Ox Alpha bills $0 today. Its listing carries a setting that only two of OpenRouter's 422 models share, both of them GLM-5.3, and that setting is what made GLM-5.3 cost 55% more per task than GLM-5.2 on an identical rate card.

What OpenRouter actually publishes about it

Start with the primary source, because almost nothing written about this model so far has used it. OpenRouter serves its whole catalogue as unauthenticated JSON at /api/v1/models, and a second endpoint gives you the per-provider detail. Ox Alpha is in there under the id stealth/ox-alpha, and the record is more informative than the model card is.

Two of those rows do more work than the rest. The endpoint reports a discount of 0, which is OpenRouter's way of saying the zero is not a markdown: there is no list price sitting behind it waiting to reappear. And expiration_date reads 2098-12-31, a placeholder shared by four other listings, so the widely repeated claim that the free window closes around August 27 is not coming from OpenRouter.

That date is worth chasing down, because it is real and it is attached to the wrong thing. ExplainX's write-up of the launch reports that OpenCode Go made the model near-unlimited and free for six more days as of the evening of August 21, which lands on August 27. That is one tool's promotion of a model, not OpenRouter's listing of it, and the two have been collapsed into a single deadline as the story travelled. OpenRouter has published no end date at all. So the honest position is that the free window could close tomorrow or run into next year, and the only dated commitment anyone has made belongs to a third party.

Eight fields match GLM-5.3 exactly. The ninth is the interesting one.

The attribution question has been answered elsewhere with tokenizer fingerprints and stack traces, which are hard to reproduce and harder to check. There is a duller method that anyone can rerun in about a minute: take nine of the descriptive fields OpenRouter publishes for every model, and diff Ox Alpha against the catalogue. Against Z.ai's GLM-5.3, eight of the nine come back byte-identical.

Notice what that last row does to the popular conclusion. If Ox Alpha were GLM-5.3 wearing a mask, the modality would match too, and it does not: Ox Alpha accepts image and video input and GLM-5.3 accepts neither. So the evidence points at something more specific and more interesting than a rebadge. It looks like a multimodal sibling built on the same serving configuration, which is the kind of thing a lab ships as a point release rather than as a rename. We are stating that as the reading the data supports, not as a fact, and no lab has confirmed anything.

For pricing purposes the distinction matters more than it looks. A rebadged GLM-5.3 would arrive on GLM-5.3's card. A newer multimodal sibling has a decent history of arriving above it, and image and video input are usually billed on their own line entirely.

Three listings out of 422 force reasoning on at maximum

Here is the row that should change how you read the free preview. Ox Alpha's reasoning block sets mandatory to true and default_effort to max. Ninety of the 422 models on OpenRouter make reasoning mandatory, which is common enough. Setting the default effort to max as well is not: across the entire catalogue exactly three listings do it, and the other two are GLM-5.3 and the ~z-ai/glm-latest alias that points at it.

Reasoning tokens bill as output tokens. So a model that will not let you turn reasoning off, and that starts you at the most expensive setting it has, is a model whose bill is decided before you write a prompt. Ox Alpha does expose reasoning_effort and will accept high or low, so you are not stuck at max, but the floor is low rather than off, and the default is the ceiling.

We do not have to speculate about what that costs, because GLM-5.3 already ran the experiment in public. It charges exactly what GLM-5.2 charged and exactly what GLM-5.1 charged before that: $1.40 in, $4.40 out, three generations on one card. And yet Artificial Analysis measured the cost of taking its index from about $0.44 a task on GLM-5.2 to about $0.68 on GLM-5.3, a rise of 54.5%, because the newer model emits roughly 170 million output tokens across the suite against 140 million. Same price sheet, 1.21 times the tokens, 1.55 times the bill. GLM-5.2's reasoning is optional and defaults to high. GLM-5.3's is mandatory and defaults to max. That is the difference, and Ox Alpha has copied the expensive side of it.

A zero with nothing behind it

Twenty-two of the 422 listings are priced at zero on both legs. Eighteen of them carry a :free suffix, which is OpenRouter's convention for a throttled mirror of a model that also has a paid row. Eleven of those eighteen do have a paid twin sitting next to them, z-ai/glm-5.2:free and z-ai/glm-5.2 being the cleanest example. The free listing is the trial and the paid listing is the price, and you can read the second one before you commit to the first.

