There was someone overseas who did a direct analysis of Claude 4.7's tokenizer.
Original source: https://www.claudecodecamp.com/p/i-measured-claude-4-7-s-new-tokenizer-here-s-what-it-costs-you
To summarize, here's what it comes down to:
Claude 4.7 New Tokenizer Analysis: Token Usage Increase and Real Cost Impact
Key Findings
According to Anthropic's official documentation, Claude 4.7's new tokenizer uses 1.0~1.35x more tokens compared to the existing 4.6. However, when measured with actual content, technical documents showed 1.47x increase, and the actual CLAUDE.md file showed up to 1.45x token usage increase. In other words, the upper end of the official range appears to be more typical in real-world usage environments.
What Token Increase Means
The price table remains the same, but the same prompt consumes more tokens.
For Max plan users, the context window depletes faster, cached prefix costs increase, and rate limits are hit sooner.
Ultimately, Anthropic sacrificed something to gain something else. The key question is what that is and whether it's worthwhile.
Ugh, seriously... these days after just a few prompts, you get hit with a 5-hour cooldown and it's scary to do anything.
Is anyone else experiencing this?
Experimental Method and Results
The author used the count_tokens API to compare token counts between 4.6 and 4.7 for identical content.
Actual Usage Content: 7 Sample Results
Content Type | 4.6 Tokens | 4.7 Tokens | Increase Rate |
|---|
CLAUDE.md File | 1,399 | 2,021 | 1.45x |
User Prompt | 1,122 | 1,541 | 1.37x |
Blog Post | 1,209 | 1,654 | 1.37x |
Git Commit Log | 910 | 1,223 | 1.34x |
Terminal Output | 652 | 842 | 1.29x |
Python Stack Trace | 1,736 | 2,170 | 1.25x |
Code Diff | 1,226 | 1,486 | 1.21x |
Overall Weighted Average: Approximately 1.33x Increase
Differences by Content Type
English and Code: 1.20~1.47x increase (technical documents show the highest)
CJK (Chinese, Japanese, Korean): 1.01x with almost no change
Structured Data such as JSON/CSV: 1.07~1.13x with relatively lower increase
Why Use More Tokens?
Based on analysis of tokenizer change patterns:
CJK, emojis, and symbols show lower token increase rates due to minimal vocabulary changes
English and code tend to be split into shorter subword units → increased token count
Code has many repeated keywords and identifiers, so they're likely split into smaller units during BPE training
In other words, it's estimated that the tokenizer was changed to represent the same text in smaller units.
Benefits of This Change: Improved Instruction Following Accuracy?
Anthropic highlighted "more accurate instruction following, particularly generalization without low effort level" as an advantage.
The author validated this using the IFEval benchmark (sampling 20 prompts):
Strict criteria: 4.6 (85%) → 4.7 (90%), +5 percentage points improvement
Loose criteria: Both models at 90% with no difference
Conclusion: There is a measurable but small improvement, though it's too modest to call it a "dramatic enhancement." With a small sample size and concurrent changes to weights and post-training, establishing causality is difficult.
Real Session Cost Simulation
Assuming an 80-turn coding session:
Item | Claude 4.6 | Claude 4.7 |
|---|
Cache Write (1x) | $0.05 | $0.06 |
Cache Read (79x) | $3.40 | $4.54 |
New Input | $0.20 | $0.26 |
Output Generation | $3.00 | $3.00~$3.90 |
Total | ~$6.65 | ~$7.86~$8.76 |
→ Approximately 20~30% cost increase per session. Though the token unit price is identical, the same task consumes more tokens.
Interaction with Prompt Cache
Cache Invalidation on Model Switch: Switching from 4.6 → 4.7 clears existing cache, and new cache creation costs increase by 1.3~1.45x
Cache Volume Increase: The tokens stored in cache itself increase, raising both read and write costs
Metric Changes: Even identical transcripts show different token counts, requiring caution in existing log-based analysis
Conclusion: Is This a Worthwhile Trade-off?
Cost: For English/code-focused work, token usage increases 1.3~1.45x → 20~30% session cost increase
Benefits: Strict instruction following accuracy improved by approximately +5 percentage points (small sample basis)
"Does this small accuracy improvement justify the added cost?" depends on your workload.
If precise instruction following is critical to your work, it might be worthwhile.
On the other hand, for general coding assistant use, the cost increase may feel more significant.
