GLM-5.3-Flash is an open-source large language model built by Z.ai (Zhipu AI) and released under an MIT licence on August 26, 2026, after a five-day anonymous preview on OpenRouter under the code name Ox Alpha. OpenRouter listed a free anonymous AI model named Ox Alpha on August 20–21, 2026, carrying a 1,048,576-token context window and a stated design focus on coding, sustained agentic work, and production workflows. The provider was identified only as “stealth” — no company, no lab, no benchmark numbers.
Ox Alpha accepts text, images, and video as input and returns text. OpenRouter lists a maximum output of 131,072 tokens and prices both prompt and completion tokens at $0 for a preview window the OpenCode team described as roughly one week. OpenRouter’s listing states that the provider retains prompts but does not use them for training. The OpenCode team confirmed “near unlimited usage, zero data retention” for the free window. A Hacker News discussion of the launch reached 178 points on August 21 and characterises the model as built for long-horizon software engineering, complex reasoning, and workflows that combine text with visual context.
“This model is stealth — developed and operated by a third-party provider who has chosen to remain anonymous during this preview.” — OpenRouter listing
Z.ai Confirms: Ox Alpha Is GLM-5.3-Flash
On August 26, 2026, Bloomberg News reported that Z.ai confirmed Ox Alpha as a new iteration of its GLM model series in response to a direct media inquiry. The model’s official name is GLM-5.3-Flash. Z.ai is the international brand of Zhipu AI, the Chinese AI company behind the GLM open-weight series. The company released model weights on Hugging Face under an MIT licence on the evening of August 26, fulfilling a same-day announcement. GLM-5.3-Flash runs on Chinese AI chips and carries a 320B-A18B architecture with a 1,048,576-token multimodal context window — the same 1M window OpenRouter listed for Ox Alpha.
On the DeepSWE Pass@1 benchmark — a coding-agent evaluation measuring end-to-end software engineering task resolution — GLM-5.3-Flash scored 80%, placing ahead of Claude’s 65% and GPT-5.6 Sol’s 52% at the time of measurement. That benchmark result explains the aggressive free preview strategy: the stealth launch gave Z.ai real production traffic and user behaviour data before the formal announcement, a pattern GLM-5.3’s predecessor (GLM-5) established on the same platform. Businesses that tested Ox Alpha during the free window were, in effect, running evaluation traffic on a GLM-5.3-Flash deployment.
Why a $0 Open-Source Model With 1M Context Changes the Experimentation Math
Ox Alpha removed the token bill from long-context agentic testing, which is the single largest cost line in codebase-scale AI experiments. A 1M-token window holds an entire mid-size repository, a full quarter of support tickets, or a multi-hundred-page contract set in one prompt — work that costs $3–15 per million tokens on Claude Sonnet 5 and GPT-5.6 Lun, and $0 on Ox Alpha during the preview.
The workloads that benefit are the ones businesses already run against paid frontier APIs: repository-wide refactors, multi-step data processing, document-set analysis, and CI failure triage. Teams comparing those jobs across providers can measure the same benchmark-backed options in our guide to the best AI agents for business tasks, then run the identical prompts through the now-public GLM-5.3-Flash weights at zero marginal cost for self-hosted deployments.
What the Open-Source Weights Mean for Business Users
GLM-5.3-Flash’s MIT licence removes the primary adoption barrier that the anonymous preview created. An MIT-licensed open-source model can be self-hosted, fine-tuned, integrated into commercial products, and redistributed without royalties — unlike the API-only access terms of GPT-5.6 Sol or the Apache 2.0 restriction that governs some Chinese open-weight releases. The 320B-A18B architecture (320 billion total parameters, 18 billion active per inference) makes self-hosting realistic on multi-GPU deployments rather than requiring a full frontier-scale cluster.
Post-preview pricing through OpenRouter is not yet announced. Teams that want to keep using GLM-5.3-Flash after the free window should check OpenRouter’s listing directly or monitor Z.ai’s official channels. For a comparison of named open-weight models with confirmed pricing and support terms, see our guide to AI tools for business.
