Quick Verdict
GLM-5.3-Flash, released by Z.ai on August 26, 2026, is a more significant release than its "Flash" name suggests. Unlike GLM-5.3 (which mostly improved post-training on the same base as GLM-5.2), GLM-5.3-Flash starts from a newly trained, natively multimodal base model with a hybrid sparse-and-linear attention architecture built to cut long-context inference costs. Z.ai says it beats GLM-5.2 across its own evaluations while costing roughly one tenth as much per task.
This is a launch-week assessment, not a hands-on usage review — the model has been public for about three days at the time of writing. What follows is based on Z.ai's own release notes, the model card on Hugging Face, and independent listings from OpenRouter, which tracks real per-provider pricing and uptime rather than just vendor claims. We'll update this page with first-hand testing once we've run it against real coding workloads.
Our early take: 8.3/10 — a genuinely cheaper, more capable successor to GLM-5.2 with a useful new multimodal capability, tempered by the fact that most of the benchmark comparisons still come from Z.ai itself.
What Is GLM-5.3-Flash?
GLM-5.3-Flash is the latest release in Z.ai's GLM-5 series, and the first model in that lineup the company describes as natively multimodal — meaning it was trained from the start to handle text and vision together, rather than having vision bolted on afterward. It's a Mixture-of-Experts model with 320 billion total parameters, of which 18 billion are active per token, which is what keeps inference cheap despite the large total parameter count.
The architecture combines sparse and linear attention, an approach aimed specifically at reducing the compute cost of long-context inference — the kind of cost that piles up fast in long coding sessions, browser agents, and computer-use workflows that keep large amounts of context in play. Z.ai has already rolled the model into its GLM Coding Plan subscriptions and published the open weights on Hugging Face (zai-org/GLM-5.3-Flash) under the permissive MIT license, runnable through frameworks like vLLM and SGLang.
Notably, before the official announcement, a mystery model nicknamed "Ox Alpha" had been quietly appearing on leaderboards and OpenRouter's traffic charts. Once GLM-5.3-Flash launched, several outlets confirmed the two were the same model — Z.ai had been testing it in the wild under a different name before the public release.
GLM-5.3-Flash Pricing: API Costs
Z.ai's own launch material describes GLM-5.3-Flash as "one tenth the price" of GLM-5.2, but doesn't provide a clean side-by-side per-million-token table in the announcement itself. The actual listed rates, confirmed via OpenRouter's model page and Z.ai's own API documentation, are below.
| Access method | Price | Notes |
|---|---|---|
| Z.ai API (direct, list price) | $0.15 / 1M input tokens, $0.50 / 1M output tokens | Listed rate at launch |
| OpenRouter (launch promotion) | $0.075 / 1M input, $0.25 / 1M output | 50% launch discount across several providers as of this writing; check current rate before budgeting |
| Cached input (OpenRouter) | As low as $0.015 / 1M tokens | Applies to repeated context, e.g. long-running agent sessions |
| GLM Coding Plan | Same tiers as GLM-5.2 (Lite/Pro/Max, roughly $3–80/month) | GLM-5.3-Flash is included in the existing Coding Plan subscriptions; verify current pricing at z.ai/subscribe |
For comparison, GLM-5.2 launched at $1.40/$4.40 per million input/output tokens. Even at Z.ai's non-discounted list price of $0.15/$0.50, that's roughly a 9x drop on input and a 9x drop on output — close enough to Z.ai's "one tenth" framing to hold up. Z.ai also reports a discounted cost of $0.045 per task on its Artificial Analysis Intelligence Index run, though that figure depends heavily on task type and shouldn't be treated as a flat per-request price.
Try GLM-5.3-Flash's open weights directly
View GLM-5.3-Flash on Hugging Face →GLM-5.3-Flash vs Claude Opus 4.8: How It Stacks Up on Benchmarks
The figures below come from Z.ai's own release evaluation, as reported by independent outlets covering the launch. As with any vendor-published benchmark, treat these as a starting point rather than a substitute for testing against your own workload — independent, third-party reproduction of these exact numbers wasn't available at the time of writing.
| Benchmark | GLM-5.3-Flash | Claude Opus 4.8 |
|---|---|---|
| Terminal-Bench 2.1 | 84.3 | 85.0 |
| DeepSWE 1.1 | 63.4 | 58.0 |
| Toolathlon Verified | 78.4 | 76.2 |
| AutomationBench | 48.8 | 41.0 |
| OfficeQA Pro | 62.4 | 48.9 |
GLM-5.3-Flash edges out Claude Opus 4.8 on four of the five published benchmarks, and comes within a point on Terminal-Bench 2.1 — a meaningful result for a model priced at a fraction of Opus 4.8's rate. On Z.ai's own DeepSWE evaluation, GLM-5.3-Flash also scores substantially above its predecessor GLM-5.2, with an even larger gap on AutomationBench, reflecting the newly trained base model rather than incremental post-training tweaks.
Key Features
- Native multimodal support: The first model in the GLM-5 series trained from the ground up on combined text and vision data, rather than adding vision capability after the fact.
- Hybrid sparse-and-linear attention: An architecture specifically designed to keep long-context inference affordable, which matters for long coding sessions, browser agents, and computer-use workflows.
