What the Alibaba AI Catalyst Program Is Reported to Give You
Alibaba Cloud AI Catalyst is reported to offer startups up to 2 billion free Model Studio tokens. A separately reported program, Startup Catalyst, is described as offering up to $120,000 in cloud credits across four tiers. Neither figure could be confirmed against an Alibaba program page during this research, so both belong in the reported column rather than the planning column.
That distinction is the first thing secondhand write-ups lose. In the coverage these are two different programs rather than two lines of one bundle, and the four-tier structure attaches to the cloud credits, not the token grant. A budget assuming one application delivers both rests on something nobody has published.
"Up to" is doing real work in both figures, and no source we could check explains what separates a full award from a fraction of one. AI Perks tracks both programs and updates them as terms get confirmed. Numbers at this confidence level are the upper bound for a conversation, not a line in a forecast.
What Qwen3.8-Max and Qwen3.8-Flash Are Reported to Cost
Qwen3.8-Max is reported at $2.00 per million input tokens and $6.00 per million output on Model Studio. That figure circulates as official pricing, but it did not survive a direct check against an Alibaba pricing page, so treat it as reported.
The same is true of every other Qwen rate below, which is why each is labeled.
| Model | Released | Input / 1M | Output / 1M | Weights |
|---|---|---|---|---|
| Qwen3.8-Max | Aug 3, 2026 (GA) | $2.00 (reported) | $6.00 (reported) | Closed |
| Qwen3.8-Flash | Aug 3, 2026 | $0.15 (reported) | $0.47 (reported) | Open |
| Qwen3.7-Max | May 20, 2026 | $2.50 (reported) | $7.50 (reported) | Closed |
| Qwen3.8-27B | Aug 17, 2026 | No vendor API price | No vendor API price | Open |
| GLM-5.3 (Zhipu, for reference) | Aug 2026 | $1.40 (official) | $4.40 (official) | n/a |
The GLM row is a sanity check, not a recommendation. Zhipu publishes its rates in its own developer docs, which makes GLM-5.3 the only price in this table resting on a primary source: $1.40 input, $4.40 output, $0.26 cached, with a GLM-5.3-Flash sibling at $0.15, $0.50 and $0.03. Against that anchor, reported Qwen3.8-Max is the pricier flagship and reported Qwen3.8-Flash sits level with GLM-5.3-Flash. Cached input for Qwen3.8-Max is reported at $0.25, unconfirmed like the rest, so do not build a caching case on it yet.
Qwen3.8-Max is reported as a 2.4 trillion parameter closed-weight flagship positioned against Claude Opus and DeepSeek frontier models, live across the Beijing, Singapore, Tokyo, Frankfurt and Virginia regions, all at the same confidence as the prices.
Qwen3.7-Max is the older flagship, unveiled at the Alibaba Cloud Summit on May 20, 2026 with API access live the day before, and it is the model that took the context window to 1 million tokens, up from 256K. It was also reported to sustain roughly 35 hour autonomous coding runs and to place fifth on the Artificial Analysis Intelligence Index at launch. If the reported rates hold, Qwen3.8-Max is both newer and cheaper, which makes 3.7-Max hard to justify for new work. Model cost comparisons live at AI Perks.
What 2 Billion Free Tokens Would Be Worth
If the 2 billion token figure holds, it converts to somewhere between roughly $300 and $12,000 of usage at reported list prices, depending entirely on which model you point it at and your input to output ratio.
This is the calculation nobody publishes, and it decides whether the program is worth your time. Grant size and rates are all reported, so read the table as a shape rather than a quote.
| How you spend the 2B tokens | On Qwen3.8-Max | On Qwen3.8-Flash |
|---|---|---|
| All input | $4,000 | $300 |
| All output | $12,000 | $940 |
| Typical 80% input, 20% output | $5,600 | $428 |
Three things fall out of that shape. First, a "2 billion token" headline is worth an order of magnitude more on the flagship than on the cheap model, which inverts the usual advice to run everything small. If a grant is token-denominated rather than dollar-denominated, the logic is to spend it on the expensive model and pay cash for the cheap one. Second, the reported $120,000 cloud-credit figure dwarfs the token figure, but it belongs to a different reported program, so the two should not be added together in a plan. Third, $5,600 of frontier inference is a real but finite runway, a serious prototype rather than a year of production. Stacking credits from several vendors is how teams actually cover a first year, and AI Perks shows which grants are compatible.
The Free Quota That Does Not Involve a Program
New Alibaba Cloud Model Studio accounts are reported to receive a free token allowance lasting 90 days, described as roughly 70 million tokens in total and 1 million tokens per model. The specifics are unverified.
