When accounting for caching behavior in typical agentic workflows (where input tokens heavily outnumber output tokens) SignalBloom's blended cost estimates put Anthropic and OpenAI at roughly $2.80 to $2.82 per million effective tokens. DeepSeek, by contrast, comes in at approximately $0.094 per million tokens under the same conditions. That is a roughly 30x price difference.
The question the analysis poses is whether frontier models are 30 times more capable than DeepSeek for the tasks enterprises actually need done. The author's conclusion: not for most coding and agentic workflows. When paired with a competent human engineer who can handle tasks where current AI still falls short — such as long-term memory management, assessing whether there is sufficient evidence to act, and genuinely autonomous decision-making — a cheaper model is often good enough.
The piece also highlights a broader trend the author calls "tokenmaxxing," in which frontier model architectures and product defaults push toward higher token consumption per task, compounding the effect of rising per-token prices. Token usage has risen substantially across the industry even as the narrative around inference costs has emphasized falling prices.
SignalBloom acknowledges the limitations of the analysis — future pricing, open-source model improvement curves, and enterprise behavior are all difficult to forecast, and market participants adjust their own behavior as conditions change. But the core argument holds regardless of precise projections: as AI API spend becomes a meaningful and growing line item in enterprise budgets, the existence of a credible 30x cheaper alternative puts a structural limit on how far frontier labs can push pricing before customers start moving.
This analysis is based on reporting from SignalBloom AI.
Image courtesy of Austin Distel.
This article was generated with AI assistance and reviewed for accuracy and quality.
