A production agent running 10,000 search-and-fetch cycles per day would spend $7.50 on DeepSeek’s new APIs. On any established search provider, that same workload costs $50 to $120. The gap is not marginal — it is a different billing model entirely.
DeepSeek’s Search and Fetch APIs, launched October 17 alongside DeepSeek-V4.1-Flash-1017, charge $0.75 per 1,000 successful requests per endpoint. Billing begins at 00:00 Beijing time on November 1 (16:00 UTC on October 31). Until then, calls through DeepSeek Harness, Claude Code, and Codex are free. Each endpoint is capped at 5 requests per second per account, and failed requests, empty results, and timeouts are not billed.
The pricing is per-call, not per-token. That distinction matters because it is the first time any company that ships a frontier model has offered agent tooling at a flat rate per request rather than metering by the volume of text that flows through the system. Every frontier lab — OpenAI, Anthropic, Google — still prices model inference by the token. OpenAI’s GPT-5.6 Sol runs $5.00 per million input tokens and $30.00 per million output. Anthropic’s Sonnet 5.5 is $2.00 and $10.00. Google’s Gemini 2.5 Pro is $1.25 and $10.00. The cost of a single agent step depends on how much text the model consumes and produces, which makes budgeting for high-frequency agentic loops inherently volatile.
Against established retrieval providers, the gap is even wider. Brave Search API charges $5 per 1,000 requests. Exa charges $7 for standard search and $12 for deep search. Tavily ranges from $8 to $16 depending on depth. Perplexity Sonar starts at $5 per 1,000 requests and adds token costs on top. DeepSeek’s $0.75 is 6.7 times cheaper than Brave and 9.3 times cheaper than Exa’s standard tier. This is the same price architecture our prior coverage identified in DeepSeek’s retrieval monetization move — the per-call model is the mechanism that makes the price floor sustainable.
Agents use models differently than humans do. A human chatting with a language model generates a long, variable sequence of tokens per exchange. An agent performing retrieval makes a structured call, receives a bounded result, and moves to the next step. The value of that step is binary: the agent got the information or it did not. Per-call pricing maps to that binary outcome. Per-token pricing does not — it forces the builder to pay for output length even when the retrieval result is a short, fixed-format object. The agent cost stack has been restructuring around exactly this divergence: Google now meters compute directly, factoring prompt complexity and effort, while Orb’s 2026 pricing study of 80 AI agent companies finds 95 percent on hybrid pricing models, with effort, compute, and discrete steps climbing as the primary billing units. Pure outcome pricing remains rare at 3.8 percent and falling.
DeepSeek’s move isolates the retrieval layer as the first component to transition from per-token to per-call. The sandbox and infrastructure layers are already commoditizing — what remains contested is the orchestration and retrieval interface above them. By pricing retrieval at a flat rate per call, DeepSeek is attempting to capture that interface before established providers can respond with their own per-call models.
Whether this pricing model propagates to core inference is the question frontier labs are now watching. The per-token model captures value across the entire lifecycle of a prompt — every input token, every output token, every reasoning step. A flat per-call rate for inference would require labs to estimate the average workload per call and price accordingly, which introduces margin risk on complex reasoning tasks. The structural incentive for OpenAI, Anthropic, and Google is to keep inference per-token while watching whether DeepSeek’s retrieval pricing forces their own search and tooling layers to compress.
For builders, the practical shift is in cost modeling. Per-call pricing makes the retrieval layer predictable: 10,000 calls at $0.75 per 1,000 is $7.50, regardless of how much text each call returns. That predictability does not exist in per-token inference, where a single complex reasoning chain can consume thousands of output tokens at variable cost. The ability to mix per-call retrieval with per-token reasoning — using DeepSeek’s tooling alongside a frontier model for the cognitive layer — gives builders a new axis of cost control. The catch is the 5-requests-per-second cap per endpoint, which limits throughput for high-concurrency agents. Whether DeepSeek raises that ceiling will determine whether the per-call model scales to production workloads or stays confined to prototyping.
Verification Note: All pricing figures cited in this article are from vendor-published documentation (DeepSeek API docs, Brave, Exa, Tavily, Perplexity, OpenAI, Anthropic, Google). OpenAI and Anthropic pricing sourced from third-party aggregators due to access restrictions on official pricing pages.