SOTA LLM Pricing Comparison

API KEY is needed to chat PDF in PapersGPT for online LLMs

Almost all the mainstream LLMs(large language models) are supported in PapersGPT, and LLM API KEY should be provided or bought by yourself. While which LLM is the most suitable for you, which is more cost-effective, the following gives you some reference information.

What's the token?

Before making a price comparison, we first need to figure out what a token is. In the field of Artificial Intelligence and Natural Language Processing, Token is the basic unit of text after segmentation. The number of English words contained in a Token is not fixed. In English, common short words such as "the" "and" are a Token, while longer words such as "hesitation" are also a Token. As a rough estimate, on average, an English Token may correspond to 3–5 letters.

What's the SOTA(State Of The Art) Model?

The "SOTA LLM model" is a constantly moving target. And there isn't one single, universally declared "SOTA" (State-of-the-Art) LLM that definitively outclasses all others across every single metric. Instead, what's considered SOTA depends heavily on specific tasks, benchmarks, and evaluation criteria. Different models excel in different areas. However, Until August 2025, the most widely used and recognized top-tier smart models include:

GPT-6 family (OpenAI): The current generation spans GPT-6 Astra (the most intelligent model, $10/$50 per million tokens), GPT-6.1 Sol (the default workhorse, $2/$10) and GPT-6 Luna (an ultra-cheap small model at $0.10/$0.50). It excels in a broad range of tasks, including reasoning, coding, creative writing, complex instruction following, and general knowledge.

Gemini 4 Argon (Google): Google's newest flagship ($2/$10 until Nov 30, 2026, then $4/$20) sits at the top of major model evaluation rankings, with full multimodal input and a huge context window. The Gemini 3.x Flash series ($0.75/$3.75) delivers hybrid reasoning at a fraction of the price.

Claude (Anthropic): A very professional model that emphasizes practical usability. The current generation — Opus 5.5 ($4/$20), Sonnet 5.5 ($2/$10) and Haiku 5.5 (from $0.10/$0.50) — is best-in-class for long-document comprehension, writing, and vibe coding.

The smartest second-tier LLMs, approaching the first tier and developing rapidly, include:

DeepSeek: The V4 generation keeps DeepSeek's selling point — frontier-class intelligence with 1M-token context at a fraction of competing prices. V4 Pro runs $0.66-$1.32 input / $1.98-$3.96 output (off-peak / peak), and V4.1 Flash starts at $0.15/$0.60, with open-source lineage and aggressive off-peak discounting.

Mistral: From Europe, known for its multilingualism and open-source nature. Mistral Large is now priced at just $0.5/$1.5 per million tokens, with batch processing at 50% off and cached input up to 90% cheaper.

Kimi: The K-series excels at coding and agent scenarios — kimi-k3 ($3/$15 with a 1M-token context window) and kimi-k2.6 ($0.95/$4), with cache-hit input pricing as low as $0.16-$0.30.

There are many evaluation rankings for LLMs. Here recommending lmarena.ai. Its main feature is to use actual and relatively subjective manual evaluation as the standard. In actual conversations and usage scenarios, humans manually score the participating black-box LLMs.

Which is the most cost-effective Model?

Currently, in the paper reading scenario, most SOTA or near-SOTA models perform very well, unless the paper is very long or contains a large number of charts, etc. Therefore, it is very necessary to choose a LLM with good cost-effectiveness. Below is the latest SOTA LLM API pricing I compiled until August 2025.

The final decision is yours

Although there are various LLMs on the market, you don't need to worry about choosing. The best and most affordable LLM is determined by its practical use. You only need to choose one in PapersGPT and use it. If you think the effect is not good or the price is too high, you can switch to another model with one click. It is very convenient.