DeepSeek's V4-Flash runs 100x cheaper than Claude Fable 5

DeepSeek has reopened the price war in AI, and this time the gap is not close. Artificial Analysis, a San Francisco research firm, found that V4-Flash, released on 31 July, costs an average of 3 cents per benchmark test to run — against $1.86 for OpenAI's GPT-5.6 Sol and $3.15 for Anthropic's Claude Fable 5. That makes it more than 100 times cheaper than Fable 5 to operate, according to Reuters.

The list price tells only part of the story: $0.14 per million input tokens and $0.28 per million output tokens. Artificial Analysis instead measures cost per completed test, because a model that looks cheap per token can still cost more overall if it needs many more steps to reach a correct answer. By that measure, V4-Flash is not just cheap — it is in a different class from every other well-known model, including Moonshot's Kimi K3 at 86 cents.

The trade-off shows up in capability. V4-Flash scores 50 out of 100 on the Artificial Analysis Intelligence Index, a composite of nine benchmarks spanning coding, reasoning and workplace tasks. That puts it level with Google's Gemini 3.6 Flash and a point behind Meta's Muse Spark 1.1 and Z.ai's GLM-5.2. Kimi K3 scores 57, and Claude Opus 5, Fable 5 and GPT-5.6 all lead V4-Flash by at least nine points. Architecturally, V4-Flash is a mixture-of-experts model with about 284 billion total parameters but only around 13 billion active per request, hybrid attention, and a 1 million token context window — design choices aimed squarely at keeping inference costs down.

DeepSeek is reportedly preparing for a possible IPO. Its R1 model triggered a global tech sell-off in early 2025 when it first demonstrated that Chinese labs could match Western performance at a fraction of the cost, according to The Next Web.

For founders running high-volume, low-stakes workloads — customer support, data extraction, bulk summarisation — V4-Flash's price makes it hard to justify paying 60 to 100 times more for a frontier model, especially where the extra nine-plus points of intelligence don't change the outcome. That puts direct pricing pressure on OpenAI and Anthropic's mid-tier offerings, which now have to justify their premium on quality alone.

The risk sits with model providers that compete on raw capability without a credible cheap tier. DeepSeek is proving that "good enough" at near-zero cost is a viable product category, not just a stunt — and with an IPO in view, it has every incentive to keep pushing the price floor down further.