Of the four zero-priced listings without that suffix, two are Google's Lyria audio previews and one is OpenRouter's own router pseudo-model. Ox Alpha is the only zero-priced reasoning model on the platform that is neither a :free variant nor a media model, which is to say it is the only one where there is no companion row to tell you the price. That is not a small bookkeeping detail. Every other free model on OpenRouter either shows you what it costs or is something you would not put in a coding agent.

A month of agent work, priced on each card it could land on

Since nobody will publish a rate, the useful exercise is to price the same work on the cards that are actually in front of us. The workload below is an assumption and we have held it constant everywhere, so the only thing moving between rows is the price sheet: 12,000 requests in a month, 60,000 input tokens each with 70% of that served from cache, and 12,000 output tokens each. That comes to 720 million input tokens, of which 216 million are fresh and 504 million are cache reads, plus 144 million output tokens.

Output is 59.4% of that $1,067.04, which is the arithmetic consequence of the forced-reasoning setting rather than a coincidence. Run the same month at GLM-5.2's measured output volume instead, which is the closest thing we have to a reasoning-off comparison, and the same card produces $955.23. The $111.81 gap is what the max default costs you at this cache rate, and it grows as the workload gets more output-heavy.

One number in that table deserves a warning label. A single Ox Alpha response can run to 131,072 output tokens, and at GLM-5.3's $4.40 that one response is $0.5767, about 6.5 times the average request in the month above. Nothing stops a mandatory-reasoning model from spending its way toward that ceiling on a hard problem, and while you are testing at $0 you will not notice it happening.

Nothing in this family caches itself

The 70% cache rate in that model is doing a lot of work, so it is worth saying where it has to come from. Ox Alpha's endpoint reports supports_implicit_caching: false. So does GLM-5.3's. So does every single one of GLM-5.2's 33 endpoints, which we checked one by one rather than assuming. Caching on this family is something you build, not something that happens because you sent a similar prompt twice.

Drop the cache rate to zero and the GLM-5.3 month goes from $1,067.04 to $1,641.60. That is another $574.56 for the same work, and it is the most likely way a team gets surprised: you prototype against a free endpoint where cache misses cost nothing, you never build the prefix discipline, and then the meter starts. The teams that will be fine here are the ones already structuring prompts for cache reuse on whichever model they are on now.

GLM-5.2 has 33 endpoints. GLM-5.3 has one.

The spread between the second and fourth rows of that table is the single biggest unknown, and it is not about the model at all. It is about whether the weights come out. GLM-5.2's did, and the result is 33 endpoints from 28 providers selling the same model at input rates from $0.60 to $2.31 and output from $2.00 to $7.26. On the month modelled above that is $493.20 at Ambient against $1,777.25 at Alibaba. Same weights, same month, and the dearest endpoint takes more than three and a half times what the cheapest one does.

It is worth saying that cheapest input does not mean cheapest bill, because this is the mistake the spread invites. Sail Research sells the lowest input rate of the 33 at $0.50, and on this workload it finishes eighth, at $619.56, because its output rate is $3.15 against Ambient's $2.00. Once reasoning is mandatory, output is where the money is, and ranking sellers by their input column will pick the wrong one.

Z.ai's own GLM-5.2 endpoint, meanwhile, is not the cheap one. It bills $1,067.04 for this month, the identical figure its GLM-5.3 card produces, because Z.ai charges $1.40 and $4.40 for both. Of the other 32 endpoints, 18 come in cheaper, 8 land on exactly the same number and 6 sell above it. The resellers are mostly underneath the lab rather than above it, which is the opposite of the usual arrangement and worth knowing before you route to the vendor out of caution.

GLM-5.3 has exactly one endpoint, Z.ai's, running fp8. That is the same finding we published on August 22 and it has not changed in two days, though the GLM-5.2 seller count has: that post counted 21 providers and today the endpoint list returns 28, so the open-weight market is still growing underneath it. Ox Alpha also has exactly one endpoint, and its quantization field reads unknown. That is not unique, since 9 of GLM-5.2's 33 endpoints also decline to say, but those nine sit beside 24 that do, and beside a published price. Ox Alpha has neither a second endpoint to compare against nor a rate to compare, so there is no way to check whether the model you eventually pay for runs at the precision you evaluated for free.

