GLM-5.3-Flash unmasked: what the stealth test means for teams using coding agents

The anonymous test model "Ox Alpha" that ran free on OpenRouter for five days has been unmasked: it is GLM-5.3-Flash from Beijing-based Z.ai — a model with a million-token context window, open weights under the MIT license, and a price of 7.5 US cents per million input tokens. For teams running coding agents, this is the most interesting price-performance combination of the year — but not yet a green light for production repositories.

What happened

On 26 August 2026, Z.ai confirmed to Bloomberg that it was behind the stealth model, as reported by blogspan.net. Shortly afterwards the model appeared with its name and price list in the OpenRouter catalogue; the free preview was switched off at the same time.

The hard facts in brief:

• 320 billion parameters in total, only 18 billion active per token — a mixture-of-experts model with 288 experts, eight of which compute per token.
• A native context window of 1,048,576 tokens — the million from the preview was the actual design, not a marketing add-on.
• Pricing via OpenRouter: $0.075 per million input tokens, $0.25 per million output tokens.
• The weights are available on Hugging Face under the MIT license, shipped in fp8 quantisation.
• No prompt caching: send the same long context twice and you pay for it in full, twice.
• Image and video tokens are part of the configuration; audio is not.

In the vendor-supplied benchmarks, GLM-5.3-Flash leads in one of six disciplines; on the coding-adjacent Terminal Bench and DeepSWE tests, GPT-5.6 Terra and Gemini 3.7 Flash sit above it. For this price class, the proximity to the frontier is still remarkable — the comparison models cost ten times more per output token and beyond.

Why this matters

The price is not the product of subsidy or dumping but of architecture. Of the model's 45 layers, only eleven use full attention; the remaining 34 use linear attention. The cost of full attention grows quadratically with context length, while linear methods do not — and that is exactly what finances a million-token window at this price. Z.ai describes it as the first hybrid architecture of this kind in the GLM series.

Two points make this story bigger than a model launch. First, the MIT license: releasing a frontier-class model of this size without usage restrictions is the most permissive option currently on offer — and it opens the door to self-hosting, which we will come back to. Second, the market strategy: a lab that lets people test its model anonymously first and then discloses it is betting on adoption speed rather than API revenue. That shifts pricing pressure across the entire market.

For comparison: the larger sibling GLM-5.3 sits in the same catalogue at $1.40 and $4.40 — a factor of 18 and 17.6 above the Flash model. The direct predecessor GLM-4.7-Flash costs $0.06 input and $0.40 output: the new version is a quarter more expensive on input but almost 40 percent cheaper on output — with roughly five times the context.

What this means for teams using coding agents

For companies running — or planning — coding agents, the maths gets concrete.

Agent runs get noticeably cheaper

Coding agents read a lot and write a lot: entire repositories, long tool outputs, multi-step patches. Long outputs in particular drive cost. That is where the new model is strong — output is significantly cheaper than the predecessor's, and the large window allows agent runs across complete codebases without constantly compressing the context. If you are currently running agents on models in the four-dollar class, benchmark your own workload against this.

The missing cache is the real cost driver

The most important limitation is not the list price but the lack of prompt caching. Typical agent setups resend the repository context with every step. With a million-token window, a single agent session can burn through hundreds of millions of input tokens — and then 7.5 cents per million becomes a budget line. Work out your read-to-write ratio before switching: workloads that read a lot and write little may be better served by the cheaper predecessor.

Open weights change the data-protection debate

In our view, this is the most interesting point for mid-market companies. The prompts from the stealth preview still sit with the provider — anyone who fed the model real code back then sent that code to a party whose name they did not know. Via the API that remains true: every prompt goes to a Chinese provider, with everything that implies for GDPR, trade secrets and customer data. The MIT license, however, offers the way out: self-hosting on your own or European infrastructure. For the first time, a model of this context class becomes available without compromising data sovereignty — an argument the closed western frontier models simply do not have.

What to do now

1. Build a test set, don't switch: take ten to twenty real tasks from your agent workload and run GLM-5.3-Flash in a sandbox against your current model. Measure success rate, token consumption and cost per solved task — not benchmark percentages.
2. Test only with non-sensitive material: as long as the API is the route, no production repositories, customer data or credentials belong in prompts. The preview phase showed how easily code ends up with unknown recipients.
3. Evaluate self-hosting: if you already have GPU capacity or an infrastructure partner, cost out on-prem operation of the fp8 variant. For organisations with strict data-protection requirements, that is the real lever of this release.
4. Wait for independent long-term tests: the benchmarks come from the vendor. Give the community a few weeks before moving the model into critical processes.

Two things you should not do: migrate your agent fleet overnight because the list price is low — changes in model behaviour partially invalidate prompt libraries and evaluations. And do not treat the stealth preview as a harmless free trial — stored prompts cannot be retrieved.

The bigger picture

GLM-5.3-Flash is one data point in a larger shift: Chinese labs are increasingly betting on open weights and aggressive pricing while western frontier providers keep their models behind APIs. For European companies this creates a paradox — it is precisely the open models from China that can be operated without data flowing abroad. Anyone aligning their AI strategy with data sovereignty should keep this on their radar: in 2026, the competition around coding agents will no longer be decided on capability, but on cost per task and control over your own data.

Sources

• reported by blogspan.net — https://www.blogspan.net/ox-alpha-glm-5-3-flash-zai-aufloesung/
• OpenRouter catalogue — https://openrouter.ai/z-ai/glm-5.3-flash
• Hugging Face — https://huggingface.co/zai-org/GLM-5.3-Flash