GPT-6 Sol Pro vs GLM 5.3: What an Open-Weight Licence Actually Buys You

GPT-6 Sol Pro and GLM 5.3 are the two halves of a genuine fork in how to buy a reasoning model, and the licence is where the fork is. GPT-6 Sol — the vendor's model that the gpt-6-sol-pro configuration runs, generally available since September 22, 2026 at $2.00 per million input tokens and $10.00 per million output tokens — is a closed API. You rent it. GLM 5.3, the company's flagship, which reached the API on August 18, 2026 and published its weights in early September, is $1.26 input and $3.96 output per million tokens on our own model page, and you can download it.

This piece is not going to argue that one of those is correct. It is going to be precise about what the download actually gets you, because the answer is narrower than the open-weights discourse usually implies, and the part that is genuinely valuable is not the part people usually cite.

What each price line says

Both are vendor lists — OpenAI's and Z.ai's own numbers, passed through unchanged by the platforms that host them.

• Input — GPT-6 Sol $2.00 per million tokens vs GLM 5.3 $1.26 per million tokens

• Output — GPT-6 Sol $10.00 per million tokens vs GLM 5.3 $3.96 per million tokens

• Cached input — GPT-6 Sol $0.20 per million tokens vs GLM 5.3 at a 90% discount to its uncached rate, which is the same fraction of a smaller number

• Cache write — GPT-6 Sol $2.50 per million tokens at a single TTL vs GLM 5.3's cache-write tier, which is where the two rate cards stop being directly comparable because the vendors meter writes differently

• Long-context pricing — GPT-6 Sol reprices a request above its input threshold at a higher rate for the whole request; GLM 5.3's 1M-token window is billed at its standard rates throughout

• Context window — GPT-6 Sol 1,050,000 tokens vs GLM 5.3 1,000,000 tokens

• Maximum output — GPT-6 Sol 128,000 tokens vs GLM 5.3 128,000 tokens

• Input modalities — GPT-6 Sol text and image vs GLM 5.3 text only

• Weights — GPT-6 Sol closed, no download vs GLM 5.3 published on Hugging Face under Z.ai's own GLM-5.3 Licence

On list rates GLM 5.3 is roughly 1.6× cheaper on input and 2.5× cheaper on output. That is a real gap and it is not a decisive one. The line that changes decisions is the last one.

What the licence actually says, and what it does not

Here is where the open-weights conversation usually goes wrong. The claim "it's open weights, so you're safe from vendor risk" is true in one specific sense and false in another, and the difference is worth spelling out.

GLM 5.3's weights are published. The licence is Z.ai's own GLM-5.3 Licence, not MIT — the previous generation, GLM-5.2, was MIT, and the change is the whole story. The new licence permits commercial use and fine-tuning. It adds a condition: a model-as-a-service provider with more than $10 billion in revenue over any twelve-month period must pass a Z.ai security review before offering the model as a service.

That condition tells you what the licence is for. It is not a restriction aimed at you. It is aimed at the handful of hyperscalers who could otherwise resell Z.ai's work as their own product without contributing to it. If you are running GLM 5.3 inside your own product, or fine-tuning it on your own data, or serving it to your own users, the clause does not touch you.

What it does mean is that the licence is a vendor-controlled document rather than an irrevocable grant. MIT cannot be revised. A custom licence can be revised for future versions, and the terms you accepted for 5.3 are not a promise about whatever Z.ai numbers next. That is a smaller risk than "the vendor can retire the model" — the weights you downloaded do not stop working — but it is not nothing, and no open-weights comparison page mentions it.

The genuinely valuable part of the download is narrower and more concrete than the slogan: it is that the checkpoint you evaluated is the checkpoint you serve, permanently, regardless of what the vendor's roadmap or rate card does next. That is a reproducibility guarantee, not a cost guarantee.

What the independent record shows

Artificial Analysis is the neutral harness both models appear on, and it re-based its index on September 7, 2026 — swapping one terminal benchmark for another and reweighting the private-task component. That re-basing is why the most-quoted GLM 5.3 figure in circulation is retired: the model was published at 60 on the previous index and reads 45 on the current one, and the move is a change of harness rather than a change of model.

