Best AI model for PDFs — November 2026
No single model is clearly best for PDFs on BenchLM's evidence. On AA-LCR, run by Artificial Analysis, Kimi K3.1 Pro (86.3%) and MiMo-V2.6-Pro (86.3%) are within 1.5 points of each other. The test uses extracted text. File upload, scanned pages, tables and figures depend on the app and the document format.
The shortlist for PDFs
Constraints on this page: a builder choosing by accuracy, hosted processing allowed, any price, ordinary input size. Change any of them under Refine. A small score gap does not establish a reliably better model.
01Kimi K3.1 ProTied at the top
Moonshot AI · Open Weight
86.3%
AA-LCR
- Highest AA-LCR result, with 1 other model within 1.5 points
- $4.00 for the stated workload
- 1.05M context
- Open weights; check hardware before treating local use as feasible
- API access is not verified. Check the provider documentation before relying on it.
02MiMo-V2.6-ProTied at the top
Xiaomi · Open Weight
86.3%
AA-LCR
- 86.3% on AA-LCR, run by Artificial Analysis
- $0.61 for the stated workload
- 1M context
- Open weights; check hardware before treating local use as feasible
- API access is not verified. Check the provider documentation before relying on it.
Model page and evidenceCompare with Kimi K3.1 ProOpen benchmark source
03Claude Opus 5.5
Anthropic · Proprietary
84.7%
AA-LCR
- 84.7% on AA-LCR, run by Artificial Analysis
- $8.00 for the stated workload
- 1M context
- Anthropic lists the model ID on each documented platform. Access details · checked September 22, 2026.
- Provider API offer: Claude API. Check which models qualify and whether payment details are required. Limits and source.
Model page and evidenceCompare with Kimi K3.1 ProOpen benchmark source
04DeepSeek V4.1 Flash 1017
DeepSeek · Open Weight
84.0%
AA-LCR
- 84.0% on AA-LCR, run by Artificial Analysis
- $0.54 for the stated workload
- 1M context
- Open weights; check hardware before treating local use as feasible
- Served as deepseek-flash on the DeepSeek API since September 10, 2026 with thinking and non-thinking modes, tool calls, the Responses and Anthropic formats, and image input. Since October 17, 2026 the ID serves DeepSeek-V4.1-Flash-1017 with the calling method and pricing unchanged. MIT-licensed weights for both checkpoints are published on Hugging Face. The legacy deepseek-v4-flash and deepseek-v4-flash-vision-exp IDs temporarily route to this model. DeepSeek announced that deepseek-v4-pro would join them from 04:00 UTC on September 14, 2026, then stated on its pricing page, read September 17, 2026, that V4 Pro service continues with billing unchanged. Access details · checked October 17, 2026.
- Provider API offer: DeepSeek API. Check which models qualify and whether payment details are required. Limits and source.
Model page and evidenceCompare with Kimi K3.1 ProOpen benchmark source
05Muse Wonder Full
Meta · Proprietary
83.7%
AA-LCR
- 83.7% on AA-LCR, run by Artificial Analysis
- 1M context
- API access is not verified. Check the provider documentation before relying on it.
Model page and evidenceCompare with Kimi K3.1 ProOpen benchmark source
Not ranked: replaced by a newer model
- Kimi K3 scored 88.7% here. Its lab has replaced it with Kimi K3.1 Pro, which scored 86.3%.
ChatGPT, Claude, Gemini, or Copilot?
No app is clearly ahead here. On AA-LCR, ChatGPT’s GPT-6 Luna (83.3%) and Claude Opus 5.5 (84.7%) are within 1.5 points of each other.
- ChatGPTGPT-6 Luna scores 83.3% (highest of the three)Plus, $20 a monthPlan lists: GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Sol and GPT-6 Luna in Codex; GPT-6.1 Sol is available in Work and Codex, not Chat.
- ClaudeClaude Opus 5.5 scores 84.7% (highest of the three)Pro, $20 a monthPlan lists: Opus, Sonnet and Haiku; Fable is listed under usage credits rather than the included allowance.
- GeminiGemini 3.8 Flash scores 81.3%Google AI Pro, $19.99 a monthPlan lists: Gemini 4 Argon, the Pro model and the Flash model.
- Microsoft CopilotNot scored. Microsoft doesn’t name the model Copilot runs, so no benchmark result can be attached to it. How it compares on plans and data termsMicrosoft 365 Premium, $19.99 a month
Each entry is the highest-scoring model that company sells, not the model your plan opens by default, so check what the plan lists before you pay. The plan shown is each app’s plan nearest $20, read October 16, 2026 for ChatGPT, November 10, 2026 for Copilot, and November 11, 2026 for Claude and Gemini.