- 1M-token context window: Z.ai lists support for context lengths up to one million tokens; OpenRouter's listing shows an extended window of 1,310,720 tokens with a maximum output of 131,072 tokens through some providers.
- MIT-licensed open weights: Downloadable from Hugging Face and runnable through vLLM or SGLang for teams that need to self-host.
- A genuinely new base model, not a fine-tune: Z.ai says GLM-5.3 (without "Flash") was mostly a post-training update on GLM-5.2's base. GLM-5.3-Flash is different — a newly trained base model with a new architecture, which is why the benchmark jump is larger than a typical point release.
How to Access GLM-5.3-Flash
- Hugging Face: Download the open weights directly and self-host (
zai-org/GLM-5.3-Flash), MIT licensed. - Z.ai API: Pay-per-token access at the rates above; check z.ai for current pricing.
- OpenRouter: A unified endpoint routing across more than a dozen hosting providers, useful for redundancy and for comparing effective pricing after promotions.
- GLM Coding Plan: Already rolled into Z.ai's existing flat-rate subscription tiers for developers who'd rather not meter token usage.
Compare pricing and providers for GLM-5.3-Flash in one place
See GLM-5.3-Flash on OpenRouter →Pros and Cons
| Pros | Cons |
|---|---|
| Roughly a 9x price drop from GLM-5.2 on both input and output tokens | Released August 26, 2026 — no long-term track record and no independent hands-on review from us yet |
| Beats Claude Opus 4.8 on 4 of 5 published benchmarks, per Z.ai's own testing | Most benchmark comparisons still come from Z.ai; independent third-party reproduction is limited so far |
| First natively multimodal model in the GLM-5 series | Z.ai's launch material doesn't include a clean standard price table, which took some digging to confirm via OpenRouter |
| Open weights under MIT license, self-hostable via vLLM/SGLang | Using the hosted Z.ai API routes data through a China-based provider, which may matter for data-residency-sensitive teams |
| New architecture built specifically to reduce long-context inference costs | Launch pricing includes a promotional discount on some providers that may not be permanent |
Who Should Use GLM-5.3-Flash?
GLM-5.3-Flash is worth evaluating if you are:
- Already running GLM-5.2 in production — the pricing drop and benchmark gains make this a near-obvious upgrade path to test against your existing workloads.
- Running high-volume coding agents or browser/computer-use workflows — the long-context-focused architecture and native multimodal support are aimed directly at these use cases.
- Cost-sensitive about API spend — at roughly a tenth of GLM-5.2's price and a fraction of Claude Opus 4.8's, it's worth benchmarking against your actual task costs rather than list price alone.
- Working under data residency or self-hosting requirements — the MIT-licensed weights let you run it on your own infrastructure.
GLM-5.3-Flash is probably not the right fit yet if you:
- Need benchmark claims verified by more than the vendor and a handful of early trackers — independent evaluation is still thin three days after launch.
- Have strict requirements against routing data through China-based infrastructure and can't self-host the open weights.
- Want to lock in pricing today — some of the lowest advertised rates are launch promotions on third-party providers, not Z.ai's standing list price.
Frequently Asked Questions
Is GLM-5.3-Flash free?
The open weights are free to download and self-host under the MIT license, but you'll need substantial GPU resources for a 320-billion-parameter model. The hosted API is pay-per-token, listed at $0.15/$0.50 per million input/output tokens directly through Z.ai, with some third-party providers currently offering a 50% launch discount.
Is GLM-5.3-Flash actually better than GLM-5.2?
According to Z.ai's own evaluations, yes — it scores higher on DeepSWE 1.1 and AutomationBench, and it's a fundamentally different, newly trained base model rather than an incremental update. It also costs roughly a tenth as much per token. We haven't yet run our own side-by-side comparison against real workloads.
Was GLM-5.3-Flash the mystery "Ox Alpha" model?
Yes. Several outlets and community trackers identified matching behavior between the anonymous "Ox Alpha" model that had been appearing on leaderboards and OpenRouter traffic charts before launch, and the officially released GLM-5.3-Flash.
Can I self-host GLM-5.3-Flash?
Yes. The weights are published on Hugging Face under the MIT license and can be run through frameworks including vLLM and SGLang, permitting both self-hosting and commercial use.
How big is GLM-5.3-Flash's context window?
Z.ai lists support for up to 1 million tokens of context. OpenRouter's listing shows an extended window of 1,310,720 tokens through some providers, with a maximum output of 131,072 tokens.
Final Verdict
GLM-5.3-Flash is a more consequential release than a "Flash" naming suggests. Rather than a cheaper, cut-down variant of an existing model, it's a newly trained, natively multimodal base model with an architecture built specifically to make long-context, agentic workloads cheaper to run — and it backs that up with a roughly 9x price drop from GLM-5.2 alongside benchmark wins over Claude Opus 4.8 on four of five published tests.
The honest caveat is the same one that applies to almost every model in its first week: these are launch-week specs and vendor-reported benchmarks, not our own extended testing. We'll revisit this review once we've run GLM-5.3-Flash against real coding and agentic workloads.
Early rating: 8.3/10 — a strong, meaningfully cheaper successor to GLM-5.2 for cost-conscious, coding-heavy teams, with the usual new-release caveats attached.