AI Aulis

AI NEWS IN ENGLISH · 21 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

← Issue 21 November 2026Paper

Artificial Analysis: Parallel Search shares top place among search APIs with 83 points

AI-generated contentWritten by AI

artificialanalysis.ai, x.com · 4 sources · Sources published: 20 November 2026 · Some source publication dates are unknown

In Artificial Analysis’s search benchmark, Parallel Search’s advanced mode scored 83, the same as Keenable Search (pro). The quicker fast mode scored 82 at a total cost of about $0.014 per task.

Artificial Analysis (AA) published the results on X on 20 November. Parallel’s score rose markedly from August, when advanced mode scored 75 and fast mode 73. In the November test, the model filled in the optional objective field and one to three search queries for every search call. In the chart in AA’s post, the August results have no objective label. AA does not say what caused the rise or whether the objective field was responsible.

An AI agent uses a search API to retrieve information from the web. AA compared 26 search products from 14 providers and kept the answer model, GPT-5.6 Luna (medium), and the settings the same. The search APIs ran in Stirrup, AA’s open-source harness: a program that controls the agent, gives the model its tools and runs its steps. Only the search service changed. OpenAI Web Search is a reference point in which the same model used OpenAI’s own integrated search tool and harness. Without search, the model scored 33.

The index is an equally weighted average of three benchmarks on a scale of 0 to 100. DeepSearchQA has 900 broad research tasks, and a language-model judge scores the answer lists using F1. AA used a subset of 200 hard tasks from BrowseComp, and AA-Omniscience is AA’s own unpublished subset of 600 questions.

The table shows the index score, costs and time per task for the leaders and the reference points. All figures were measured with a single answer model.

Search service Index Search / 1,000 tasks Model / 1,000 tasks Time / task
Parallel (advanced, objective) 83 $23.00 $7.04 24.1 s
Keenable (pro) 83 $31.87 $8.35 16.2 s
Keenable (realtime) 82 $31.94 $8.52 15.7 s
Parallel (fast, objective) 82 $4.63 $9.40 15.4 s
Exa (auto, dynamic highlights) 80 $53.17 $10.60 36.4 s
Perplexity (medium) 80 $62.30 $8.04 30.7 s
Octen (highlights) 77 $9.07 $14.50 15.3 s
Brave (LLM context) 75 $61.96 $16.95 22.1 s
OpenAI Web Search (medium) 74 $40.19 $9.40 32.8 s
DeepSeek (summary) 69 $8.47 $20.19 22.0 s
Model without search 33 – $2.92 25.8 s

Source: Artificial Analysis’s Search API page, 19 November 2026. Search costs are calculated at the providers’ list prices. Time includes measured search time and calculated model time; OpenAI’s search time is an upper bound because it includes OpenAI’s processing time. OpenAI Web Search used its own search tool and harness; the others used Stirrup.

A search service’s price alone does not show the total cost. DeepSeek’s search, for example, cost $8.47 per 1,000 tasks, but model costs came to $20.19. AA explains differences in model costs in general terms: services return different amounts of text, and weaker results make the model search and reason more. It does not break down the cause for individual services.

Parallel’s advanced search costs five times as much as fast ($5 versus $1 per 1,000 requests). Yet advanced used about 82,000 tokens per task and fast about 116,000. Almost all of these are input tokens, which include the search results returned to the model. Advanced cost about $0.030 per task in total and fast about $0.014. The median across the comparison was $0.063.

Previous coverage

Report an error in this story