Can I ask an LLM to find papers?

Use a real index (Semantic Scholar, etc.). LLMs invent references. Then summarize the PDFs you actually downloaded.

Which model is best for papers?

The cheapest one that obeys “only from these excerpts” on your eval set, with enough context or RAG to see the paper.

Should I use long context or RAG?

Single papers: long context is simple. A corpus: RAG. See both guides.

Is Claude better for careful hedging?

Often in style. Still ground it. Style is not a bibliography.

Best LLM for research — literature, notes, caveats

Research assistants fail by inventing citations and overstating papers they did not read. The best LLM for research is the one that stays inside the PDFs you give it and is cheap enough to run on every paper in a folder.

Combine a long-context or RAG setup with a model that follows “quote or don’t claim.” Quality index helps a little; citation faithfulness is the eval that matters. Pair this with the RAG and summarization guides.

Longest context windows right now

Live OpenRouter pricing and Arena quality, cached about an hour. Not a static blog table.

Model	Provider	Blend $ / 1M	Quality	Context
Grok 4.20 Multi-Agent	xAI	1.88	—	2,000,000
Grok 4.20	xAI	1.88	—	2,000,000
Auto Router (Beta)	OpenRouter	—	—	2,000,000
Pareto Code Router	OpenRouter	—	—	2,000,000
Auto Router	OpenRouter	—	—	2,000,000

Never trust a fabricated DOI

If the prompt allows the open web in the model’s head, you will get fake papers. Ground in retrieved PDFs. Ask for quotes with page numbers from the prompt context only.

Literature review at scale

A cheap model that extracts structured notes (problem, method, result) from each PDF beats a flagship that writes a rhetorical essay from three abstracts. Use cost ranking for the extract step; use a stronger model only to synthesize the notes.

Math and proofs

Reasoning models may help, but they still err. Treat outputs as conjectures. See o-series vs Claude if the work is proof-heavy.

Multimodal figures

Papers are figures. A vision-capable model or a captioning preprocess is required. Text-only rows will skip the actual result plot.

Cost

A lab group summarizing 500 papers should look at 100M-token monthly cost on model pages. That number decides whether you use flash or flagship for extraction.

FAQ

Can I ask an LLM to find papers?: Use a real index (Semantic Scholar, etc.). LLMs invent references. Then summarize the PDFs you actually downloaded.
Which model is best for papers?: The cheapest one that obeys “only from these excerpts” on your eval set, with enough context or RAG to see the paper.
Should I use long context or RAG?: Single papers: long context is simple. A corpus: RAG. See both guides.
Is Claude better for careful hedging?: Often in style. Still ground it. Style is not a bibliography.

best LLM for researchAI for academic researchLLM for papersbest model for literature review

Rank by context and quality