DeepSeek Search debuts at $0.75/1k searches, cheapest paid Search API, ranks 17th of 27
ArtificialAnlys · x · 2026-10-22
Artificial Analysis benchmarked DeepSeek's Search API (launched Oct 17) in its Stirrup agent harness with GPT-5.6 Luna (medium) as the calling model across BrowseComp, DeepSearchQA and AA-Omniscience.
Key results
- Search Index score: 69, ranking 17th of 27 — a 36-point lift over the model-only baseline (33), between TinyFish (71) and Linkup standard (68)
- Strongest on factual QA (AA-Omniscience 70%, 10th, 3 points behind leaders Exa/Firecrawl); weakest on multi-hop BrowseComp (63%, 25th, far behind Perplexity/Octen at 85-87%)
- Pricing: $0.75 per 1k searches, the cheapest paid Search API on the board; $0.032 per task, 5th cheapest of 27 (median $0.066)
- Each search returns 7k tokens, driving model inference cost to $0.023/task (2nd highest on the board); search itself is only 27% of total cost
- Constraints: 128-character query limit (6% of searches hit it, 87% fixed by a shorter rewrite), 5 searches/sec rate limit, empty results not billed
Overall DeepSeek Search is an ultra-cheap option for factual retrieval but sits below the cost frontier — Octen scores 77 at $0.024 per task.
More from Models
- Users complain Anthropic models have 'suddenly got really stupid' — gaganghotra_ · 2026-10-22
- The '4' in Gemini 4 was a randomly chosen number, insider says — prajdabre · 2026-10-22
- Artificial Analysis benchmarks 27 Search APIs across 14 providers; DeepSeek Search 3 points off the lead — ArtificialAnlys · 2026-10-22
- Lafa scales flow language models to 8B, matching AR quality at 2.5x throughput — AdtRaghunathan · 2026-10-22
- Jev 1.14 lands on Databricks, targeting fast decisions and complex tasks — matei_zaharia · 2026-10-22
- Reddit asks: which 50-150B MoE models are worth running locally besides Qwen Flash Next — FlyingFishMakeAWish · 2026-10-22