DISPATCH29 AUG 2026

GLM-5.3-Flash — a 1M-context multimodal model under MIT at $0.15 per million

DATELINE29 AUG 2026 — Bengaluru

The anonymous free 1M-context model developers had been hammering since 20 August turned out to be Z.ai's GLM-5.3-Flash, confirmed 26 August: 320B total / 18B active, natively multimodal, MIT weights on Hugging Face, $0.15 per million in. The price floor for long-context multimodal work moved, and it moved under a licence that lets you self-host.

1. 01

   "Ox Alpha" — free, 1M context, no claimed owner — was GLM-5.3-Flash all along

   It appeared on a third-party API platform on 20 August at a price of zero and developers spent six days benchmarking a model nobody would admit to. Z.ai confirmed it on the 26th. The stealth launch is a distribution tactic worth noticing on its own: six days of unpaid, unbiased evaluation before anyone knew whose model to be sceptical of.

2. 02

   320B total / 18B active, natively multimodal, 1,048,576-token context, MIT weights

   Text, image and video in; 131,072 max output; hybrid sparse-plus-linear attention. Day-one serving support in SGLang, vLLM, TokenSpeed and KTransformers, which is the detail that decides whether "open weights" means anything in practice.

3. 03

   $0.15 per million in, $0.50 out — half that until 9 September

   Cached input is $0.03 per million. Whatever you make of the benchmarks, that is the number your existing cost model now has to justify itself against, and the promo ends in under two weeks.

Two things happened here and the second one is the story.

First, a stealth launch: on 20 August a model with a 1M-token context window, image and video input, and a price of zero showed up on a third-party API platform under the name Ox Alpha, with no claimed owner. Developers spent six days hammering it. On 26 August Z.ai confirmed it was theirs.

That's a distribution tactic worth stealing, incidentally — six days of enthusiastic unpaid evaluation from people who had no brand to be sceptical of yet, and by the time the name was attached the reputation was already built by strangers.

Second, and more consequentially: the confirmed pricing and licence.

What actually changed

Not your reasoning path. A 320B/18B mixture-of-experts model at $0.15 per million input tokens is not the thing you put in front of a decision that matters, and the benchmark table won't tell you otherwise in a way you should trust.

What changed is the economics of the high-volume, low-stakes layer — and in most real products that layer is the majority of tokens by a wide margin:

• Classification and routing of incoming text.
• Extraction from documents where a schema validates the output anyway.
• First-pass drafting that a human or a second model will revise.
• Summarising long context where the cost of the context window was the entire reason you weren't doing it.

That last one is the interesting case. A 1M-token window at fifteen cents per million makes "just send the whole thing" viable for workloads where you previously had to build a retrieval layer to avoid the bill. Sometimes that's a genuine simplification and sometimes it's a trap — a retrieval layer is also what makes the output attributable, and dumping a million tokens in gives you an answer you can't trace.

The two-tier architecture — cheap model for volume, frontier model for the calls that matter, with an explicit rule for which is which — stops being an optimisation you'll get to later and starts being the obvious default.

Open weights, and what that's worth

MIT on Hugging Face is unusually permissive for a model this size, and the day-one support in SGLang and vLLM means self-hosting is a real option rather than a licence-page technicality.

Where that matters: data residency, air-gapped deployments, anything where "we send customer data to a third-party API" is the blocker rather than the cost. Those constraints didn't previously have an answer at this capability level, and now they do.

Where it doesn't: your margin. 320B total parameters is not cheap to serve, even at 18B active. Unless you have sustained volume and someone who knows how to run inference infrastructure, the hosted price at fifteen cents is almost certainly cheaper than your own GPUs plus the person minding them. Self-hosting is a compliance answer here far more often than a cost one.

Also worth being precise about: open weights is not open data. You can run it, modify it and ship it. You can't audit what it was trained on — 30 trillion tokens of undisclosed corpus — and for some regulated buyers that's the question that actually gets asked.

What to actually do

Take one high-volume, low-stakes task you're currently paying frontier prices for. Pull fifty real inputs from your own logs. Run both. Compare on your own criteria, not on the benchmark table.

That's an afternoon, and it either saves you a recurring bill or it tells you the cheap tier isn't good enough for that specific job — which is also worth knowing and is not something a leaderboard can tell you. Do it before 9 September if you want to test at the promotional price, though don't build a business case on a promo; the list price is the number that has to work.

The week's other half

This and Claudeforce are the same story from opposite ends. One says the value is moving up, into the system of record and the governance boundary. This one says it's draining out of the inference itself, because a 1M-context multimodal model is now MIT-licensed and costs fifteen cents a million.

Both are bad news for a product whose entire substance is the call in between.

Sources: MarkTechPost on the release · launch pricing, promo window and open weights.