Opus 4.7: Community Data Shows 39% Increase in Token Costs

Community data comparing Opus 4.6 and 4.7 shows a significant 38.9% average increase in token usage and costs. Based on 164 submissions, the average request size has expanded from 246 to 337 tokens. This indicates that the newer model version is considerably more expensive to run for identical inputs.

Key Points

• Opus 4.7 consumes an average of 38.9% more tokens than Opus 4.6 for the same inputs.
• The financial cost of API requests has increased by 38.9% in tandem with the token growth.
• Average request sizes have grown from 246 tokens to 337 tokens across 164 community submissions.
• Individual data points show a range of token increases, with some specific requests jumping by as much as 46.7%.

Sentiment

The overall sentiment is predominantly negative and frustrated. The community largely agrees that input costs have measurably increased and that forced adaptive thinking is a step backward. However, there is a meaningful minority pushing back on the framing, arguing that input token counts alone are misleading and that total task costs may not have increased as dramatically. The discussion is more nuanced than pure outrage, with some users sharing positive experiences and others offering technical explanations for the observed behavior.

In Agreement

• The 39% input token increase is real and confirmed by community data, meaning the same prompts cost significantly more to process with Opus 4.7
• Users on various subscription tiers report burning through usage limits much faster than with previous Opus versions, consistent with the token inflation data
• Opus 4.7 replacing both 4.5 and 4.6 in VS Code Copilot with a 7.5x multiplier represents a price hike for many workflows
• The forced adaptive thinking in 4.7 compounds the cost issue by sometimes producing lower-quality output that requires additional correction rounds

Opposed

• The data only measures input tokenizer differences and ignores output and reasoning tokens, where 4.7 may be cheaper — the right metric is cost per completed task, not cost per input token
• Artificial Analysis benchmarks show 4.7 cost about 11% less than 4.6 overall to complete their test suite, suggesting net costs may actually be lower for some workloads
• Much of the increased usage consumption may be due to the higher default effort level (xhigh vs old medium), and switching back to medium largely resolves the issue
• A smarter model that completes tasks in fewer turns could offset higher per-token costs through reduced total token usage
• Attributing the cost increase to an intentional Anthropic strategy to squeeze customers is uncharitable — Anthropic is a PBC with an AI safety mission and competitive incentives to maintain quality