Headline: DeepSeek quietly ships V4 Pro-0813 — marginally behind top U.S. models but thousands of percent cheaper, a big deal for crypto builders DeepSeek rolled out a finished version of its flagship LLM this week with almost no fanfare — no blog post, just an updated model name on its API pricing page: DeepSeek-V4-Pro-0813. The new tag signals updated weights under the hood. Pricing is explicit: roughly $0.435 per million input tokens and $0.87 per million output tokens (with cached input at $0.003625). Those rates haven’t changed; what did is the model itself. Why this matters - Performance: DeepSeek published a sheet comparing V4-Pro-0813 across 10 agent-style benchmarks against Anthropic’s Fable 5 and other models. On the eight benchmarks where Fable 5 led, the gaps were small. Averaged across all tests, Fable 5’s relative lead is about 5.3%. Excluding one outlier test (Humanity’s Last Exam without tools, where DeepSeek scores 42.7 vs Fable’s 53.3), the remaining comparisons average a 2.8% gap. In short: performance is in the same neighborhood. - Verification caveat: the V4 family has been publicly available in preview builds since April, and DeepSeek has previously called those previews unfinished. The V4-Pro API was listed as “unchanged” when V4-Flash hit GA on July 31. The Hugging Face model card still labels V4 as a preview, and nobody outside DeepSeek has yet benchmarked the new 0813 build independently. Two of the 10 DeepSeek benchmark sets are internal (DSBench-FullStack and DSBench-Hard), so some results aren’t publicly verifiable. The economics: where the story really changes - Price vs. price: Anthropic’s Fable 5 charges about $10 per million input tokens and $50 per million output tokens. DeepSeek’s V4 Pro sits at $0.435/$0.87 per million — orders of magnitude cheaper. - Blended comparison: using blended-rate math reported, that’s roughly $30 per million effective for Fable 5 versus $0.65 for DeepSeek — about 46× cheaper (4,600% of the cost). - Per-task cost: independent analyses highlight even wider gaps when measured per completed benchmark task because the pricier models tend to “think longer” and produce longer outputs. Artificial Analysis estimated $3.15 per benchmark task for Fable 5 vs $0.03 for DeepSeek V4-Flash (≈105× cheaper). Hugging Face CEO Clément Delangue put a per-task spread at over $31 vs ~$0.04. There’s not yet a public per-task figure for 0813 specifically. Competitive context - Anthropic itself complicates the premium story: Claude Opus 5 outperforms Fable 5 on many benchmarks while running at roughly half the price. - Chinese open-weights groups keep closing the gap on U.S. frontier models while dramatically undercutting prices. Examples: Kimi K3 beat Fable 5 and GPT-5.6 Sol on release, and DeepSeek and Xiaomi have been pushing frontier-cost reductions as large as 99% compared with U.S. offerings. - Openness: DeepSeek’s weights are MIT-licensed and available on Hugging Face, meaning independent teams can download and evaluate them directly once they run benchmarks on the released 0813 weights. What this means for crypto projects - For decentralized apps, DeFi analytics, on/off-chain oracles, smart-contract auditing, and bot-driven strategies, inference cost is a direct operational expense. A 40–4,600× difference in cost can change which features are economically viable at scale. - Cheaper, nearly frontier-capable models unlock higher-frequency, lower-margin use cases: automated monitoring, live trading signals, transactional metadata generation, and large-scale data labeling for on-chain analytics. - But caveats remain: independent benchmarking of the exact 0813 build is still pending, and some of DeepSeek’s tests are private. Teams should verify performance against their own tasks before migrating mission-critical workloads. Bottom line: DeepSeek’s V4-Pro-0813 tightens the tradeoff between raw accuracy and price. If independent benchmarks confirm DeepSeek’s claims on the released weights, the model could reshape cost-sensitive AI usage in crypto ecosystems — enabling many new LLM-driven services that were previously too expensive to run at scale. Read more AI-generated news on: undefined/news