What Remains Unverified After the Confirmation
The DeepSWE 80% figure was reported as the benchmark score at preview launch; an independent third-party replication on the published weights had not been confirmed as of August 27, 2026. Post-preview API pricing through OpenRouter is TBD. The data-retention terms that applied during the anonymous Ox Alpha window — “prompts retained, not used for training” — now require review against Z.ai’s formal GLM-5.3-Flash terms, since the anonymous operator assurance carries no contractual backing. Businesses that ran regulated, customer, or proprietary data through the free preview should verify what terms applied to their specific requests via OpenRouter’s provider agreement.
The Stealth-Drop Pattern: Four Confirmed Chinese Labs
Z.ai’s GLM-5.3-Flash is the fourth Chinese AI lab confirmed to have used OpenRouter’s stealth program for a production-traffic preview before a formal launch. The four confirmed cases are Zhipu AI (GLM-5, 2025), Xiaomi (MiMo-V2-Pro), Ant Group (Lingxi Ling-2.6-flash), and Meituan (LongCat-2.0). Xiaomi’s MiMo team was the most-cited community candidate for Ox Alpha before Z.ai’s confirmation — that speculation is now definitively closed.
The mechanism remains consistent: OpenRouter’s stealth program gives a lab real production traffic and real user behaviour data before a formal launch, without the PR scrutiny or competitive signalling that a named release triggers. The reveal timeline for all four cases was within weeks of the preview start, not months. GLM-5.3-Flash fits the pattern exactly — five days from OpenRouter listing (August 21) to Bloomberg confirmation (August 26).
For Context: Our Coverage of Z.ai and Chinese Frontier Models
- Stripe’s $7B OpenRouter acquisition — who now owns the router that hosts every stealth drop, and what that means for your AI stack.
- GLM-5.2 review — Zhipu AI’s 1M-context model, the predecessor now confirmed as the same lab that built Ox Alpha.
- GLM-5.3 coding results — the open-source coding leader Ox Alpha was benchmarked against, now confirmed as the same model family.
- Qwen3.8-27B open weights — the free-to-run alternative that carries a named lab and a published licence.
- DeepSeek V4-Pro pricing — what a Chinese frontier model charges once it exits preview, and the rate GLM-5.3-Flash’s free window will be judged against.
Frequently Asked Questions
What is Ox Alpha?
Ox Alpha was the code name for GLM-5.3-Flash, an open-source AI model built by Z.ai (Zhipu AI). It was listed anonymously on OpenRouter on August 20–21, 2026, with a 1,048,576-token context window and a free preview period. Z.ai confirmed its authorship to Bloomberg News on August 26, 2026, and released MIT-licensed weights on Hugging Face the same evening.
Who built Ox Alpha?
Z.ai (the international brand of Zhipu AI, a Chinese AI company) built Ox Alpha, confirmed via Bloomberg News on August 26, 2026. The model is officially named GLM-5.3-Flash and is part of Z.ai’s GLM open-weight model series. Before Z.ai’s confirmation, the builder was unknown — OpenRouter listed the provider as “stealth.”
How did Ox Alpha score on benchmarks?
GLM-5.3-Flash (Ox Alpha) scored 80% on the DeepSWE Pass@1 benchmark, which measures end-to-end software engineering task resolution. That placed it ahead of Claude (65%) and GPT-5.6 Sol (52%) at the time of measurement. Independent third-party replication on the public weights had not been confirmed as of August 27, 2026.
Is Ox Alpha / GLM-5.3-Flash free to use?
The free preview on OpenRouter ended around August 27, 2026. Post-preview API pricing through OpenRouter has not been announced as of the time of this update. The model weights are open-source under an MIT licence and available on Hugging Face, meaning the model can be self-hosted at no licensing cost — compute and hosting costs apply.
What is GLM-5.3-Flash’s context window and architecture?
GLM-5.3-Flash has a 1,048,576-token (1M) multimodal context window and a 320B-A18B architecture — 320 billion total parameters with 18 billion active per inference. It accepts text, images, and video as input and returns text. The model runs on Chinese AI chips and is released under an MIT open-source licence.
