Anthropic Tokenizer Can Add Up To 73 Percent More Claude Tokens
The Register reported that Playcode found Anthropic's newer Claude tokenizer can emit up to 73 percent more tokens than OpenAI's GPT-5.x family on a TypeScript file. The same report said Anthropic's introductory Sonnet pricing runs through August 31, 2026, while independent customer-bill and task-completion evidence remains absent.
Anthropic's newer Claude tokenizer can make the same code file consume materially more tokens than OpenAI's GPT-5.x family, The Register reported, putting tokenizer behaviour into enterprise AI cost planning rather than treating it as a technical footnote.
In the Playcode test cited by the publication, Claude's token output for a 2,888-character TypeScript file was 73% higher than the GPT-5.x family from OpenAI.
Anthropic's own documentation notes that the updated tokenizer can map the same input to roughly 1.0-1.35x more tokens depending on content type.
Playcode Found Claude Used 1.73x More Tokens On TypeScript
Playcode compared Anthropic's latest tokenizer with OpenAI's GPT-5.x tokenizer across code samples.
In that comparison, the TypeScript example produced 1.73x more Claude tokens than the GPT-5.x tokenizer and 1.32x more than Claude's older tokenizer.
Other code examples came in lower but still above the OpenAI baseline.
Rust measured 1.58x, JavaScript 1.52x and Python 1.50x in the comparison cited by The Register.
Providers charge for model input and output by token.
Token slicing varies by model, which can make bills hard to predict when the same prompt, file or workflow is moved between model families.
Anthropic Sets Introductory Sonnet Pricing Through August 31, 2026
Anthropic acknowledged the tokenizer change when Sonnet 5 shipped at the end of June.
Its note states that Sonnet 5 uses an updated tokenizer similar to the one introduced with Claude Opus 4.7.
The introductory Sonnet schedule runs until August 31, 2026, with Anthropic listing $2 per million input tokens and $10 per million output tokens.
After that date, the listed rates move to $3 per million input tokens and $15 per million output tokens.
Anthropic's own note warns that the new tokenizer may raise bills by as much as a third compared with the tokenizer used by older company models.
For engineering teams, that means model migration tests need to measure tokenisation as well as nominal per-token prices.
Ploy Account Claimed GPT-5.6 Finished Pages Faster
The Register also cited marketing platform Ploy, which described moving production work across OpenAI's GPT-5.6 Sol and Anthropic's Opus 4.8.
In that account, Ploy claimed GPT-5.6 finished pages 2.2x faster, cost 27% less and used about half the output tokens; The Register did not include an independent benchmark methodology for the claim.
Those performance and cost claims remain vendor- or platform-provided evidence rather than independent benchmark results.
AI costs should also be judged by task completion and by the model harnesses used around systems such as Claude Code, Codex, Pi and OpenCode.
The practical lesson is that token counts are part of model economics.
The Register did not include Anthropic customer bills, independent benchmark methodology, audited task-completion rates or enterprise deployment data for the tokenizer change.
