Photo: ianwelsh.net
· rights & takedowns According to Ian Welsh's blog post on IanWelsh.net, the author compares token costs and performance between US-hosted models and the Chinese service DeepSeek V4. The post reports pricing figures: Claude at $5 input and $25 output per million tokens, DeepSeek at $0.28 input and under $1 output per million tokens (with discounts), and a reported cached cost of $0.0002 per million tokens. Ian Welsh reports personal spending of $3,000$5,000 per month on Claude Code before switching to DeepSeek and says his work now costs about $5/week versus an asserted $1,000/week for the same workload on Claude. The post argues that Chinese AI offerings, combined with open-source models and efficiency, create a competitive advantage over Western models. What happened According to Ian Welsh's blog post on IanWelsh.net, the author describes switching from Claude to DeepSeek V4 after observing large cost differences. The post gives pricing comparisons: Claude at $5 input and $25 output per million tokens, DeepSeek at $0.28 input and under $1 output per million tokens (with current discounts), and a reported cached cost of $0.0002 per million tokens. The author reports previously spending $3,000$5,000 per month on Claude Code and now paying about $5/week on DeepSeek, contrasted with a claimed $1,000/week for the same workload on Claude. Editorial analysis Industry-pattern observations: cost-per-token and caching efficiency materially change economics for high-volume, token-intense workloads. Organizations evaluating cloud vs self-hosting often find that aggressive caching and lower inference costs reduce marginal price pressure and can enable different trade-offs between latency, control, and total cost of ownership. Industry context Industry-pattern observations: public commentary emphasizing low-cost, open-source models reflects a broader debate on whether proprietary frontier models retain value when near-parity performance is available at far lower running cost. Analysts and practitioners have repeatedly noted that cheaper inference shifts the levers for product design, deployment scale, and operational budgets. What to watch For practitioners: monitor independent benchmarks for quality parity, reproducible latency and cost measurements for comparable workloads, and availability of permissive licenses or self-hosting tooling that materially reduce vendor lock-in. For investors and builders: track whether cost reductions are sustained at scale and whether open-source projects maintain active maintenance and security practices. Scoring Rationale The post highlights a material cost argument that is relevant to practitioners evaluating self-hosting and vendor lock-in. The piece is an opinionated synthesis with no independent benchmarking beyond the author's claims, so relevance is moderate rather than industry-shaking. Practice interview problems based on real data 1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with. Try 250 free problems