It’s a well-known indisputable fact that completely different mannequin households can use completely different tokenizers. Nonetheless, there was restricted evaluation on how the method of “tokenization” itself varies throughout these tokenizers. Do all tokenizers end in the identical variety of tokens for a given enter textual content? If not, how completely different are the generated tokens? How vital are the variations?
On this article, we discover these questions and study the sensible implications of tokenization variability. We current a comparative story of two frontier mannequin households: OpenAI’s ChatGPT vs Anthropic’s Claude. Though their marketed “cost-per-token” figures are extremely aggressive, experiments reveal that Anthropic fashions might be 20–30% costlier than GPT fashions.
API Pricing — Claude 3.5 Sonnet vs GPT-4o
As of June 2024, the pricing construction for these two superior frontier fashions is extremely aggressive. Each Anthropic’s Claude 3.5 Sonnet and OpenAI’s GPT-4o have equivalent prices for output tokens, whereas Claude 3.5 Sonnet provides a 40% decrease price for enter tokens.
Supply: Vantage
The hidden “tokenizer inefficiency”
Regardless of decrease enter token charges of the Anthropic mannequin, we noticed that the entire prices of operating experiments (on a given set of fastened prompts) with GPT-4o is less expensive when in comparison with Claude Sonnet-3.5.
Why?
The Anthropic tokenizer tends to interrupt down the identical enter into extra tokens in comparison with OpenAI’s tokenizer. Because of this, for equivalent prompts, Anthropic fashions produce significantly extra tokens than their OpenAI counterparts. Consequently, whereas the per-token price for Claude 3.5 Sonnet’s enter could also be decrease, the elevated tokenization can offset these financial savings, resulting in greater total prices in sensible use instances.
This hidden price stems from the best way Anthropic’s tokenizer encodes data, typically utilizing extra tokens to symbolize the identical content material. The token rely inflation has a big influence on prices and context window utilization.
Area-dependent tokenization inefficiency
Various kinds of area content material are tokenized in a different way by Anthropic’s tokenizer, resulting in various ranges of elevated token counts in comparison with OpenAI’s fashions. The AI analysis neighborhood has famous comparable tokenization variations right here. We examined our findings on three in style domains, particularly: English articles, code (Python) and math.
| Area | Mannequin Enter | GPT Tokens | Claude Tokens | % Token Overhead |
| English articles |  | 77 | 89 | ~16% |
| Code (Python) |  | 60 | 78 | ~30% |
| Math |  | 114 | 138 | ~21% |
% Token Overhead of Claude 3.5 Sonnet Tokenizer (relative to GPT-4o) Supply: Lavanya Gupta
When evaluating Claude 3.5 Sonnet to GPT-4o, the diploma of tokenizer inefficiency varies considerably throughout content material domains. For English articles, Claude’s tokenizer produces roughly 16% extra tokens than GPT-4o for a similar enter textual content. This overhead will increase sharply with extra structured or technical content material: for mathematical equations, the overhead stands at 21%, and for Python code, Claude generates 30% extra tokens.
This variation arises as a result of some content material sorts, reminiscent of technical paperwork and code, typically include patterns and symbols that Anthropic’s tokenizer fragments into smaller items, resulting in the next token rely. In distinction, extra pure language content material tends to exhibit a decrease token overhead.
Different sensible implications of tokenizer inefficiency
Past the direct implication on prices, there may be additionally an oblique influence on the context window utilization. Whereas Anthropic fashions declare a bigger context window of 200K tokens, versus OpenAI’s 128K tokens, attributable to verbosity, the efficient usable token area could also be smaller for Anthropic fashions. Therefore, there may probably be a small or giant distinction within the “marketed” context window sizes vs the “efficient” context window sizes.
Implementation of tokenizers
GPT fashions use Byte Pair Encoding (BPE), which merges often co-occurring character pairs to type tokens. Particularly, the newest GPT fashions use the open-source o200k_base tokenizer. The precise tokens utilized by GPT-4o (within the tiktoken tokenizer) might be considered right here.
JSON
{
#reasoning
"o1-xxx": "o200k_base",
"o3-xxx": "o200k_base",
# chat
"chatgpt-4o-": "o200k_base",
"gpt-4o-xxx": "o200k_base", # e.g., gpt-4o-2024-05-13
"gpt-4-xxx": "cl100k_base", # e.g., gpt-4-0314, and many others., plus gpt-4-32k
"gpt-3.5-turbo-xxx": "cl100k_base", # e.g, gpt-3.5-turbo-0301, -0401, and many others.
}
Sadly, not a lot might be stated about Anthropic tokenizers as their tokenizer will not be as instantly and simply out there as GPT. Anthropic launched their Token Counting API in Dec 2024. Nonetheless, it was quickly demised in later 2025 variations.
Latenode studies that “Anthropic makes use of a novel tokenizer with solely 65,000 token variations, in comparison with OpenAI’s 100,261 token variations for GPT-4.” This Colab pocket book accommodates Python code to investigate the tokenization variations between GPT and Claude fashions. One other software that permits interfacing with some widespread, publicly out there tokenizers validates our findings.
The power to proactively estimate token counts (with out invoking the precise mannequin API) and price range prices is essential for AI enterprises.
Key Takeaways
- Anthropic’s aggressive pricing comes with hidden prices:
Whereas Anthropic’s Claude 3.5 Sonnet provides 40% decrease enter token prices in comparison with OpenAI’s GPT-4o, this obvious price benefit might be deceptive attributable to variations in how enter textual content is tokenized. - Hidden “tokenizer inefficiency”:
Anthropic fashions are inherently extra verbose. For companies that course of giant volumes of textual content, understanding this discrepancy is essential when evaluating the true price of deploying fashions. - Area-dependent tokenizer inefficiency:
When selecting between OpenAI and Anthropic fashions, consider the character of your enter textual content. For pure language duties, the fee distinction could also be minimal, however technical or structured domains might result in considerably greater prices with Anthropic fashions. - Efficient context window:
As a result of verbosity of Anthropic’s tokenizer, its bigger marketed 200K context window might provide much less efficient usable area than OpenAI’s 128K, resulting in a potential hole between marketed and precise context window.
Anthropic didn’t reply to VentureBeat’s requests for remark by press time. We’ll replace the story in the event that they reply.
