What was the Ox Alpha model?

Ox Alpha was a stealth-tested model that appeared free in coding tools during late August and early September 2026. Z.ai revealed it as GLM-5.3-Flash, the natively multimodal small model in the GLM-5 family released on August 26, 2026. Stealth releases like this let a vendor gather real-world usage traces at scale before committing to a price, and the anonymity keeps early impressions free of brand expectations.

How much does GLM-5.3-Flash cost?

Z.ai lists GLM-5.3-Flash at $0.15 per million input tokens, $0.03 per million cached input tokens and $0.50 per million output tokens. A 50% launch promotion halves those to roughly $0.075 / $0.015 / $0.25 through September 9, 2026 at 24:00 UTC+8. Some trackers list slightly different figures near $0.071 / $0.238, reflecting currency conversion and reseller spreads — cite the source you actually billed against.

Is GLM-5.3-Flash good enough to replace a frontier model?

For scoped, high-volume work yes; for open-ended agent work no. It posts an Artificial Analysis Intelligence Index of 59.4, which is competitive with much more expensive models, and it is currently among the cheapest capable APIs tracked. But frontier models still lead on long-horizon agent suites, and cross-vendor benchmark claims from Chinese labs frequently use vendor harnesses that do not reproduce elsewhere. Pilot it on your own tasks before committing.

Does GLM-5.3-Flash have open weights?

Not confirmed at the time of the reveal. Z.ai has staged open-weight releases for the GLM-5 family rather than shipping weights at launch, and GLM-5.3 itself followed that pattern in August 2026. Treat GLM-5.3-Flash as an API-only model for planning purposes until Z.ai publishes a weights release, and do not build a self-hosting roadmap on an unshipped promise.

Quick Answer

Ox Alpha Revealed as GLM-5.3-Flash: Price and Specs

Published: September 5, 2026

The Short Answer

Ox Alpha was GLM-5.3-Flash. The anonymous model that showed up free inside coding tools in late August and early September 2026 was Z.ai’s small natively-multimodal GLM-5 model, released August 26, 2026 and identified publicly in early September.

The free window is over. Real pricing:

Tier	Input	Cached input	Output
List	$0.15	$0.03	$0.50
Promo (through Sep 9, 2026)	$0.075	$0.015	$0.25

The promotion ends at 24:00 UTC+8 on September 9, 2026. After that the list rate applies.

Last verified: September 5, 2026.

What a Stealth Release Actually Is

Ox Alpha followed a pattern that became routine in 2026: a vendor ships an unnamed model into third-party coding tools at zero cost, collects real usage traces at scale, then reveals the name and switches on billing.

The mechanics are deliberate. Anonymity strips brand priors — developers evaluating “Ox Alpha” have no expectations about a Chinese lab, an American lab, or a specific model family, so the feedback is closer to a blind test. And free usage generates the one dataset that benchmarks cannot: what people actually ask a coding model to do, in their own repositories, at volume.

The cost to the vendor is inference on a large free tier for a couple of weeks. The return is a pricing decision informed by real demand elasticity rather than a spreadsheet. Reporting around the reveal indicated Ox Alpha was served on roughly 100,000 domestic Chinese accelerators, which is its own signal about how far the non-NVIDIA serving stack has come.

What GLM-5.3-Flash Is

It is the small, fast, natively multimodal member of the GLM-5 line — the cheap tier beneath full GLM-5.3, which Z.ai released on August 14, 2026 at $1.40 / $4.40 per million tokens.

The gap between the two is roughly 9x on input and 9x on output. That is a much wider flagship-to-flash spread than Western vendors run, and it is the core of Z.ai’s positioning: the flash tier is priced to win high-volume work outright rather than to upsell.

On the Artificial Analysis Intelligence Index, GLM 5.3 scores 59.4 — for context, Claude Fable 5.1 is 65.6, Claude Opus 5 is 63.0, Gemini 3.8 Flash is 58.7, and Qwen 3.8 Max is 57.7. The GLM family is genuinely in the conversation, not a discount curiosity.

How the Price Compares

Against the other cheap tiers as of September 2026:

Model	Input	Output	Notes
GLM-5.3-Flash (promo)	$0.075	$0.25	Ends Sep 9, 2026
GLM-5.3-Flash (list)	$0.15	$0.50	Cached input $0.03
DeepSeek V4 Flash	$0.22	$0.66	Off-peak; peak is 2x
GPT-5.6 Luna	$0.20	$1.20	Cut 80% Jul 30, 2026
Qwen 3.8 Flash	$0.15	$0.47	Open weights
Gemini 3.8 Flash	$0.75	$3.75	⚠️ Doubles Jan 1, 2027

At list price GLM-5.3-Flash is matched by Qwen 3.8 Flash on input and slightly beaten on output, but Qwen ships open weights while GLM-5.3-Flash does not — which one wins depends entirely on whether self-hosting is on your roadmap.

⚠️ Numbers vary by source. Some trackers list GLM-5.3-Flash around $0.071 / $0.238, which is close to the promo rate and reflects currency conversion plus reseller spreads. Z.ai’s own listing is the $0.15 / $0.50 figure. Cite whichever source your invoice comes from — vendor and reseller rates for Chinese models have diverged repeatedly through 2026.

What Changed on September 9

Nothing about the model — only the bill. If you adopted Ox Alpha during the free window and left it in a pipeline, your cost per million output tokens went from $0 to $0.25 on September 9, 2026, and to $0.50 when the promotion lapsed.

That is still cheap. But “free model quietly becomes a billed model” is a recurring 2026 failure mode, and the damage is proportional to volume. If you wired a stealth model into a high-throughput job, check the model string in your config now rather than at the end of the billing period.

Should You Use It?

Good fit:

High-volume scoped work — classification, extraction, summarization, boilerplate generation
Multimodal input at flash-tier prices, since it is natively multimodal rather than a text model with vision bolted on
Cost-floor workloads where per-token price genuinely dominates

Poor fit:

Long-horizon autonomous agents, where frontier models still lead decisively
Anything with data-residency constraints that exclude Chinese API endpoints
Roadmaps that require self-hosting, until weights actually ship

⚠️ A standing caution on benchmarks: claims that a cheap model beats a frontier model on agent benchmarks are common and frequently rest on vendor-run harnesses that do not reproduce in independent setups. The Intelligence Index figure of 59.4 is third-party and credible. Vendor comparison tables should be treated as marketing until someone else reproduces them.