DeepSeek Raised Prices and Revenue Doubled; Qwen Cut Voice Prices Up to 95% -- API Pricing Is No Longer One Race

On September 24, The Information reported that DeepSeek has reached an annualised revenue run rate of about $1 billion, up from under $500 million a few months earlier. The report cites people with direct knowledge; DeepSeek had not commented. The notable detail is how it got there. In August, DeepSeek raised model usage fees by 2.3 to 4.5 times depending on the model, and founder Liang Wenfeng reportedly told investors that demand held up. Essentially all of that revenue comes from API customers, while the consumer chatbot stays free. The same reporting says DeepSeek is finalising a second funding round of roughly $7.5 billion at a valuation of about 500 billion yuan.

A day earlier, on September 23, Alibaba's Qwen team released Qwen-Audio-3.1. It is a five-model stack: upgraded ASR, TTS and Realtime models, plus two new ones, TTS-Next for audio creation and ASR-Next for audio understanding. Alongside it came steep price cuts: about 95% for speech recognition, about 70% for TTS, and about 85% for the Realtime voice model. The capability claims, such as filler-word cleanup, speaker and emotion detection, and single-pass generation of voice, effects and background audio, come from Alibaba's own descriptions. No independent benchmarks accompanied the release.

These are not contradictory moves; they are pricing different products. DeepSeek's text and reasoning models had been priced far below what customers were evidently willing to pay, and a 2-4x increase that does not lose customers reads like the market discovering a price floor rather than a lab overreaching. Speech is a different contest. Recognition and synthesis have many credible providers, switching costs are low, and a platform like Alibaba Cloud can treat voice as a loss-leader that pulls developers into its broader stack.

For teams building on these APIs, the practical implication is that per-token cost is no longer a one-way ratchet downward. Budget models should carry a price-change scenario per provider and per modality, and multi-provider routing is worth the engineering if a single vendor's repricing can move a unit cost 4x in a month.

Still uncertain: run-rate figures project a strong month forward and are not audited revenue, and heavily discounted voice prices may not last once adoption targets are met.