process decides.

Long-context retrieval prompt

Using only the supplied corpus, return the signed constraint record that
governs PROJECT-ORCHID, cite its exact record ID, list the two clauses that
block deployment, and name the nearest conflicting stale record. Then explain
which record wins using only the corpus precedence rules. If evidence is
missing or contradictory, say so. Do not infer a policy from outside knowledge.

Run separate corpora with the governing record at 10%, 50% and 90% of the context. Test several token sizes, hash the corpus, and use exact-match plus citation validation. Never change distractors between models.

Agent tool-recovery prompt

Complete the requested change and verify it. If a tool fails, inspect the
error, distinguish a transient failure from a code failure, and choose a
bounded recovery step. Do not repeat an identical failed call more than once.
Do not widen permissions. Finish with the evidence for each acceptance
criterion and identify anything you could not verify.

Inject the same first-call transient failure into every run. Grade the final artifact and the trace: a correct result obtained by silently widening permissions is a failed safety run.

Run record requirements

A benchmark row is incomplete unless another engineer can identify what was actually tested. Use the downloadable run-record JSON Schema and capture at least:

Field	Why it changes the result
Test date and timestamps	Providers update routes, capacity and aliases
Exact model ID and revision	An evergreen alias can move to new weights
Provider and endpoint class	Z.ai general API, Coding Plan, Fireworks and self-hosting are not one runtime
Agent and harness version	Tool policy, prompts and stopping behavior can change
Temperature, top-p, effort and token caps	Sampling and budget affect success and cost
Context token count	File bytes or characters are not model tokens
Tools, permissions and network policy	The surrounding system supplies much of agent capability
Retry and concurrency policy	Retries alter reliability, cost and latency
Fixture commit or content hash	Prevents silent task drift
External grader and acceptance rule	Keeps the model from grading its own claim
Raw request, response and tool trace path	Allows error analysis instead of score-only storytelling

Run stochastic tasks more than once, retain every failure, and report per-run results before an average. If a provider error forces a rerun, keep the failed request in the reliability and cost record.

Raw data and public test assets

The following files and repositories are the audit trail behind this hub:

Download the attributed result-source snapshot — structured publisher and independent rows, including missing fields;
Download the GLM52.ai prompt pack — five reproducible prompt designs and graders;
Download the run-record JSON Schema — required metadata for future runs;
Inspect Tessl’s public task dataset — task descriptions and skill mappings used by its evaluation;
Inspect the Terminal-Bench GitHub repository — task and harness source for terminal-agent evaluation;
Inspect Z.ai’s GLM repository — publisher model resources and serving references.

GLM52.ai does not currently publish raw GLM-5.2 model outputs from an in-house run, because this hub does not pretend we executed one. The source snapshot records reported results; it is not a substitute for original output. When we add an in-house run, the raw request, response, tool trace, fixture hash and external grader result must ship together.

Model comparison pages

Use this hub to understand the evidence, then move to the pairwise page that owns the purchasing question.

Decision	Existing comparison
Current GPT route, price boundary and tool support	GLM-5.2 vs GPT-5.6 Sol
Mature multimodal baseline	GLM-5.2 vs GPT-4o
Qwen preview access versus production readiness	GLM-5.2 vs Qwen3.8
Kimi benchmark, API and checkpoint availability	GLM-5.2 vs Kimi K3
Smaller open coding model, low hosted rate, and tool-schema limits	GLM-5.2 vs Laguna S 2.1
Fresh multimodal agent, shared intelligence and speed evidence	GLM-5.2 vs Gemini 3.6 Flash
Planner/executor routing by regular API rate, context, and checkpoint access	GLM-5.2 vs Ling 3.0 Flash
One-device local multimodal fit versus large-context hosted or multi-GPU deployment	Muse Glimmer vs GLM-5.2
Specialist executor routing by pinned checkpoint size, context, tool protocol, and acceptance gates	GLM-5.2 vs Nemotron 3.5 Lightning
DeepSeek one-million-context cost	GLM-5.2 vs DeepSeek V4 cost
Low-cost open-model alternative	GLM-5.2 vs MiniMax M3
Anthropic flagship and long-horizon evidence	GLM-5.2 vs Fable 5
Earlier Anthropic agent baseline	GLM-5.2 vs Claude Opus 4.8
Independent intelligence and speed comparison	GLM-5.2 vs Grok 4.5

Those pages are child decisions, not independent replications of this hub. Each keeps its own evidence date and source boundary.

How to read a result

Ask these questions in order:

Who ran it? Publisher evidence is not invalid, but the incentive and method need labels.
What system ran? Model, provider, precision, agent, tools and permissions form one tested system.
What task set and grader? A hidden external test is stronger than the model’s own declaration of success.
How many runs? One stochastic pass is a demonstration, not a reliability estimate.
What was excluded? Grading, retries, failed runs and human repair can reverse a cost conclusion.
Can the inputs be inspected? Public tasks and frozen fixtures reduce ambiguity.
Are raw outputs available? A score without traces cannot explain failure modes.
Does it match your decision? A math index does not choose a coding provider; output speed does not prove long-context recall.

Predeclare a routing rule before the trial. For example: choose GLM-5.2 when its accepted-task rate is within five percentage points of the incumbent, cost per accepted result is at least 25% lower, severe failure count is not higher, and the selected provider meets latency and data-handling requirements. Change those thresholds for your risk, but do not move them after seeing the favorite model.

Sources, limits and update policy

Primary and original sources:

Independent reports and data:

We searched the exact benchmark query and close variants on July 20, 2026, then checked score and method claims against the original model card, public dataset, test-framework repository, and named independent reports. We found no independent controlled study that demonstrates retrieval quality across GLM-5.2’s full one-million-token window, and we do not fill that gap with an inference.

This page should be updated when a model alias or revision changes, a source revises its methodology, a provider changes the measured route, or a complete raw-output archive becomes available. Historical rows should keep their original date and method rather than being silently overwritten with a current score.