DeepSeek V4: The Open-Source Model Frontier Labs Feared | Hello, AI<br>← Back to Hello, AIDiscovery3 minMay 15, 2026<br>DeepSeek V4: The Open-Source Model Frontier Labs Feared<br>DeepSeek V4 ships under MIT with $0.30/M output tokens — 83x cheaper than Claude Opus 4.7 — while scoring 80.6% on SWE-bench Verified. The agentic-coding price floor just moved an order of magnitude.<br>DeepSeek V4-Pro charges $0.30 per million output tokens. Claude Opus 4.7 charges $25. GPT-5.5 charges $30. That is an 83-to-100x gap, and the model on the cheap end of it just scored 80.6% on SWE-bench Verified — 0.2 points behind Claude Opus 4.6. The weights shipped April 24 on Hugging Face under an MIT license with no commercial restrictions.<br>The architecture is what makes the pricing defensible rather than promotional. V4-Pro is a 1.6-trillion-parameter MoE that activates only 49 billion parameters per token, and DeepSeek cut single-token inference FLOPs to 27% of V3.2 while shrinking KV cache occupancy at 1M-token context to 10% of the previous generation. The cost structure isn't a loss-leader subsidy — it reflects an inference profile that a competent infra team can roughly replicate on their own hardware. Self-hosting a 1.6T MoE is still not trivial, but the per-token economics finally make sense for teams that already run GPU fleets.<br>The coding numbers are the part frontier labs will find hardest to spin. V4-Pro posts the highest LiveCodeBench Pass@1 of any model at 93.5, a Codeforces rating of 3206 that edges out GPT-5.4 xHigh's 3168 and Gemini 3.1 Pro's 3052, and the SWE-bench Verified result lands inside the closed-frontier band. Agentic coding — the workload that has justified $25-per-million pricing for the last two years — is no longer a closed-model moat. It is a tier where an open-weight Chinese model is competitive on quality and roughly two orders of magnitude cheaper on API.<br>The caveats are real and worth stating plainly. DeepSeek's benchmark transparency is thinner than Anthropic's or Google's — the reports are credible but less audited, and independent replications are still arriving. The lab is Chinese, which carries data-governance implications some buyers cannot ignore regardless of license terms. And 1.6T parameters means self-hosting requires multi-node inference; the $0.30 API price assumes you accept DeepSeek's hosted endpoint, with whatever logging and jurisdictional exposure that entails. Teams with sensitive code will weigh those costs against the savings differently than a startup burning runway on Claude bills.<br>What this resets is the price floor for frontier-grade coding intelligence. When the cheapest credible option for 80%+ SWE-bench was $15 per million output tokens, closed labs could hold the line. With $0.30 weights sitting on Hugging Face under MIT, every procurement conversation between now and Q4 starts from a different anchor. Expect Anthropic and OpenAI to either compress output pricing on their next tier or sharpen the agentic and tool-use capabilities that benchmarks still don't capture — because parity on SWE-bench at one-hundredth the cost is the kind of pressure that ends a pricing regime.

← More from Hello, AI