How Many Free Tokens Does the Z.ai Startups Program Give?
The Z.ai Startups Program grants approved startups up to 1 billion free GLM API tokens.
Most vendor startup programs are denominated in dollars. Zhipu AI, the lab behind Z.ai and the GLM models, denominates this one in tokens instead. That single difference makes the offer awkward to compare on a spreadsheet, and it is probably why the grant gets so little attention relative to its size.
Alongside the tokens, Z.ai lists volume pricing, priority rate limits and technical support for approved companies. AI Perks tracks the current terms next to the rest of the vendor programs.
This page does the conversion work, and labels any figure that is reported rather than confirmed on Z.ai's own pages.
GLM-5.1, 5.2 and 5.3: What Shipped and When
Zhipu AI shipped three flagship GLM releases in about four months, and the version numbers make them easy to confuse.
| Release | Date | What it is |
|---|---|---|
| GLM-5.1 | 8 April 2026 | Open-source flagship agentic model, stronger coding, vision support and file input |
| GLM-5.2 | 13 June 2026 | 744B-parameter mixture-of-experts flagship, fully open source under a plain MIT license |
| GLM-5.3 and GLM-5.3-Flash | 14 August 2026 | Reasoning flagship pitched at coding and cyber defense, plus a cheaper, faster Flash sibling |
One secondary source dates GLM-5.3 to 17 August rather than 14 August. The discrepancy has not been reconciled, so treat the exact day as unsettled.
GLM-5.3 carries the tagline "Built to Code. Ready for Cyber Defense." and is post-trained on a 743B-parameter base. Parameter counts around this family move between sources by a billion or two, so the practical advice is to quote the model card in the weights repository rather than a launch write-up. Context window and maximum output length are worth reading off Z.ai's own documentation at the moment you need them, because they are the specifications most often restated incorrectly second-hand.
The useful pattern here is the cadence, not the version numbers. Three flagship releases in roughly four months means any model you lock in today is one to expect to revisit inside a quarter, which is an argument for keeping the model name in configuration rather than scattered through your code.
What GLM Costs at List Price
GLM-5.1, GLM-5.2 and GLM-5.3 all cost exactly the same: $1.40 per million input tokens, $4.40 per million output tokens, and $0.26 per million cached input tokens.
These figures come from the official Z.ai pricing documentation.
| Model | Input / 1M | Output / 1M | Cached input / 1M |
|---|---|---|---|
| GLM-5.3 | $1.40 | $4.40 | $0.26 |
| GLM-5.2 | $1.40 | $4.40 | $0.26 |
| GLM-5.1 | $1.40 | $4.40 | $0.26 |
| GLM-5.3-Flash | $0.15 | $0.50 | $0.03 |
The upgrade path from GLM-5.1 to GLM-5.3 is free in pricing terms. If you are still calling GLM-5.1 out of habit, you are paying the GLM-5.3 rate for the older model, which is a one-line change in most codebases and the cheapest upgrade on this page.
GLM-5.3-Flash deserves a second look: at $0.15 input and $0.50 output, it runs at roughly a tenth of the flagship rate on both sides of the meter. Compare that against the rest of the market at AI Perks.
Cached input at $0.26 per million is the number most teams ignore, and it decides whether a long system prompt is affordable to resend on every turn. Prices move, so re-read the pricing page before committing a budget to any figure here.
Where the Flash Tier Fits
The gap between GLM-5.3 and GLM-5.3-Flash is roughly ten to one, which makes model routing worth more than prompt golfing.
A billion tokens sounds unlimited until an agent loop starts retrying. The practical consequence is a sequencing decision rather than a single model choice: put prototyping, evaluation harnesses, classification, routing and other low-stakes background work on the cheaper Flash tier, and reserve the flagship for traffic that genuinely needs the stronger reasoning.
Done properly, that split lowers the bill on the paid path and stretches a token-denominated grant across far more real usage than a single-model deployment would.
Rate limits are one of the things the startup program explicitly improves for approved companies, and we could not confirm a published limit table for either tier outside the program. Build the fallback path before you depend on one. AI Perks tracks which vendors publish their limits and which leave them to support tickets.
What a Billion Free Tokens Is Worth in Dollars
Between $1,400 and $4,400 at GLM-5.3 list pricing, depending on how the grant splits between input and output tokens. Z.ai has not published that split.
The arithmetic, using the verified list prices above:
- 1B tokens, all input: 1,000 x $1.40 = $1,400
- 1B tokens, 70% input and 30% output: $980 + $1,320 = $2,300
- 1B tokens, all output: 1,000 x $4.40 = $4,400
A realistic agent or coding workload lands near the middle case, so roughly $2,300 in equivalent list-price value is the honest number to plan against, not the headline billion.
