Chinese AI Models Are Now Over 100x Cheaper Than Anthropic

The 105x gap is real, but narrower than the headline sounds

The AI model price war has moved into a new phase. Reuters reported this week that DeepSeek V4-Flash costs about $0.03 per Artificial Analysis benchmark test, compared with $3.15 for Anthropic's Claude Fable 5. On that workload, the Chinese model is roughly 105x cheaper.

The underlying token prices make the gap just as striking. DeepSeek lists V4-Flash at $0.14 per million uncached input tokens and $0.28 per million output tokens. Anthropic lists Claude Fable 5 at $10 per million input tokens and $50 per million output tokens. That makes Fable 5 about 71x more expensive on standard input and about 179x more expensive on output.

The distinction matters. Saying that every Chinese AI model is more than 100x cheaper than every Anthropic model would be wrong. The 105x figure is a specific cost-per-benchmark comparison, and different workloads produce different ratios. But the broader signal is still hard to ignore: the low end of China's model market is now operating on a radically different cost curve from premium US frontier models.

Qwen shows this is bigger than one DeepSeek release

DeepSeek is not the only Chinese provider pushing inference toward commodity pricing. Alibaba Cloud lists Qwen3.5 Flash in its global Virginia deployment at $0.029 per million input tokens for requests up to 128,000 tokens and $0.287 per million output tokens.

Against Claude Fable 5's current API rates, that is roughly 345x cheaper on input and 174x cheaper on output. Qwen3.6 Flash is more expensive than Qwen3.5 Flash, but it still comes in at $0.165 per million input tokens and $0.99 per million output tokens for shorter contexts, far below Fable 5's rates.

These are not capability-equivalent products, and the cheapest model is not automatically the best choice. Still, the fact that both DeepSeek and Alibaba can sell useful API access at these prices suggests the pricing pressure is structural, not a one-company stunt. Chinese labs are competing not only on benchmark scores and model size, but on how cheaply developers can run intelligence at scale.

China is competing on the cost curve, not just the leaderboard

The pricing gap is becoming part of the product strategy. DeepSeek says V4 uses token-wise compression and sparse attention to reduce compute and memory costs, while making a one-million-token context window standard across its official services. V4-Flash is explicitly positioned as the smaller, faster and more cost-efficient member of the family.

That technical story sits inside a much more aggressive market. Reuters notes that DeepSeek is fighting for attention with Moonshot AI, MiniMax, Z.AI, ByteDance and Alibaba. Alibaba's Qwen lineup spans everything from low-cost Flash models to much larger Max systems, giving developers multiple price and capability tiers inside the same ecosystem.

For US builders, the strategic implication is more important than the exact architecture. If capable models can be offered at cents instead of dollars for the same class of automated task, then model calls stop looking like a scarce premium resource. They start looking more like cloud primitives that can be invoked, retried and routed in large numbers. That changes the economics of coding agents, customer support automation, document processing and other high-volume workflows.

The catch: price per token is not the same as cost per result

The temptation is to turn a pricing table into a quality ranking. That would be a mistake. Claude Fable 5 is Anthropic's premium model for long-running agents, coding, vision and complex enterprise workflows. DeepSeek V4-Flash and Qwen Flash models target a different part of the market, where throughput and cost efficiency matter more.

There is also a measurement problem. A March 2026 study of reasoning models found that the model with the lower listed token price ended up costing more in 21.8% of model-pair comparisons. The main reason was variation in how many thinking tokens different models consumed. In some cases, the same query triggered dramatically different amounts of hidden reasoning work.

That means buyers should measure cost per successful task, not just cost per million tokens. Reliability, retry rates, latency, context handling, tool use and output length can erase part of a headline price advantage. The 105x DeepSeek-versus-Fable benchmark gap is useful precisely because it tries to move beyond sticker price and measure the cost of completing a standardized evaluation workload.

For US builders, model routing is becoming the default answer

The most practical response is not to replace every premium model with the cheapest Chinese API. It is to stop treating one model as the default for every step of a workflow.

A company processing 100 million output tokens would pay about $28 at DeepSeek V4-Flash's listed output rate, versus $5,000 at Claude Fable 5's rate. Qwen3.5 Flash would be about $28.70 for the same token volume at its lowest listed output tier. At that scale, even a modest quality gap can be worth managing with routing, retries and selective escalation.

A sensible architecture is increasingly obvious: use low-cost models for extraction, classification, summarization, code transformations and parallel agent subtasks, then route hard planning, sensitive decisions, multimodal work or low-confidence outputs to a premium frontier model. The result is not a winner-take-all model market. It is a layered stack where the expensive model is called only when its extra capability is worth the premium.

The bigger risk for Anthropic is therefore not that DeepSeek or Qwen replaces Claude outright. It is that developers learn they no longer need Claude for most tokens. If the majority of routine inference moves to models priced one or two orders of magnitude lower, premium US labs may have to justify their margins task by task rather than model by model.

Sources