Best AI for research

Who this page is for

This page is for people who use AI models to gather, read and summarise information: literature reviews, market and competitor research, fact-finding across long reports, or answering questions that need current sources from the web. It covers models you can use through an API or a chat app. It does not cover dedicated research products built on top of these models.

Using a chat app rather than an API? The AI subscriptions scored for research compare consumer plans per price tier on model quality, deep research, web search, usage and price.

How this ranking works

The ranking combines six criteria into one score. The weights are fixed and add up to 100%. Our methodology explains how each criterion is scaled and combined.

• Humanity's Last Exam without tools, 35%. HLE is a set of 2,500 expert-written questions across many academic subjects, organised by the Center for AI Safety and Scale AI ^[1]. Without tools, it measures what a model knows and how well it reasons on hard expert questions.
• Humanity's Last Exam with tools, 20%. The same questions with web search or code execution allowed, which is closer to how a research assistant works.
• Context window, 15%. Research means reading long papers and many sources at once. Larger is better, on a log scale.
• Document and PDF input, 10%. Uploading papers and reports directly saves copying text by hand.
• Reasoning mode, 5%. Step-by-step reasoning helps with multi-part research questions.
• Blended price, 15%. Long documents use many tokens, so price matters, but less than answer quality. Lower is better, on a log scale.

Which HLE results count

HLE results depend heavily on how a model was run, so each HLE criterion uses a setting group: a list of stored settings we treat as comparable for this ranking.

• Without tools: full-question-set runs without tools count, at any reasoning effort and with any grader. Runs on a subset of the questions, and runs whose tool use is not stated, are left out.
• With tools: runs with tools count, at any reasoning effort, whichever tools the run allowed, such as web search, web fetch or code execution.

Runs with and without tools are never mixed. Mixing them would mostly measure who used tools: Anthropic reports 64.4% for Claude Opus 5.5 without tools and 67.7% with web search, web fetch and code execution ^[2], and DeepSeek reports 36.8% without tools and 63.9% with tools for the first release of DeepSeek V4.1 Flash (September 2026) ^[3].

Within each group, a model counts with its best result. Independent results are preferred over results a provider reports itself; if only a provider result exists, it is used and marked "self-reported". Runs from the Scale Labs leaderboard ^[4] are not part of either group, because their stored settings do not state whether tools were used. The methodology box on this page lists every stored setting in each group.

A model needs data for at least 60% of the total weight to be ranked. Context, file input, reasoning and price together make up 45%, so a model needs at least one HLE result in the groups.

How to read the results

• Scores are relative. Each criterion is rescaled from 0 to 100 across the candidates on this page. 100 means best among these models, not perfect, and adding or removing a model can shift the others.
• Check the setting. The table shows the setting behind each score, including effort, tools and grader where the source states them.
• "Self-reported" marks a result that comes from the model's own provider, run and graded on a setup the provider chose.
• Small gaps are noise. Graders, harnesses and efforts differ inside each group, so treat models a few points apart as roughly equal.
• Capabilities barely separate the leaders. Many current models offer file input, a reasoning mode and a context window of about one million tokens, so the benchmark criteria and price drive most of the differences.

Limitations

Grouping lets us compare more models, at the cost of comparing runs that were not set up identically:

• Different tool sets. Google's with-tools result for Gemini 3.1 Pro (51.4%) used search with a blocklist plus code ^[5], Anthropic's runs used web search and fetch with code execution ^[2], and OpenAI reports the best of several effort levels for GPT-6 Astra (57.2%) ^[6].
• Different graders. The public Scale Labs leaderboard uses o3-mini as the judge ^[4]; Anthropic grades its own runs with Claude Opus 4.6 ^[2]. The same model can land far apart: Anthropic reports 60.9% for Claude Fable 5.1 without tools at max effort ^[7], while the Scale Labs leaderboard lists 46.5% at xhigh effort, with tool use not stated ^[4].
• Mostly self-reported. Most HLE results in the groups come from the providers, each under its own harness, grader and effort level.
• Narrow measure. HLE tests answers to hard questions, not whether a model cites sources correctly. Our capability data says whether a feature exists, not how well it works.
• Coverage. Only models with an HLE result in at least one group can be ranked. A missing result doesn't mean a model is weak, only that we have no comparable measurement.
• The data behind this page is as of October 2026. Our benchmark pages show every result with its setting.

How to choose for your own work

1. Check for web search. If your research needs current information, the model must be able to search. Some providers offer it as an API tool: Anthropic documents a web search tool for Claude ^[8], and OpenAI and Google list search support on their model pages ^{[9] [10]}. Web search is not a ranking criterion, so check it on the model pages.
2. Check file input and context. For reading reports and papers, file input and a large context window matter. A large window is a ceiling, not a guarantee of recall; see context windows explained.
3. Decide whether you need reasoning. Reasoning modes help with multi-step questions but cost more output tokens; see reasoning models explained.
4. Estimate the cost. Research means long inputs. Compare list prices on AI pricing and run your expected volume through the cost calculator.
5. Test on your own material. Give two or three candidates the same questions with sources you already know, and check every citation. Benchmarks do not tell you how often a model misquotes a source.

Related guides

• How to choose an AI model
• How to compare AI models
• Multimodal AI explained
• AI token pricing explained
• Reduce AI API costs