Ox Alpha Was GLM-5.3-Flash: Specs, Price, and How to Use It

Ox Alpha showed up on OpenRouter on 20 August 2026 as stealth/ox-alpha, with no lab name, a 1,048,576-token context window, and a price of $0. For six days it was the mystery coding model of the week. On 26 August, Z.ai published GLM-5.3-Flash and said they had been testing that model anonymously as ox-alpha on OpenRouter and OpenCode. OpenRouter now labels the stealth listing as Z.ai GLM-5.3-Flash. The free preview is over.

People also search for 0x Alpha and 0xAlpha. The official name uses the letter O. The callable model ID today is z-ai/glm-5.3-flash on OpenRouter, or glm-5.3-flash on Z.ai's own API. stealth/ox-alpha is gone from the active catalog. If your agent config still points at the old slug, change it.

This is a recap of the stealth week, the reveal, the numbers that survived contact with independent runs, and what we'd actually do with the model on a client job. We have not put GLM-5.3-Flash through our own harness yet. Vendor tables and other people's evals are cited as such.

What Ox Alpha is, in numbers

• Developer: Z.ai (Zhipu AI). Confirmed 26 August 2026. Previously listed only as an anonymous "stealth" provider.
• Public name: GLM-5.3-Flash. Alias during preview: Ox Alpha. OpenRouter slug now: z-ai/glm-5.3-flash. Z.ai API code: glm-5.3-flash.
• Architecture: mixture-of-experts, 320 billion total parameters, 18 billion active per token. First natively multimodal model in the GLM-5 series. Hybrid sparse plus linear attention.
• Context window: 1,048,576 tokens. Max output is listed as 131,072 tokens on OpenRouter and 128K in Z.ai's docs. Check the route you actually call.
• Inputs: text, images, and video. Output: text. Tool calling and structured JSON are documented. Thinking cannot be turned off on the Z.ai API; effort levels are low, high, and max.
• Weights: MIT-licensed on Hugging Face. Day-one serving in SGLang, vLLM, and TokenSpeed. This is a 320B MoE, so self-hosting is a datacenter job, not a laptop job.
• List price: $0.15 per million input tokens, $0.50 per million output, $0.03 cached input. A 50% launch promo ($0.075 / $0.25 / $0.015) runs through 9 September 2026 at 16:00 UTC. Budget against the list price.

Z.ai's own pitch is that Flash outperforms GLM-5.2 at about a tenth of GLM-5.3's token price, and that it approaches Claude Opus 4.8 on coding and agent benches. Those are company numbers. The independent picture is more mixed, and it is the one that should drive a routing decision.

The six-day stealth window

OpenRouter announced Ox Alpha on 20 August as a frontier reasoning model for coding, sustained agent work, and production use, with a 1M context window and text, image, and video input. A follow-up note said it was free, and that this time the provider would not train on prompts or completions. TechCrunch caught the speculation that followed: GLM, Xiaomi MiMo, even an unreleased Microsoft MAI. Stripe CEO Patrick Collison called it "very impressive." Nobody had a name.

The marketing sites that appeared that week, oxalpha.com and oxalpha.io, are third-party chat and API wrappers, not Z.ai. They still talk like the preview is the product. Treat them as untrusted front-ends. The named model lives on OpenRouter, the Z.ai API, the GLM Coding Plan, and Hugging Face.

OpenCode listed the free route as x-preview-f-free (or opencode/x-preview-f-free in config). OpenCode's own usage dashboard now files the traffic under GLM-5.3-Flash, formerly ox-alpha: as of 2 September, fourth by volume in their observed set, about 10% of traffic, 633,000 unique users, and an average session around 3.6 million tokens. That is a lot of agent loops for a model that was nameless for less than a week.

A stealth preview is a load test with an audience. The model you got for free is the model you will pay for, under a different slug, once the lab is done watching.

How people identified it before Z.ai did

The community got there first, with fingerprints rather than leaks. AI/ML API counted tokenizer test packs and matched GLM-5.2 and GLM-5.3 on every string they tried; MiMo and Qwen matched two. Other researchers reported a 44-of-44 and later 95-of-95 match on discriminating tokenizer probes. Malformed requests surfaced Java class paths and numeric error codes associated with Z.ai's serving stack. Video-token accounting lined up with GLM vision models. A reasoning_effort: "none" call reportedly returned Z.ai error code [1210]. Public GLM-5.3 is text-only, which is why the working theory during the week was an unreleased multimodal GLM variant. That theory was right.

