Alibaba Cloud is dangling up to $5,000 in credits for AI workloads just as its Qwen3.7-Max model closes the gap with Claude and GPT on core benchmarks.
If you're building anything that touches AI infrastructure, Alibaba Cloud just made it cheaper to find out whether its models can do the job. The company is running a credit program that hands new customers $1,000 after their first $1,000 in spending, with two more $2,000 tranches available if usage climbs to $6,000 over the following months, adding up to as much as $5,000 in total credits. There's no minimum spend required to start, and the offer covers compute, storage, and the AI infrastructure driving an increasing share of Alibaba Cloud's revenue.
A separate Qwen-specific offer sweetens the deal further. Companies that build AI as their core product, run a live website, and were founded within the last decade can claim $5,000 in credits over a year, redeemable against roughly 2 billion tokens on Alibaba's Model Studio platform. For a startup burning through inference costs on GPT-5.5 or Claude Opus 4.6, that's a real incentive to benchmark Alibaba's models against the ones they're already paying for.
The timing isn't an accident. Alibaba just gave developers a genuine reason to look. At its Cloud Summit on May 20, the company unveiled Qwen3.7-Max, a reasoning and agent model with a full 1 million token context window that Alibaba's own engineers ran unsupervised for 35 hours straight. During that run it made 1,158 tool calls and delivered a 10x speedup on a kernel optimization task, working from its own judgment about what to try next.
What makes that launch notable isn't only the marathon run. According to VentureBeat's reporting on the release, Qwen3.7-Max supports external agent harnesses, including Anthropic's own Claude Code, so developers can plug Alibaba's model directly into tooling built around a rival lab's workflow. That's an unusual move for a frontier lab, and it signals Alibaba is competing on compatibility as much as raw capability.
The benchmarks back up the ambition, even if they don't settle the argument. On the Apex Math Reasoning benchmark, Qwen3.7-Max scored 44.5 against Claude Opus 4.6 Max's 34.5. On HMMT's February 2026 competition math set, it hit 97.1, edging out Opus 4.6 Max's 96.2. On SWE-Verified, the benchmark most engineers actually trust for real coding work, Qwen3.7-Max posted 80.4 to Opus 4.6 Max's 80.8, close enough to call a tie. GPT-5.5 still leads on most tasks, but Qwen3.7-Max beat it on SWE-Bench Pro, 60.6% to 58.6%.
None of that makes Qwen3.7-Max the best model available. It's text-only, so anyone needing vision input has to reach for the separate Qwen3.7-Plus-Preview instead. Frontier benchmarks are also notoriously easy to game and hard to translate into an everyday coding session. But a model that's statistically tied with Opus 4.6 Max on software engineering tasks, priced at $2.50 per million input tokens and $7.50 per million output tokens, changes the math for teams watching their API bills.
That pricing matters more than the credits do, honestly. A startup running high volume inference through Claude or GPT can burn through five figures a month without much to show for it. Alibaba's rate card undercuts both, and the API went live on Model Studio on May 19, a day ahead of the public unveiling, with same day listings on OpenRouter, Together AI, and Qubrid AI. Developers didn't have to wait for access. They could route a live workload to it within hours of the announcement.
There's a broader pattern here worth naming plainly. Chinese AI labs spent much of 2024 and 2025 catching up on raw capability. Now they're competing on distribution: free credits, compatible tooling, aggressive pricing, and same day availability across the third party platforms most developers already use. Alibaba isn't asking anyone to abandon Anthropic or OpenAI. It's asking them to run one workload on Qwen3.7-Max, look at the invoice, and decide for themselves.
Whether that's enough to shift real production traffic away from Claude and GPT is a separate question from whether it's smart marketing. It is. Free credits get a developer to try something once. A model that's genuinely close on the benchmarks that matter, at a fraction of the price, is what gets them to stay.
Also read: Alibaba Times Its Qwen Price Cuts to Catch American Coders at Work • Kling AI nears a $3 billion round as Tencent buys into a rival's video bet • Google's Electricity Demand Jumped 37 Percent in 2025 on AI Buildout
