Z.ai's GLM 5.3 shipped on 18 August with a 1,310,720-token context window, larger than anything OpenAI or Anthropic offers, at $0.84 per million input tokens. Eight days later the Flash variant arrived at $0.15 with the same enormous window and a 943,718-token output ceiling. The largest context windows on the market are no longer coming from the labs that made long context a selling point.
What landed
GLM 5.3 costs $0.84 per million input tokens and $2.64 output, with a 1.31M context window and 131K maximum output, text only. GLM 5.3 Flash, on 26 August, costs $0.15 and $0.50, keeps the 1.31M window, adds image and video input, and raises the output ceiling to 943,718 tokens.
Qwen 3.8 Flash arrived the same day at $0.15 and $0.47, with a 1M context window and image and video input.
For comparison, Claude Opus 5 and the GPT-5.6 family sit at 1M and 1.05M context respectively, at several times the price.
Specifications are not capability
It would be easy to read that finding as "the cheap model is better", and it is not what it means. Context window describes how much the model will accept. It says nothing about how well it uses what it is given.
Models attend unevenly across long contexts. Information in the middle of a very long prompt is reliably harder for a model to use than information at either end, and this affects every model to some degree. A larger window raises the ceiling on what you can send; it does not guarantee the model made good use of it.
So the honest reading is narrower and still interesting: the structural limits are no longer where cost and capability are decided. If you need to put 1.2 million tokens in front of a model, your options are no longer restricted to the expensive end of the market. Whether the answer is any good remains a separate question you have to test.
What long context is actually for
Having worked with these windows for a while, our view is that most long-context use is a workaround for not having retrieval, and it is usually the worse option.
Sending a million tokens costs a million tokens on every turn. A retrieval step that finds the relevant 10,000 tokens costs a hundredth as much and frequently produces a better answer, because the model is not hunting for the signal.
Long context genuinely wins when the relevant information cannot be identified in advance: a question that could touch anything in a large codebase, a document where the answer depends on relationships between distant sections, an analysis where you do not know what matters until you have read everything. Those cases are real and they are a minority.
The useful rule: if you can describe what to retrieve, retrieve it. If you genuinely cannot, that is what the big window is for.
The open-weight angle
GLM and Qwen both publish weights, which matters beyond the price. It means the model can be run on your own infrastructure, that it cannot be retired out from under you, and that its behaviour cannot change without you choosing to update.
That last point is underrated. A hosted model can be updated by the provider, and your carefully tuned prompts can quietly start behaving differently. An open-weight model you host does exactly what it did yesterday. For anything where consistency matters more than having the newest thing, that is a real advantage.
On Deepest we offer these hosted alongside the frontier models, because for most people running inference is not the point. But the option existing at all changes the balance of power, and it is worth knowing which models give you that exit.
Frequently Asked Questions
Is GLM 5.3 Flash good enough to replace a frontier model?
For high-volume, well-specified work, it is worth testing seriously. For hard reasoning and tasks where subtle errors are costly, expect the frontier models to remain ahead. The only way to know where your work falls is to run both on your own prompts and compare.
Does a 1.31M context window mean I can stop using retrieval?
Almost certainly not. Retrieval is usually cheaper and often produces better answers, because a focused prompt beats a haystack. Treat the large window as an option for cases where you genuinely cannot predict what is relevant, not as a replacement for thinking about what to send.
What does "open weight" actually get me?
The ability to run it yourself, immunity from the provider retiring or silently updating it, and no dependence on their availability. What it does not get you is free inference: you still pay for the hardware, and for most workloads hosted inference is cheaper than doing it yourself.