Z.ai's launch post closed the loop: they served the anonymous traffic on Chinese AI accelerators, with a custom SGLang stack they say reached per-token cost comparable to mainstream NVIDIA GPUs after a 3× serving-performance gain on that hardware. They have not named the chip vendor.

Benchmarks, without the sweep

Two scoreboards circulated during the stealth week, and they do not describe the same test.

The viral one was a 10-task DeepSWE sample, the same table oxalpha.com still publishes: Ox Alpha 8/10 (80%), Fable 5 at 65%, GLM-5.3 at 62%, Grok 4.6 xhigh at 62%, GPT-5.6 Sol at 52%. Ox Alpha was scored pass/fail; the others were passes out of four attempts. That is a small, one-shot sample. Treat it as a teaser, not a ranking.

The fuller community run, 66 of 113 DeepSWE tasks (58.4%) under pier 0.3.1 and mini-swe-agent, is the number to remember from the preview. The official DeepSWE leaderboard at the time had Claude Opus 5 (max) around 74%, GPT-5.6 Sol (max) around 73%, Fable 5 (max) around 70%, and GLM-5.3 (max) around 69%. On that comparison, Ox Alpha sat near Claude Opus 4.8, not at the top of the table. Other independent sets were harsher: LiveCodeBench release_v6 at 49/175 (28.0% pass@1) with greedy decoding and no tools, and SlopCodeBench at 7/39 strict checkpoints (17.9%), behind GLM-5.3 on the same 39-checkpoint set.

Z.ai's own post-reveal table is stronger, as vendor tables usually are. On their published chart:

• DeepSWE v1.1: 63.4, ahead of GLM-5.2 (46.2) and Claude Opus 4.8 (58.0), behind GPT-5.6 Terra (69.6).
• Terminal-Bench 2.1: 84.3, a point behind Opus 4.8 (85.0) and further behind Terra (87.4).
• AutomationBench v1.0.6: 48.8, well ahead of GLM-5.2 (26.2) and Opus 4.8 (41.0).
• Toolathlon Verified: 78.4 vs GLM-5.2 at 59.9 and Opus 4.8 at 76.2.
• GDPval-AA v2: 1773 Elo, the top row they show, ahead of Opus 4.8 (1582) and Terra (1571).
• Artificial Analysis Intelligence Index v4.1.1: 57, at $0.045 per task on the discounted rate.

Z.ai's DeepSWE footnote uses mini-swe-agent, temperature 0.95, a six-hour timeout, and 400K context. The community 58.4% run was a leaner harness. Both numbers can be true. MindStudio's frontend write-up from the stealth week is still useful colour: testers liked typography, spacing, and Three.js / SVG work more than they liked unconstrained chat. Aniruddha Adak's week of real open-source chores on rebases, CI archaeology, and first-pass reviews matches what we would try it on first. Speed during the preview sat around 35–40 tokens per second in AI/ML API's timing, with large swings during the day. OpenRouter's current provider table for z-ai/glm-5.3-flash shows median throughput from the high teens to over 100 tokens per second depending on host. Pin a provider if latency is part of the product.

What it costs now

The $0 window ended with the name. Current rates, per million tokens:

• Z.ai API and the discounted OpenRouter / Novita / DeepInfra / GMICloud routes, through 9 September 2026 16:00 UTC: $0.075 input, $0.25 output, $0.015 cached input.
• List price after the promo, and the undiscounted OpenRouter hosts already: $0.15 input, $0.50 output, $0.03 cached input.
• GLM-5.3, the flagship, lists at $1.40 / $4.40. Flash at list is roughly a tenth of that, which is the comparison Z.ai wants you to make.
• GLM Coding Plan: Flash is in the subscription with 3× the quota of GLM-5.3. Off-peak and weekend calls consume half points.
• Self-host: MIT weights, your GPUs. Do not plan this on a single workstation.

That is still cheap next to Grok 4.6 at $2/$6 or Claude Opus 5 at $5/$25. It is not free, and a million-token prompt is a real invoice even at $0.15. Cache hits are doing a lot of work on OpenRouter's Z.ai route (about 90% in their recent mix), which is why the effective input price there is far below the sticker. Set a cache key. Stay under the context you actually need. Recheck the live OpenRouter listing before you quote a client a unit cost, because the promo has a date on it.