On the current board, GPT-6 Sol scores 48 at max effort at $1.06 per index task. GLM 5.3 scores 45 at $2.01 per index task. Three points of index separation is the honest summary, and it comes with an unusual wrinkle: on this pair the cheaper model is the closed one. GPT-6 Sol is roughly half the cost per completed task, which is the opposite of the assumption most open-weights comparisons start from.

Two further caveats belong attached.

• The open-weights ranking is a category result, not a general one. GLM 5.3's position is a statement about the field of downloadable models as much as about GLM 5.3.

• Most of the benchmark figures circulating for GLM 5.3 are Z.ai's own. The independent numbers are the index score and the per-task cost; the Terminal-Bench, DeepSWE and cyber-capability figures on the model card are vendor-run and should be labelled that way when quoted.

One point in GLM's favour that the per-task figure understates: the cheaper sibling. GLM 5.3 Flash is a 320-billion-parameter model with 18 billion active parameters, MIT-licensed, at $0.15 input and $0.50 output per million tokens, and it scores 42 on the same index at $0.25 per task. That is a quarter of GPT-6 Sol's per-task cost for six index points, and it is the model on this comparison that actually sits on the index's cost-efficiency frontier.

The cost argument for self-hosting is weaker than the cost argument against it

It is tempting to frame the download as a cost play. At $1.26 and $3.96 per million tokens, the API is already cheap enough that a 753-billion-parameter mixture-of-experts model is not obviously cheaper to run yourself — and GLM 5.3 is a 753B MoE, which is not a single-GPU deployment under any quantization scheme that preserves its behaviour. The break-even is a utilization question, and for most teams the honest answer is that it never arrives.

The reason to take the download is the one the licence section implies: control over where inference happens. Self-hosting answers a data-residency question by where you run the model rather than by what a vendor's contract says, and for regulated workloads that is the only version of the answer that holds. It also permits fine-tuning on proprietary data with no data leaving your infrastructure, which a closed API does not offer at any price.

If neither of those applies to you, the API is the correct choice and the licence is irrelevant. That is a legitimate conclusion and it is the one most teams should reach.

Where a routing layer changes the arithmetic

GLM 5.3 is on OrcaRouter's catalogue at $1.26 input and $3.96 output per million tokens, with a 1,000,000-token context window, a 128,000-token output cap, a p50 time to first token of 3.99 seconds as measured on our own model page. Provider list pricing passes through with no markup on top, so a Z.ai price change is live on our side the same day it is live on theirs.

GPT-6 Sol is not one of our routes, so nothing here is a claim about its price through us.

The reason a single endpoint matters in this specific matchup is that the two models are not substitutes so much as tiers. GLM 5.3 is text-only; GPT-6 Sol takes images. GLM 5.3 is downloadable; GPT-6 Sol is not. Neither of those is a price decision, and neither is settled by a benchmark. What a team actually wants here is one key that can reach the cheap text model for the volume and a different model for the calls that need vision, without a second contract, a second SDK and a second set of rate limits. That is a routing problem, and the failover behaviour matters for the same reason it always does — the cheap tier your pipeline depends on should not be the tier that takes the whole pipeline down.

The decision rule

• Volume text work at a known quality bar — GLM 5.3. Three index points below GPT-6 Sol at roughly 40% of the output rate, on the same context window and the same output ceiling.

• Anything that needs to read an image — GPT-6 Sol, or neither of these. GLM 5.3 does not accept images on the API, and no price advantage survives a modality you cannot use.

• Regulated data, or a workload you must be able to reproduce in two years — GLM 5.3, self-hosted from the published checkpoint. Not because it is cheaper, but because the checkpoint is fixed and the residency question is answered by your own infrastructure.

• Fine-tuning on proprietary data — GLM 5.3. A closed API offers no equivalent at any price.

• A team with no residency constraint and no fine-tuning plan — GPT-6 Sol, and treat the licence as a non-issue. It is a genuine non-issue for you, and pretending otherwise is how open-weights arguments lose credibility.

The licence is not a cost lever. It is a control lever, and it is worth exactly as much as the control is worth to you — which for some teams is the entire decision and for others is nothing at all.