Spent on GLM-5.3-Flash instead, the same billion tokens is worth closer to $150 to $500, because the model is cheaper. That inverts the usual instinct: with a token-denominated grant, you extract the most value by running it on the most expensive model, which is the opposite of how dollar-denominated credits behave. AI Perks is built around exactly these comparisons.
Read the License File, Not the Launch Coverage
GLM-5.2 is fully open source under a plain MIT license. That is confirmed. Do not assume the same terms carry forward to every later release in the line.
The GLM family is genuinely open weight, which is most of why it is interesting to a startup: you can self-host, fine-tune and keep inference on infrastructure you control, so the API grant becomes a convenience rather than a dependency. But "the GLM models are MIT" is a generalization, and a generalization is not a license.
Coverage of a new release routinely carries the previous version's licensing framing forward, because it is the fastest sentence to write and nobody checks. We could not confirm the exact license attached to the GLM-5.3 weights, so this page does not state one. Open the license file in the weights repository for the specific model and the specific version you intend to ship, and have legal read that file rather than a blog post, this one included.
Licensing terms move as often as prices do, which is why AI Perks re-checks them rather than archiving a launch-day snapshot.
What Z.ai Has Not Published
Three things materially change what this offer is worth, and none of them are answerable from Z.ai's public pages today.
Whether the grant expires. No expiry window is published for the startup token grant. An unexpiring billion tokens and a 90-day billion tokens are very different products, and the second one should change your rollout order: you would front-load the expensive traffic instead of saving the grant for later.
Whether the grant is model-restricted. Nothing states whether the tokens can be spent on the current flagship or only on older models, which is the difference between roughly $2,300 and a few hundred dollars of equivalent value. Ask before you plan a budget around the higher number.
The new-user free tier. New Z.ai accounts are reported to receive 5 million free tokens plus a 100 yuan API credit. That is a third-party report, not confirmed on Z.ai's own pages, so treat it as unverified and do not put it in a forecast.
Separately, and also as reported rather than confirmed, Baseten added GLM-5.3 to its Model APIs in late August 2026 and a faster variant in early September. Baseten's pricing table lists GLM-5.3 input at $1.40 per million, matching Z.ai's own rate, but the output figure rendered inconsistently across fetches, so verify it directly on Baseten's pricing page before relying on it. Baseten's standing new-workspace offer is reported as $30 in trial credits, which is a general signup credit rather than anything specific to GLM.
A token-denominated grant has more undocumented surface area than a dollar-denominated one, because dollars carry their own definition and tokens do not. Ask the expiry and model-scope questions in the same message, before you design around the answer.
Frequently Asked Questions
How many free tokens does the Z.ai Startups Program give?
Up to 1 billion free GLM API tokens for approved startups, plus volume pricing, priority rate limits and technical support. At GLM-5.3 list rates that is worth roughly $1,400 to $4,400 depending on the input and output mix. Current terms are tracked at getaiperks.com.
How much does the GLM-5.3 API cost?
$1.40 per million input tokens, $4.40 per million output tokens, and $0.26 per million cached input tokens, per Z.ai's official pricing documentation. GLM-5.3-Flash costs far less at $0.15 input and $0.50 output per million, roughly a tenth of the flagship rate.
Is GLM open source?
GLM-5.2 is fully open source under a plain MIT license, and GLM-5.1 shipped as an open-source flagship. Licensing is set per release, though, so do not assume MIT carries forward to a newer version. Open the license file in the weights repository for the exact model you plan to ship.
What is the cheapest GLM model?
Of the models with confirmed pricing, GLM-5.3-Flash is the cheapest at $0.15 per million input tokens and $0.50 per million output, with cached input at $0.03. That is roughly a tenth of the flagship rate, which makes it the sensible default for prototyping, evaluation runs and background work. Compare it against other vendors at getaiperks.com.
Do the Z.ai startup tokens expire?
Z.ai has not published an expiry window or a model restriction for the startup token grant. Both would substantially change what the grant is worth, so assume nothing until the vendor states it. AI Perks tracks program terms as vendors confirm them.
How does Z.ai compare to OpenAI and Anthropic credits?
Z.ai denominates its grant in tokens rather than dollars, so it does not line up cleanly against dollar-denominated programs on a spreadsheet. Converted at Z.ai's own list prices, a billion tokens is roughly $2,300 of equivalent value on a mixed workload. Whether it can be combined with any other vendor's program depends on that vendor's terms, which are worth reading individually rather than assuming.
A grant denominated in tokens is harder to compare than one denominated in dollars, which is exactly why fewer people bother to claim it.