Use the free window, but use it to measure

None of this is an argument against trying it. A million-token context with video input at no cost is a genuinely good offer and the uptime has held at 99.99% over the last day. It is an argument against letting the zero do your thinking. The thing that makes a free preview dangerous is not that it ends, it is that it teaches you nothing about the bill while you build the habits that determine it.

So log your token counts now, while they are free. Record input, cached input and output separately for every call, because on the nearest real card fresh input is 5.38 times a cache read and output is 16.92 times one, and an undifferentiated token total will not let you reprice anything later. Set reasoning_effort explicitly instead of inheriting max, and measure whether low actually loses you anything on your work: on GLM-5.3's card that choice is most of the difference between $955 and $1,067 a month, and possibly a lot more on output-heavy jobs. And if you are choosing between this and something with a published price, our calculator will price your own logged numbers against every card in the catalogue, which is a better basis for a decision than a zero with no end date.

Everything here came from two API endpoints, and you can rerun them

• OpenRouter: model catalogue API - Pulled August 24, 2026, 422 records. Source of the Ox Alpha listing in full, the created timestamp of 1787256295 which decodes to August 20, 2026 at 20:04:55 UTC, the $0 prompt and completion prices, the 1,048,576 context, the 131,072 max completion, the text-image-video modality, the mandatory reasoning block with max as its default, and the 2098-12-31 expiration placeholder. Also the source of every count in this post: 22 zero-priced listings, 90 with mandatory reasoning, 3 of those defaulting to max. This endpoint needs no key, so every number above is checkable in one curl
• OpenRouter: Ox Alpha endpoint detail - Read August 24. One endpoint, provider name Stealth, quantization unknown, discount 0, supports_implicit_caching false, uptime 99.9889% over the last day. The discount field is what establishes that the zero is a list price rather than a promotion, and it is the single most useful cell on the page
• OpenRouter: GLM-5.2 endpoint detail - Read August 24. All 33 endpoints, spanning $0.60 to $2.31 on input and $2.00 to $7.26 on output. Priced on this post's workload the range runs from Ambient at $493.20 to Alibaba at $1,777.25, a 3.60x spread, and Sail Research's category-leading $0.50 input rate finishes only eighth at $619.56. Every one of them reports supports_implicit_caching false. The equivalent GLM-5.3 call returns a single Z.ai endpoint at $1.40, $0.26 cached and $4.40, running fp8
• Z.ai: model pricing - The vendor card behind the GLM-5.3 column: $1.40 input, $0.26 cached input, $4.40 output, identical to GLM-5.2 and GLM-5.1. Used here as the price sheet Ox Alpha would inherit under the rebadge reading, which is the reading our own modality finding argues against
• ExplainX: OpenRouter ships Ox Alpha - Read August 24, and the only piece of coverage we found that tracks where the August 27 date actually comes from: OpenCode Go made the model near-unlimited and free for six more days as of the evening of August 21. It is also more careful than the summaries of it have been, describing the Zhipu attribution as serving-layer forensics rather than confirmation and warning readers not to architect compliance around a guessed identity. We disagree with none of that; we simply think the listing settles more than the forensics do
• TokenCost: GLM-5.3 bills the same $1.40 and $4.40 that GLM-5.2 billed - Our August 22 post, and the source of the two figures this one leans on hardest: Artificial Analysis measuring cost per task rising from about $0.44 to about $0.68 on an unchanged rate card, and the 140 million against 170 million output tokens across the index that explains it. That post found the one-seller problem first; this one finds the setting that causes the token growth
• Four things we could not establish, said plainly. Nobody has confirmed which lab operates the Stealth provider, and our metadata diff narrows it without settling it: eight identical fields is strong evidence of a shared serving stack, and the ninth field rules out the simple answer that it is GLM-5.3 itself. The August 27 end date circulating in other coverage appears in no OpenRouter field, and the nearest thing to a source for it is a six-day OpenCode Go promotion dated from the evening of August 21, which is a different company making a different promise. The monthly workload is our assumption rather than a measurement, deliberately held identical across all four cards so the comparison is about price sheets and not about token counts; your own mix will move every dollar figure and none of the ratios. And the $955.23 reasoning-off counterfactual borrows GLM-5.2's measured output volume as a stand-in for what Ox Alpha would emit at a lower effort setting, which is a reasonable proxy and not the same thing as running the test