GLM-5.3-Flash Is Ox Alpha: The Reveal, the Specs, and the Real Pricing

§ 01The reveal, hour by hour

One detail worth noticing: the Hugging Face repository’s creation timestamp is August 25 at 06:43 UTC, a full day before the reveal. This was a prepared launch, not a leak forced into the open. The six-day free window, in hindsight, looks exactly like what we guessed it was: launch marketing with the launch list pre-assembled.

§ 02What GLM-5.3-Flash actually is

Three architecture notes stand out from Z.ai’s own materials. First, it is the first natively multimodal model in the GLM-5 series: vision is not a bolted-on encoder for chat, it is wired into the coding loop, and Z.ai’s docs lean hard on “visual coding,” the model observing rendered interfaces and iterating against what it sees. Second, the hybrid attention design is the cost story: Z.ai calls it the first open-source frontier model to combine sparse and linear attention, and quantifies the win as 3.01x less attention computation and a 4.44x smaller KV cache than GLM-5.3. At a million tokens of context, KV cache is the bill, so that number is the entire economic argument. Third, the training corpus is 30T tokens. Keep that separate from the 16T figure that circulated during the stealth window: that was inference usage on one client in three days, not training data. Two different numbers that happen to sound alike.

§ 03What Z.ai claims, and what exists independently

Z.ai’s model card makes two headline claims: that GLM-5.3-Flash “outperforms GLM-5.2 across benchmarks and real-world workloads at one-tenth the price,” and that it approaches Claude Opus 4.8 on coding and agentic benchmarks. The card’s own chart lets you check both. These are Z.ai’s numbers, transcribed exactly:

Read against the claims, the chart is honest and, in places, undersells. The GLM-5.2 claim is unambiguous: Flash wins every row, with the two biggest jumps exactly where agent work lives, DeepSWE (46.2 to 63.4) and AutomationBench (26.2 to 48.8). The Opus claim is the modest one: “approaches” is accurate on Terminal-Bench, Agents’ Last Exam, and HLE, but on the chart’s own numbers Flash outscores Claude Opus 4.8 outright on DeepSWE, AutomationBench, and GDPval-AA. The vendor caveats still apply, these are Z.ai’s runs with Z.ai’s harness choices, and the card’s methodology footnotes (harnesses, temperatures, timeouts, per benchmark) at least make the runs reproducible. One reconciliation worth spelling out: the official DeepSWE 63.4 and the community stealth-window run’s 58.4 are both real, produced under different harnesses and budgets; Z.ai’s footnote specifies the mini-swe-agent harness with a 6-hour timeout and 400K context, more generous than the community setup.

The independent record is thinner but real, because the stealth window was itself a giant public evaluation. The completed 113-task DeepSWE community run resolved 58.4 percent with one attempt per task. Needle-retrieval tests held to roughly 934K tokens, so the million-token window is substance, not metadata. And the consistent shape of user reports held from day one: people who put it in a real agent harness with tools and a terminal rated it far higher than people who chatted with it. Whatever the official chart says, this is a model built for harness work.

§ 04What it costs now

The $0 era ended with the name. Here is the full picture:

For scale: GLM-5.3, the flagship, lists at $1.40 input and $4.40 output. Flash at list price is roughly a tenth of that, which is exactly the ratio Z.ai’s marketing leads with. Against the broader market, $0.15/$0.50 list positions it in the aggressive open-weight tier rather than anywhere near frontier-flagship rates, which is what the stealth-era GLM theory predicted the pricing would look like.

§ 05If you built on the free window

Three practical consequences, in order of urgency. First, the slug: stealth/ox-alpha is out of OpenRouter’s active catalog and no alias is documented, so anything calling the old model ID should switch to z-ai/glm-5.3-flash now rather than waiting for requests to fail. Second, the meter is on: the same workload that cost nothing on Monday costs real money today, and the 50 percent promo window ends September 9, so the price you calibrate against should be the list price, not the promo. Third, there is now a third option that did not exist during the window: the weights are MIT-licensed with day-one support in SGLang, vLLM, TokenSpeed, and KTransformers, so teams with datacenter GPUs can take the model in-house entirely.

§ 06The scoreboard on the stealth-era evidence

Worth a moment of respect for how this was called. Nobody leaked it: the identification was assembled in public from technical fingerprints, and every class of evidence held up. The 95-of-95 tokenizer probe match pointed at the GLM-5 vocabulary: correct. The API error envelopes matched Z.ai’s serving stack: correct, and that was operator evidence, not just lineage evidence. The video-token accounting matched GLM-5V-Turbo’s encoder: correct in direction, since the revealed model is natively multimodal GLM. The code-style fingerprints matched GLM-5.3: correct family. The competing Gemini theory, which ran on vibes and researcher winks, was wrong. The lesson for the next stealth model, and there will be one: tokenizers and error strings do not know how to lie.

§ 07The question the reveal did not answer

During the anonymous window, the retention terms disagreed by route: OpenCode said zero retention, OpenRouter’s stealth listing said the provider retains prompts and completions. The reveal converts “unknown counterparty” into “Z.ai,” which is a real improvement, because there is now a named company with published terms to hold to account. But as of reveal day we found no Z.ai statement about what happened to the prompts sent during the stealth window itself: whether they were retained, whether they touch training, whether deletion applies retroactively. If you sent sensitive code to a mystery box in exchange for free tokens, the mystery is solved and the question is not. That trade-off was the real price of the window, and it was on the label the whole time.

§ 08The layer above the reveal

Our standing argument through the whole Ox Alpha episode was that the harness layer, not the model layer, is where working AI gets decided, and reveal day is the cleanest demonstration of it. The model did not change on August 26. The same weights answered the same prompts. What changed was everything around it: the name, the price, the accountability, the slug your code points at. Teams whose work lived in the harness absorbed that as a config change; teams whose work lived on the free endpoint got a migration project. That layer is where we build, so discount accordingly: CellCog is the employee layer above models like this one, and GLM-5.3-Flash is now one more strong, honestly priced option in the rotation underneath it. We ranked the harness layer itself in our agent harness ranking, and the full stealth-window record stays preserved in the Ox Alpha explainer.