DeepSeek V4 Flash
1,000 / 1,000 completed
- Total inference
- $3.99
- Per paper
- $0.004
- Relative
- 1.00×
The year in AI papers
1,000+ papers mapping AI research from 2025–2026.
Explore by research lab
View all labs01 Trending this year
The papers that moved AI forward
Selected for citation impact, official-code adoption, recency, and field-wide significance.
View all top papers
01 / 07
Explore the year
View all months →Aug2026, 3 papers, opens in a new tabJul2026, 84 papers, opens in a new tabJun2026, 86 papers, opens in a new tabMay2026, 115 papers, opens in a new tabApr2026, 77 papers, opens in a new tabMar2026, 100 papers, opens in a new tabFeb2026, 86 papers, opens in a new tabJan2026, 86 papers, opens in a new tabDec2025, 66 papers, opens in a new tabNov2025, 73 papers, opens in a new tabOct2025, 98 papers, opens in a new tabSep2025, 79 papers, opens in a new tabAug2025, 63 papers, opens in a new tab
1,000+papers
curated and indexed
curated and indexed
Picking a paper for you…
The cost benchmark
What 1,000 summaries cost.
Each model summarized the same extracted paper text with the same prompt, chunk boundaries, and output contract.
GPT-5.6 Luna
1,000 / 1,000 completed
- Total inference
- $6.00
- Per paper
- $0.006
- Relative
- 1.50×
Claude Haiku 4.5
1,000 / 1,000 completed
- Total inference
- $35.76
- Per paper
- $0.036
- Relative
- 8.95×
Completed-summary model inference only, using provider-reported token counts and standard prices frozen on August 5, 2026. This is a cost comparison, not a factual-quality ranking.
About & methodology
How we turned one year of AI research into a navigable atlas.
The project began with a simple cost experiment: what would it take to summarize 1,000 real AI papers? We froze the inputs, ran a controlled model comparison, and shaped the results into this field guide.
Freeze a reproducible corpus
We started with 8,262 discovery candidates from Hugging Face Daily Papers, guaranteed verified official research from five frontier labs, deduplicated by canonical arXiv ID, checked every version-pinned PDF, and froze exactly 1,000 papers published from August 4, 2025 through August 4, 2026.
Read every PDF
The final corpus contains 30,681 pages and 102.7 million extracted characters. Papers that fit the conservative context budget were summarized in one pass; oversized papers used a deterministic 50,000-character map-reduce flow instead of silent truncation.
Hold the comparison constant
DeepSeek V4 Flash, GPT-5.6 Luna, and Claude Haiku 4.5 received the same extracted text, chunk boundaries, prompt, and final-summary contract with reasoning disabled. Costs use provider-reported token counts and prices frozen on August 5, 2026.
Build the research atlas
We combined the 1,000-paper benchmark with 18 verified Together AI research papers, then added publication details, citations, official repositories, labs, and editorial topics to create this 1,018-paper atlas.
This is a reproducible, popularity-weighted field guide and cost benchmark—not an exhaustive history of AI research or a factual-quality ranking. Reported costs cover model inference only; downloading, PDF extraction, storage, networking, and development time are excluded. Thank you to arXiv for use of its open access interoperability.