How a 60% Discount and Full Rebates Turn Enterprise LLM Calls Into Profit

Wholesale Pricing for Enterprise LLM Calls

Recent releases such as DeepSeek V4, Qwen3.6 and GLM‑5.1 have pushed open‑source model capabilities forward, yet many teams that integrate these APIs into production complain that the bills are hard to understand. A conversation with an AI‑outbound startup revealed that five backend engineers each spend a day testing APIs, consuming nearly ¥20,000 per month, not because the models are weak but because price reductions lag behind the surge in usage.

iFlytek Starry MaaS addresses this mismatch with a tiered rebate mechanism. Both Qwen3.6‑35B‑A3B and Qwen3.5‑35B‑A3B are listed at a 60% discount. On top of that, the platform returns weekly vouchers equal to 30% of the actual spend when daily token consumption reaches 30 billion, and returns the full amount when consumption reaches 100 billion. The vouchers are model‑agnostic and can be used across more than 60 models in the marketplace.

Cost example : a team that consumes 50 billion tokens per day pays 60% of the list price for the first week, then receives a 30% voucher, effectively paying about 40% of the original price. A team that reaches 100 billion tokens gets a full rebate, making the marginal cost of inference close to zero because every yuan spent is returned as a voucher.

This is not a simple promotion; it is a "pay‑first, get‑back" wholesale model that turns the price‑cut from a one‑off discount into a continuous positive feedback loop—more usage yields larger rebates, which encourages even more usage.

Can the Discounted Model Handle Real‑World Workloads?

The author evaluates Qwen3.6‑35B‑A3B, which uses a Mixture‑of‑Experts (MoE) architecture: 35 billion total parameters, about 3 billion active per inference, yielding roughly one‑tenth the compute cost of a dense model of similar size. Benchmarks reported by the model include SWE‑bench Verified 73.4, Terminal‑Bench 2.0 51.5, AIME 2026 math 92.7, and OmniDocBench 89.9, confirming strong reasoning and document‑understanding abilities.

To test production readiness, the author asked the model to generate a fully commented quick‑sort implementation in Python, which it returned in a few seconds with correct logic. Next, the model was tasked with converting a 3,000‑line synchronous Python micro‑service into an asynchronous version. Without any reference code, the model suggested a library‑swap table (Flask→FastAPI, requests→httpx, pymysql→aiomysql) with rationales, identified the common pitfall of a synchronous lock blocking the event loop, and supplied a concrete code snippet for asyncio.Lock. It then proposed a four‑stage migration plan—IO replacement, async business logic, async external calls, entry‑point refactor—each with a verification method.

One shortcoming noted was the lack of discussion about cache‑stampede risk in the async scenario, which the author had to infer independently.

Overall, the model’s engineering judgment on this task surpassed many flagship models, and the token cost for a typical request (≈8,000 tokens) drops from ¥3‑5 at full price to under ¥3 after the 60% discount, with rebates potentially covering the remainder.

Integration Workflow

Connecting to the platform requires three steps: (1) log in to the Starry MaaS portal and obtain API credentials; (2) select the desired model (e.g., Qwen3.6‑35B‑A3B) and create an application; (3) configure the client (the author demonstrates with Cherry Studio) by entering the endpoint, API key and model ID. The OpenAI‑compatible interface allows a simple configuration change to work with tools such as Cursor, Claude Code or OpenClaw.

The UI screenshots (see images) show the model marketplace, the model‑service addition dialog, the configuration page, and the connection test result. After a successful test, the model appears in the chat list and can be used immediately.

Thresholds and Market Impact

The 30 billion‑token daily threshold is reachable for large‑scale use cases: high‑volume intelligent customer service (hundreds of tokens per dialogue), enterprise knowledge‑base RAG pipelines (tens of thousands of documents), and code‑assistant workloads. Once usage exceeds 100 billion tokens per day, the rebate makes inference effectively free, turning the pricing model into a “use‑more‑save‑more” flywheel.

Smaller teams that cannot meet the threshold will not benefit from the rebate, highlighting that the program targets high‑value B‑side customers. Nevertheless, the platform’s unified OpenAI‑compatible API across dozens of models simplifies switching—changing a single modelId is enough—providing genuine flexibility rather than a mere SKU list.

From an industry perspective, the shift from flat per‑token pricing to tiered rebates redistributes pricing power toward large enterprises. If other providers adopt similar mechanisms, the overall cost of large‑model inference could continue to decline, accelerating adoption across sectors.

In conclusion, the combination of a 60% base discount, volume‑based weekly rebates, strong model performance, and a straightforward integration path makes iFlytek Starry MaaS a compelling choice for enterprises seeking cost‑effective large‑model services.