▌
Benchmarks & experiments
controlled comparisons and exploratory studies of search and research systems. each study documents its setup, evidence, and limitations.
Benchmarks
DSE-246oct 2026
The same GPT-6 Astra agent with Exa Search + Contents or DeepSeek Search + Fetch. Compare reviewed accuracy, completion time, and itemized costs on LiveBrowseComp and EvoBrowseComp.
246 tasks · 2 search stacks · 984 attempts · reviewed answer key
Astra-26sep 2026
One GPT-6 Astra agent, five web tool stacks. Exa and built-in web tie at 11/26 on LiveBrowseComp, with Keenable one answer behind. Explore accuracy, run duration, and every public task verdict.
26 tasks · 5 tool stacks · 130 runs · 30-call budget · formerly Astra-20
LBC-30aug 2026
Claude Code vs Codex CLI under three search backends (builtin, Exa MCP, Valyu MCP) on a frozen 30-task LiveBrowseComp subset. claude-exa wins at 60%; the search backend moves accuracy up to 33 points, the agent at most 7.
30 questions · 2 agents × 3 search backends · 180 runs, 0 failures · blinded BrowseComp grading
RB-30aug 2026
Five deep-research APIs plus a Perplexity calibration anchor on a frozen 30-task ResearcherBench subset. Coverage is a three-way race; faithfulness splits the field in two; parallel has the best overall profile.
30 questions · 6 systems · 180 runs · ~32k claims verified · 2 judge tracks
DRB2-20aug 2026
Nine deep research APIs on a frozen 20-task subset of DeepResearch-Bench-II. Exa effort tiers against each other, then one flagship tier per provider.
20 tasks · 20 domains · 1,504 rubrics · 180 runs
Experiments & findings
Exploratory results from bounded studies. Scope and uncertainty accompany each finding.
Published elsewhere
Benchmarks research.site does not host — catalogued with authorship, grading, and sourced scores.