How to call it

Three routes cover most builders. All of them use an OpenAI-compatible Chat Completions shape.

• OpenRouter: model z-ai/glm-5.3-flash, base URL https://openrouter.ai/api/v1. Do not send stealth/ox-alpha.
• Z.ai API: model glm-5.3-flash. Docs recommend temperature: 1, top_p: 0.95, reasoning_effort: max, and thinking.type left enabled. For 1M context in Claude Code, Z.ai's coding-plan notes use the glm-5.3-flash[1m] suffix and CLAUDE_CODE_AUTO_COMPACT_WINDOW of 1000000.
• Self-host from zai-org/GLM-5.3-Flash via SGLang, vLLM, or TokenSpeed if you have the hardware and a reason not to send prompts to a lab.

OpenCode users who selected Ox Alpha Free during the preview should switch the model picker to GLM-5.3-Flash. Same weights, named vendor, metered bill. Keep a fallback model in config. Stealth routes disappear; named routes get repriced.

Privacy is better than it was. It is not settled.

During the preview, the data story disagreed by route. OpenCode Zen advertised zero retention and no training. OpenRouter's model page said the provider retained prompts and completions but did not train on them. OpenRouter's general stealth-program terms grant broader collection and training rights. Wavect's buyer guide had this right on day three: when a product page and a contract conflict, follow the stricter boundary.

The reveal replaces "unknown counterparty" with Z.ai, which is an improvement you can put in a vendor review. We have not seen a Z.ai statement about prompts sent during the anonymous week: whether they were kept, whether they touch training, whether deletion applies retroactively. If you pasted a private repo into Ox Alpha between 20 and 26 August, that question is still open. Going forward, use the named endpoint, read the current terms, and keep secrets out of any route you have not reviewed.

Skip oxalpha.com and similar no-login chat pages for anything you would not paste into a public gist. "We don't store chats" on a wrapper is not a data-processing agreement.

Where we'd use it, and where we wouldn't

The fit is high-volume agent work where cost per completed task decides whether the feature exists: repo mapping, conflicted rebases, CI archaeology, first-pass review, visual UI iteration when you can give it screenshots and a definition of done. Native vision is the part of the pitch we take seriously for frontend and design-to-code loops, because Z.ai trained for a generate-render-inspect cycle rather than a captioning bolt-on. That is also why chat-only evals undersold it during the preview.

• Good fit: coding-agent volume, screenshot-grounded UI work, and long-context chores where Kimi K3's open weights or DeepSeek V4 Flash would also be on the shortlist, and you want a named multimodal option at Flash-tier prices.
• Good fit: a self-hosted path, now that MIT weights exist, for teams that already run 300B-class MoE models and need the prompts to stay in-house.
• Poor fit: a single-model stack that has to win DeepSWE-shaped repo work against Opus 5 or GPT-5.6 Sol. The independent 58.4% run is not that win. Z.ai's 63.4 is closer, still behind Terra.
• Poor fit: anything regulated, customer-identified, or secret on a route whose retention you have not read. The stealth week is the cautionary example.
• Poor fit: quoting the $0 preview or the 50% promo as a durable unit cost. The list price is $0.15/$0.50, and it is dated 9 September.

Against the rest of the cheap-and-fast shelf: DeepSeek V4 Flash 0731 is still the price floor we reach for when instruction-following is under control. Grok 4.6 is the hosted agent we try when the job is long, non-visual, and already in Cursor. GLM-5.3-Flash is the one we would add when the job is visual-plus-code and the budget is Flash-shaped. Most production systems should keep using more than one model on purpose.

How we'd test it on your workload

Same method as every release. Take one real task from a live project, run it against the incumbent and against z-ai/glm-5.3-flash with the same harness, and count cost per completed task (including cache), how often a person had to step in, and how many failures were recoverable without a human. If the task is visual, add a fourth: whether the first pass is something you would show a client. Do that on a redacted copy of the repo. Do not discover the data policy in production.

If you're weighing an AI feature and want to know whether this price tier changes what's affordable, or whether this vendor should touch your data at all, that's a scoping conversation we're glad to have. Send us the constraint you're stuck on and you'll get a written scope and a fixed quote before any commitment.