GLM-5.3-Flash Revealed: The Anonymous "Ox Alpha" Was Zhipu's Open Multimodal Frontier Model All Along

The model that spent a week atop third-party coding-platform usage charts finally has a name and a price. On August 26, 2026, Zhipu AI confirmed that the free "Ox Alpha" model developers had been hammering since August 20 is GLM-5.3-Flash — a 320-billion-total, 18-billion-active (320B-A18B) mixture-of-experts model, the first natively multimodal release in the GLM-5 series, and one of the cheapest frontier-grade models on any rate card at launch. Zhipu is pricing GLM-5.3-Flash at a tenth of its flagship GLM-5.3 API rate, or a twentieth during a limited-time discount, and the open weights — MIT-licensed — went to Hugging Face the same day. Unlike GLM-5.3, whose weights are still pending for the end of the month, this model is something you can self-host this week, not next.

Here is the short version: Zhipu open-sourced its cheapest large model yet, it beats its own GLM-5.2 across benchmarks despite running a fraction of the active parameters, and the vendor puts its Artificial Analysis Intelligence Index score at 57 — matching Claude Opus 4.8, per Zhipu's launch materials. Every benchmark figure attached to this launch is vendor-reported and unreproduced as of writing; the prices and the parameter counts are current fact.

What actually shipped

GLM-5.3-Flash is a 320B-total / 18B-active MoE with 45 layers, which puts its active footprint at a little over half of GLM-5.2's ~32B active parameters. The headline capability is modality: it natively takes text, image, and video input — the first GLM-5 model to do so — rather than routing vision through a separate adapter. Context goes to 1 million tokens, and Zhipu claims long-context serving cost lands at roughly a third of GLM-5.3's.

On the architecture side, Zhipu describes it as the first open-source frontier model to use a hybrid sparse-attention-plus-linear-attention design, paired with Manifold-Constrained Hyper-Connections (mHC) for scaling and an IndexPool mechanism that compresses four indexer key vectors into one weighted pool to cut long-context latency. Pretraining ran on a 30-trillion-token multimodal corpus. None of this is independently audited yet — it is Zhipu's description of its own model — but the serving-efficiency claims are specific enough to test once the weights are in front of the open-source inference stack.

The launch-day spec sheet, at a glance:

• Parameters — 320B total / 18B active, 45 layers, MoE.

• Modality — native text, image, and video input; text output.

• Context window — up to 1,000,000 tokens.

• License — MIT, open weights live on Hugging Face at launch.

• Price — $0.15 / $0.50 per million tokens (input / output), $0.03 per million cached input; a 50%-off promo through September 9, 2026 cuts that to $0.075 / $0.25 / $0.015. At list that is roughly a tenth of GLM-5.3's $1.40 / $4.40.

The Ox Alpha week

The reveal closes a week of speculation. On August 20, an anonymous model appeared on a third-party AI API platform with a 1-million-token context window, text/image/video input, and a price of zero, and it quickly became the most-called model on the platform's weekly charts. The community fingerprinted it back to Zhipu's GLM family within 48 hours — the same anonymous-then-claimed pattern that had previously identified models from other Chinese labs — and analysts widely guessed a Flash-variant GLM-5.3. The August 26 announcement confirmed the guess and added the detail that the free week was an intentional test: Zhipu says the anonymous run set call-volume records on the coding platforms where it was offered, with all traffic served from domestic Chinese chips.

That free-week context matters for pricing expectations. A model given away at $0 for a week, then launched at a tenth of the flagship's rate with a limited-time twentieth-price discount, is a deliberate market-making move in the budget tier — and the "limited-time" qualifier is the part to watch. The discount is a pricing decision that can be walked back; the MIT open weights cannot.

What it costs, in actual dollars

Zhipu's rate card is out in actual dollars, not just ratios: $0.15 per million input tokens, $0.50 per million output tokens, and $0.03 per million cached input tokens. Against GLM-5.3's published $1.40 / $4.40, that lands at roughly a tenth at list, and a 50%-off launch promo running through September 9, 2026 brings it to $0.075 / $0.25 / $0.015 — roughly a twentieth. Zhipu also puts the model's cost per AA Intelligence Index task at about $0.045, vendor-reported, which is the number behind the "1/40 of Opus 4.8" framing. For a model that scores in the same band as models priced 10–40x higher, the headline economics are real; the fine print is that the launch discount is temporary and per-task cost depends on how verbose the model turns out to be in production.

The efficiency story is where Zhipu claims the cost advantage comes from. Against GLM-5.3, the vendor reports attention compute down 3.0x and KV cache down 4.4x, and claims the lowest attention compute of the compared budget-tier models — GLM-5.3, DeepSeek V4 Flash, and Kimi K3 — while conceding its KV cache is still slightly larger than DeepSeek V4 Flash's and Kimi K3's. On the serving side, Zhipu reports a 3x end-to-end serving-performance improvement over baseline on domestic Chinese chips, reaching per-token cost it calls comparable to mainstream NVIDIA GPUs. Those are all vendor-reported and none are independently reproduced yet; they are the right numbers to check against the open weights.

Why the reveal matters

Strip the benchmarks away and the launch still changes the open-model landscape in two ways. First, it is an open-weights multimodal model in the frontier band at a budget price — the combination of MIT license, 1M context, native image/video input, and single-digit-cent per-task cost has no direct equivalent from the US frontier labs, whose equivalent multimodal models cost an order of magnitude more and ship closed. Second, it undercuts Zhipu's own flagship: GLM-5.3 is the 743B model that scored an independently measured 60 on the AA Intelligence Index this month, and GLM-5.3-Flash is positioned as "frontier intelligence, flash cost" against that same family. Zhipu's story is that a smaller, cheaper model can capture most of the flagship's capability — which, if the vendor's coding and agentic claims hold up, is the same post-training-economy argument the industry has been circling all year.

One caveat keeps this in perspective. GLM-5.3-Flash's AA Index score of 57 comes from Zhipu's own launch materials, not yet from the Artificial Analysis leaderboard, and the coding comparison to Claude Opus 4.8 is Zhipu's internal Z.ai Code Bench evaluation. Treat the 57 and the coding parity as claims until independent measurement lands — the model was widely tested anonymously, so third-party numbers should arrive quickly.

Try it, and how the routing layer fits

Day-one access is via Zhipu's own API, the open weights on Hugging Face (zai-org/GLM-5.3-Flash, MIT), and Zhipu's coding products — ZCode and the GLM Coding Plan, which includes a batch of daily experience cards. For self-hosting, the MIT license plus vLLM and SGLang support means the serving-efficiency claims above are directly checkable.

On the routing side, GLM-5.3-Flash itself is not on OrcaRouter's roster yet as of this update. The rest of the GLM family is: GLM-5.3 went live on our side the same day Zhipu's API opened, at Zhipu's list price with 0% markup passed through. That pass-through is exactly the mechanism that matters for a launch like this one — a vendor price change, whether the discount sticks or the rate card moves later, shows up on our side the same day, no renegotiation. If you want to run the Flash against the rest of the GLM line on one key once it lands, that is the setup; until then the weights on Hugging Face are the fastest way to test it yourself.

What to watch next

Three things settle the real story in the next couple of weeks. First, an independent Artificial Analysis measurement of the 57 — the model was already running in the wild as Ox Alpha, so the leaderboard should catch up fast. Second, whether the 50%-off promo — currently set to end September 9, 2026 — is extended past that date, since that decides the Flash's steady-state cost. Third, the GLM-5.3 weights, still promised for around August 28 — the same week the Flash's MIT weights went live, which would give the open-weights community two tiers of the same family to compare at once.