ChatGPT vs ClaudeChatGPT vs GeminiClaude vs GeminiChatGPT vs Copilot
The top ten for PDFs, with API prices
The first ten models on the same evidence and constraints as the shortlist, with each provider's API price in US dollars per million tokens. GPT-6 Luna costs the least of the ten for a workload of 1M input and 200k output tokens: $0.10 per million input tokens and $0.50 per million output tokens. Two of the ten have no API price on file.
| # | Model | AA-LCR | Input | Output |
|---|---|---|---|---|
| 01 | Kimi K3.1 ProMoonshot AI | 86.3% | $2.00 | $10.00 |
| 02 | MiMo-V2.6-ProXiaomi | 86.3% | $0.435 | $0.87 |
| 03 | Claude Opus 5.5Anthropic | 84.7% | $4.00 | $20.00 |
| 04 | DeepSeek V4.1 Flash 1017DeepSeek | 84.0% | $0.30 | $1.20 |
| 05 | Muse Wonder FullMeta | 83.7% | No API price on file | |
| 06 | Claude Fable 5.2Anthropic | 83.3% | $8.00 | $40.00 |
| 07 | GPT-5.3 CodexOpenAI | 83.3% | $1.75 | $14.00 |
| 08 | GPT-6 LunaOpenAI | 83.3% | $0.10 | $0.50 |
| 09 | Inkling 2 SmallThinking Machines Lab | 83.3% | $0.30 | $1.00 |
| 10 | Muse Glimmer 30BMeta | 83.3% | No API price on file | |
Prices are list API rates before caching, batch discounts and special tiers. Compare API pricing for every model.
Refine for your situation
Each link opens the selector with one answer changed. The address carries the answers, so your version is as shareable as this page.
- Change any answer in the selector
- Must run on my own hardware
- Cost matters most
- Speed matters most
- Long documents or a large codebase
- Choosing an everyday assistant, not an API
- Describe your own job instead
We keep what you type here, drafts included, for 180 days, and after that only the job and requirements read from it. Nothing is tied to your IP address or account. Leave out names and other personal details. Privacy
What to verify before choosing
- The test uses extracted text. File upload, scanned pages, tables and figures depend on the app and the document format.
- Ranked on one benchmark: AA-LCR, run by Artificial Analysis. 192 models have a published result; a model without one is not listed.
Try three representative examples of your own work. Compare errors, time, cost, and the tools available in your actual setup.
What this shortlist rests on
One benchmark: AA-LCR. Questions that can only be answered by reasoning across several long real-world documents, about 100,000 tokens of text per question. A result a model's own provider ran is left out, and a model nobody has run there is not listed.
- AA-LCRSets the order. Run by Artificial Analysis. 192 models have a published result.
- GDP.pdfRelated evidence, not used for the order. Professional tasks that start from PDF files, run by Artificial Analysis on fewer models.
- LongBench v2Related evidence, not used for the order. Long-context questions across documents, code and dialogue.
Questions
Which AI model is best for PDFs?
No single model is clearly best for PDFs on BenchLM's evidence. On AA-LCR, run by Artificial Analysis, Kimi K3.1 Pro (86.3%) and MiMo-V2.6-Pro (86.3%) are within 1.5 points of each other. The test uses extracted text. File upload, scanned pages, tables and figures depend on the app and the document format.
Which models come next for PDFs?
Claude Opus 5.5 (84.7%), DeepSeek V4.1 Flash 1017 (84.0%), Muse Wonder Full (83.7%) follow on the same benchmark. Compare the leaders on three representative examples of your own work.
How does BenchLM pick the best ai model for pdfs?
The page runs the LLM Selector with fixed answers: a builder choosing by accuracy, hosted processing allowed, any price, ordinary input size. The selector uses AA-LCR, one benchmark run by Artificial Analysis on every listed model, filters by the stated constraints, and orders by that evidence. It never adds a hidden fit score or a bonus for open weights or reasoning style.
Why does this page use AA-LCR and not a general score?
A general score answers several jobs at once. AA-LCR is closer to this job. Questions that can only be answered by reasoning across several long real-world documents, about 100,000 tokens of text per question. Only results Artificial Analysis published itself are used, so a model nobody has run there is not listed. 192 models have a result.
What do the top models for PDFs cost?
Among the ten highest-ranked models on this page, GPT-6 Luna costs the least for a workload of 1M input and 200k output tokens: $0.10 per million input tokens and $0.50 per million output tokens. Claude Fable 5.2 costs the most, at $8.00 and $40.00. These are API rates in the November 13, 2026 data. Assistant subscriptions are priced separately.
Can I change the constraints?
Yes. Every link under "Refine" opens the selector with one answer changed, and the address carries the answers so a result can be shared or reopened against the current dataset.
Method: bench-align-v5.8-2026-09-30. Read the methodology and benchmark confidence pages for how scores and verification statuses are produced.
Watch the PDFs shortlist
One weekly email when rank, price, or benchmark evidence changes make this shortlist worth revisiting.
Read a sample issueJoin 7,000+ subscribers.