The structure matters more than the headline, and none of it survived a check against Alibaba documentation:
The allowance is described as per model rather than pooled. Coverage puts it at 1,000,000 tokens per model, which is where the roughly 70 million aggregate appears to come from: dozens of models rather than one shared pot. If that is right, evaluating six models gives six separate million token budgets.
Ninety days is the reported window. What starts that clock, account activation or the launch date of the model itself, is not something any source we could check states.
The previous general developer free trial was reportedly discontinued on April 15, 2026, and this allowance is described as its replacement. Both the date and the structure are reported, not confirmed.
Because all of it is secondhand, treat it as a figure to confirm in your own account before planning an evaluation around it. Free tiers across model vendors are compared at AI Perks.
Qwen3.8-27B and the Zero Dollar Option
Qwen3.8-27B, released August 17, 2026, is an open-weight model built to run on laptops and consumer hardware, which means the vendor API price is not a number you need at all.
Alibaba's claim is that it matches models roughly ten times its size on coding, research and long-horizon agentic work. That is a vendor statement, not an independently benchmarked one.
The strategic point survives the uncertainty. If a 27B model handles classification, extraction and routing locally, paid tokens only get spent on requests that genuinely need a frontier model, and a grant covering only the hard requests lasts far longer than one absorbing everything. Qwen3.8-Flash is also distributed as open weights, reported at 125 billion parameters, and Alibaba compares it favourably to Opus 4.6 and DeepSeek V4-Flash in its own materials, which is again a vendor comparison. No hosted endpoint price for 27B turned up during this research, so self-hosting is the route.
What Is Not Confirmed
Four things a team planning around Qwen credits will want, and none of them survived a check against an Alibaba source as of September 2026.
The gaps are more useful stated than guessed at:
The token grant itself. Up to 2 billion Model Studio tokens is widely repeated and has no primary confirmation, and nothing published explains what determines an individual award.
The cloud-credit program and its tiers. Up to $120,000 across four tiers is reported for Startup Catalyst. Amount and tier structure are both unconfirmed.
Expiry on granted tokens. The 90 day window belongs to the reported new-account allowance. Whether granted tokens carry the same clock, a longer one, or none is not stated anywhere we could check.
Primary-source prices for any Qwen model. None of the three rate pairs above survived a direct check against Alibaba's own pricing documentation.
The honest summary: the Qwen headline numbers are generous and secondhand, and the only primary-source price on this page belongs to a competitor. Terms as they get confirmed are tracked at AI Perks.
Frequently Asked Questions
How many free tokens does Alibaba AI Catalyst give?
Coverage puts it at up to 2 billion free Model Studio tokens for startups, with a separate Startup Catalyst program reported at up to $120,000 in cloud credits across four tiers. Neither figure was confirmed against an Alibaba program page, and "up to" is doing real work in both. Terms as they get confirmed are tracked at AI Perks.
What does the Qwen API cost per million tokens?
Qwen3.8-Max is reported at $2.00 per million input tokens and $6.00 per million output, with cached input reported at $0.25. Qwen3.8-Flash is reported at $0.15 and $0.47, and Qwen3.7-Max at $2.50 and $7.50. None were verified against an Alibaba pricing page.
Can I use Qwen for free without applying for credits?
Reportedly yes. New Model Studio accounts are described as receiving a 90 day free token allowance, put at roughly 70 million tokens in total and 1 million per model. Those specifics are unverified, so confirm them in your own account first.
Is Qwen3.8-Max cheaper than other frontier models?
Against the one primary-source price here, Zhipu's GLM-5.3 at $1.40 input and $4.40 output, the reported Qwen3.8-Max rate of $2.00 and $6.00 is the more expensive of the two. Comparisons against Western frontier models get repeated constantly, but this page has no verified price for them and will not make that claim. Real cost is workload dependent anyway, because output ratio and caching move the bill more than headline rates do.
Are Qwen models open weight?
Some are. Qwen3.8-27B and Qwen3.8-Flash are distributed as downloadable open weights, so they cost nothing from the vendor and you pay only for the hardware. Qwen3.8-Max and Qwen3.7-Max are closed-weight, which leaves the API as the only route to them.
Does the reported cloud-credit program cover Qwen tokens?
Unknown, and it is the most expensive unknown on this page. The two are described as separate programs, and no source we could check states whether Model Studio usage can be paid from a cloud-credit balance or whether the two are ring-fenced. That one answer changes the value of the package more than any other detail here.
The headline numbers are reported. The only confirmed price here belongs to a competitor.