GLM-5.3: the same base model, a much better coder — and a cyber result nobody planned

The short answer: GLM-5.3 is GLM-5.2's base model with a month of extra post-training on top — and it is a large jump in agentic coding. It is live now on the GLM Coding Plan and the API. Weights are promised in two weeks. If your code sends thinking.type: "disabled", your requests will fail on GLM-5.3 — that migration is the only urgent thing in this launch.

Verified 14 August 2026 against Z.ai's launch post, "GLM-5.3: Frontier Coding with Emergent Cyber Capabilities", including its benchmark table and methodology footnotes. All numbers below are Z.ai's own self-reported figures unless stated. There is no independent evaluation of GLM-5.3 yet, so it is not on our LiveBench snapshot — a lab's own benchmark is a claim, not a third-party score.

What actually shipped

The same base model. This is the headline most coverage will miss. Z.ai states plainly that GLM-5.3 "uses the same base model as GLM-5.2 — every gain comes from post-training." No bigger model, no new pre-training run. The gains come from scaling reinforcement learning on long-horizon task environments.

That is a genuinely interesting result, and it is also why the improvements are lopsided: they are enormous where long-horizon agentic work is measured, and modest elsewhere.

A live API, model id glm-5.3, rolled out to all GLM Coding Plan subscribers.

Open weights in two weeks, once "safety evaluation and hardening are complete." A date-bounded commitment, not an indefinite gate.

The benchmark jumps that matter

Z.ai's self-reported figures, GLM-5.3 vs GLM-5.2:

Terminal-Bench 3.0 going from 4.6 to 28.3 is the clearest signal: on the hardest long-horizon agentic tasks, GLM-5.2 was essentially failing, and GLM-5.3 is not.

Note the shape of it. Terminal-Bench 2.1 moved 81.0 → 88.2, a normal increment. Terminal-Bench 3.0 moved 4.6 → 28.3, a 6× jump. The harder and longer the task, the bigger the gain — which is exactly what you would expect if the improvement is long-horizon RL rather than raw capability.

Where it still loses. On Terminal-Bench 3.0, Z.ai's own table puts GPT-5.6 Sol at 34.6 and Fable 5 at 33.7, both ahead of GLM-5.3's 28.3. On its in-house Z.ai Code Bench, GLM-5.3 reaches 34.5% at Max effort against Claude Fable 5's 39.5%. Z.ai says so directly, which is worth some credit.

The token-efficiency result is the business story

Buried under the benchmark table is the number that actually affects a bill.

On Z.ai Code Bench at Max effort, GLM-5.3 reaches 34.5% using roughly 75K output tokens per task — against GLM-5.2's 23.4% at 96K. Better results, fewer tokens.

Against a closed model: at High effort GLM-5.3 hits 31.4% at around 50K output tokens, versus Claude Opus 4.8 at 29.5% with 120K. Slightly better score, roughly 2.4× fewer output tokens.

For anyone running coding agents at volume, output tokens are the bill. A model that scores the same for a fraction of the tokens is a cost story, not a leaderboard story.

The cyber result, and why it is double-edged

Z.ai added vulnerability-discovery data to the training mix expecting modest gains, and reports being surprised: GLM-5.3 began reasoning across full exploitation chains, not just spotting isolated flaws. On ExploitGym it completes 105 tasks in two hours and 130 in six, against GLM-5.2's 29 and 39.

Run against real codebases with Chinese security teams, the model surfaced 2,436 vulnerabilities across 269 open-source projects — 107 critical and 990 high. The average flaw had survived 26.6 years undiscovered; the oldest dated to 1981. Z.ai has published a Security Disclosure Ledger tracking these, with 53 disclosed and 2,383 still under embargo.

Two honest caveats. First, Z.ai's own comparisons show it well behind the closed frontier here — the page cites competitors at 181 and 247 ExploitGym tasks against GLM-5.3's 105 and 130. Z.ai's summary of its own gap is candid: capability is growing fastest exactly where it is furthest behind.

Second, a capable open-weights vulnerability-discovery model is dual-use by construction. The two-week weights delay for "safety evaluation and hardening" is presumably about precisely that.

(A labelling inconsistency worth flagging: Z.ai's prose refers to "Mythos 5" for the cyber comparisons while its own table column is headed "Fable 5". We have quoted GLM-5.3's own figures rather than guess which competitor is meant.)

The migration that will break your calls

GLM-5.3 supports three thinking effort levels — low, high, max. Disabling thinking is no longer supported.

If your application currently sends thinking.type: "disabled", change it to enabled and set reasoning_effort to low before switching the model id to glm-5.3. Otherwise the request fails. Z.ai recommends max for coding tasks.

This is the one item in the launch with a deadline attached to it.

What it costs

Z.ai did not publish per-token pricing. The GLM Coding Plan has moved to a points-based quota system, with points counted separately for input, cached input, and output tokens.

The pricing detail that matters most: calls outside peak hours consume 50% of the standard points. Peak is 14:00–18:00 UTC+8, Monday to Friday. Everything else — including all weekend — is half price.

If you run batch work, that is a scheduling decision worth making deliberately. Overnight in Europe or the Americas is off-peak in UTC+8.

Through ZCode, Z.ai also advertises a 98%+ cache hit rate on repeated context and a 1.5× limited-time quota boost until 31 August.

What to do this week

If you are on a GLM Coding Plan. GLM-5.3 has already been rolled out to you. Check your code for thinking.type: "disabled" first, then try it — the token-efficiency gain is the thing to measure, not the benchmark number.

If you self-host GLM-5.2. Wait two weeks for weights. Nothing to pull yet.

If you use a coding harness. GLM-5.3 works in ZCode, Claude Code, OpenCode and others — you are not locked into Z.ai's own client.

If you were comparing GLM against Claude or GPT. GLM-5.3 narrows the gap on agentic coding substantially and wins clearly on tokens-per-result. It does not take the top spot on Z.ai's own table. Decide on the cost-per-outcome, not the ranking.

The honest caveats

These are the lab's own numbers. Every figure here is self-reported. Z.ai documents its methodology unusually thoroughly — harness versions, temperatures, context lengths, timeouts, anti-cheat measures — which is more than most labs publish, and it still is not an independent evaluation. GLM-5.3 is on our model tracker with its figures labelled as Z.ai's own, and it will not join our benchmark page until a third party scores it.

Several benchmarks are in-house. Z.ai Code Bench is private and unauditable by design. Its contamination argument is reasonable; it is still Z.ai grading Z.ai.

"Open weights" has not happened yet. Two weeks is a stated intention. GLM-5.2 shipped weights on day one; this did not.

Data and jurisdiction. Unchanged: know where your tokens go before routing customer data through a hosted endpoint.

The move this week

Do the migration check. Measure output tokens, not scores. If you batch, move the work off-peak and take the 50%.

If you want the "which model for which job" split decided against your actual workflow and bill, a Business AI Plan ($99 one-time, optional $29/month to keep it current) does that. The free 2-minute AI plan finder is the fast first pass.

Sources: Z.ai — GLM-5.3: Frontier Coding with Emergent Cyber Capabilities, Z.ai developer docs, ZCode. Retrieved 14 August 2026.