DeepSeek V4 GA undercuts Fable 5 by 57x with Opus-level coding

DeepSeek's V4 GA release brings Opus 4.8-level coding performance at roughly 1.7% of the output cost of Anthropic's Fable 5, threatening to reset pricing across the $30 billion AI model market.

"Overall performance is close to Opus 4.8 level, with coding capability approaching GPT-5.6 Sol," developer Pankaj Kumar said after testing the model.

The release includes two tiers: V4 Pro at $0.435 per million input tokens and $0.87 per million output tokens, and V4 Flash at $0.14 and $0.28 respectively. Anthropic's Fable 5 charges $50 per million output tokens, while OpenAI's GPT-5.6 Sol is priced at $30. DeepSeek also introduced peak-valley billing for the first time, with peak rates doubling to $1.74 per million output tokens for Pro and $0.56 for Flash. Cache-hit pricing remains near-zero at $0.0036 per million tokens for Pro and $0.0028 for Flash.

The pricing gap pressures every major AI lab to justify premium pricing. Anthropic trades no public shares, but Nvidia, whose H100 GPUs powered much of the training infrastructure, could see demand shift if cheaper inference drives broader adoption. DeepSeek's MIT-licensed weights also allow self-hosted deployment, a feature none of its closed-source rivals offer.

Opus-Class Performance at a Fraction of the Cost

On SWE-bench Verified, DeepSeek V4-Pro scores 80.6%, within single digits of Claude Opus 4.8's roughly 87.6% and GPT-5.5's high-80s result. On LiveCodeBench, it posts 93.5%, and its Codeforces Elo of 3,206 ranks among the best in competitive programming. The model carries a 1-million-token context window matching Opus 4.8 and GPT-5.5.

Early testers reported strong results in 3D rendering and SVG generation, though some noted V4 required more iteration rounds than Fable 5 for the same task. The gap is narrowest on coding benchmarks and widest on long-context retrieval and specialized software engineering tasks, where Opus 4.8 still leads. Moonshot's Kimi K3, a 2.8-trillion-parameter open-source model released July 16, topped Fable 5 on front-end coding benchmarks but costs $15 per million output tokens — still 17 times more than V4 Pro.

Peak-Valley Billing Reshapes API Economics

DeepSeek's introduction of time-based pricing marks a departure from its flat-rate model. Peak hours cost double the base rate. Off-peak and cache-hit pricing remain the cheapest in the industry. For teams running batch inference, data labeling, or overnight test generation, the effective cost can fall below $0.01 per session.

The move comes as DeepSeek prepares to deprecate its older deepseek-chat and deepseek-reasoner models on July 24. The new pricing structure targets enterprise teams running continuous agent pipelines, where the difference between peak and off-peak routing could cut monthly API bills by 40 percent or more.

For investors, the calculus is straightforward. DeepSeek's pricing forces a choice on every AI lab: match the cost curve or lose the volume segment. OpenAI's GPT-5.6 Sol and Anthropic's Fable 5 target premium enterprise workloads where reliability and safety outweigh price. But for the 80 percent of coding tasks that don't require frontier-level reasoning, DeepSeek V4 Pro at $0.87 per million output tokens makes the premium hard to justify. Nvidia, whose GPUs underpin both training and inference across all four major labs, stands to benefit from any scenario that expands total AI compute demand, regardless of which model wins the pricing war.

This article is for informational purposes only and does not constitute investment advice.