Simon Willison upgraded his Claude Token Counter to compare models side by side, and the results sting. Claude Opus 4.7 ships with a new tokenizer that inflates token counts by 1.46x for system prompts compared to Opus 4.6. Since both models charge the same $5 per million input tokens and $25 per million output tokens, that means Opus 4.7 is roughly 40% more expensive for identical text. Anthropic acknowledged the tradeoff, saying the same input can map to "roughly 1.0–1.35x" more tokens depending on content type. Willison's testing came in well above that upper bound.
Images tell a more dramatic story. Willison tested a 3456x2234 pixel PNG and found Opus 4.7 used 3.01x the tokens compared to 4.6. That spike comes from Opus 4.7 supporting much higher resolution images, up to 2,576 pixels on the long edge versus previous models' tighter caps. When he tested a small 682x318 pixel image, the counts were nearly identical: 314 tokens for 4.7 versus 310 for 4.6. The cost explosion only hits when you actually use the new resolution headroom.
This is a strange move in an industry racing toward cheaper inference. OpenAI optimized its tokenizer in GPT-4o to reduce token counts and cut effective costs. Anthropic went the other direction, keeping per-token pricing the same while making real-world bills bigger. A 30-page text-heavy PDF showed a more modest 1.08x increase (60,934 tokens versus 56,482), so the impact varies by workload. For developers building Anthropic Managed Agents that hammer the API with long system prompts, this is a real cost increase that compounds fast. Measure your own usage before assuming Opus 4.7 is a drop-in upgrade.