Claude Opus 4.7's Flat Price Hides a 20–47% Cost Increase

Anthropic shipped Claude Opus 4.7 last week at the same $5-per-million-input-tokens and $25-per-million-output-tokens price as Opus 4.6. The press release called it "unchanged pricing." Developers paying the bills have a different view: they're seeing 20 to 47 percent higher costs on the exact same workloads.

The culprit isn't a hidden surcharge or a pricing change Anthropic forgot to mention. It's the tokenizer. Opus 4.7 uses a different tokenization scheme than 4.6, and on certain content types — especially code and technical documentation — it splits the same input into substantially more tokens. More tokens at the same per-token rate means higher bills.

Why this matters: This is a real-world lesson in how AI pricing actually works. "Per-token pricing" is only comparable across models if the tokenizer is the same. When the tokenizer changes, the sticker price can stay flat while your costs climb. Enterprise buyers signing annual contracts need to model token counts for their specific content types, not just check the pricing page.

What Developers Are Measuring

The clearest analysis comes from developer Abhishek Ray, who ran Opus 4.6 and Opus 4.7 side-by-side on identical content and published the token counts. His numbers, documented at Claude Code Camp:

Real-world Claude Code content: 1.325x more tokens on Opus 4.7. CLAUDE.md configuration files: 1.445x. Technical documentation: 1.47x — meaning the same doc costs nearly 50% more to process.

Ray also modeled the total cost on an 80-turn coding session, which is a realistic workload for an agent doing meaningful software work. Opus 4.6 ran the session for $6.65. Opus 4.7 ran the identical session for $7.86 to $8.76, depending on the content mix. That's a 20-30% bill increase with no price change and no additional capability used.

A separate community evaluation aggregated 483 side-by-side submissions and found an average 37.4% token-count increase across mixed content. Anthropic's own migration guide acknowledges a range of 1.0x to 1.35x, which falls at the low end of what independent measurements are showing.

Why the Tokenizer Changed

Anthropic hasn't published detailed technical documentation on the new tokenizer, but the likely motivation is model quality rather than revenue. Tokenizer design directly affects how models represent structured content — code syntax, file paths, identifier naming, mathematical notation. A tokenizer that produces finer-grained tokens for these content types can make training more sample-efficient and improve model reasoning on technical tasks.

That tracks with where Opus 4.7's capability gains are concentrated: software engineering and complex coding workflows. Anthropic's own benchmarks show 4.7 outperforming 4.6 on SWE-bench Verified and agent tool-use evaluations. If the new tokenizer is part of what's enabling those gains, it's a real engineering tradeoff — better performance costs more tokens.

The issue is how Anthropic framed the pricing. "Same price as 4.6" is technically true per token. It's misleading per request. A more accurate framing would acknowledge the token-count shift and let customers decide whether the capability improvements justify the effective cost increase.

Content Types Hit Hardest

The tokenizer change doesn't affect all content equally. Ray's measurements and the community evaluation both show:

Code files see the largest bumps — often 40-47% more tokens. This makes sense given the tokenizer's apparent optimization for fine-grained code representation.

Natural-language prose shows smaller increases, typically 15-25%. English documentation, email drafting, and conversational exchanges are less affected.

Chinese and Japanese text showed minimal to no change. The tokenizer optimization appears concentrated on English-language code and technical content.

For enterprise customers, the practical implication is that software engineering workloads — the ones Anthropic is pushing hardest into — are the ones most affected by the cost shift. Teams using Claude Code, agentic development tools, or coding assistants will see the largest bill increases.

What Opus 4.7 Actually Delivers

The performance gains are real but modest. Anthropic reports Opus 4.7 sticking to strict instructions five percentage points more reliably than 4.6 on IFEval. Vision resolution triples, which helps for document analysis and screenshot-heavy agent tasks. The new "xhigh" reasoning mode offers deeper thinking for hard problems at a higher compute cost.

For customers running Claude Code or building agents, the improvements compound. A model that follows instructions five points more reliably across a 20-turn agent loop is meaningfully less likely to go off the rails. The coding benchmark gains translate to fewer retries and less debugging.

The question isn't whether 4.7 is better than 4.6. It is. The question is whether it's enough better to justify a 20-30% cost increase per workload — and that depends entirely on what you're using it for.

Mitigations and Workarounds

Anthropic offers two meaningful cost-reduction mechanisms that apply to Opus 4.7: prompt caching, which can reduce input costs by up to 90% for repeated system prompts and context, and batch processing, which cuts costs by 50% for workloads that don't need real-time responses.

Teams running large agentic systems on Opus should audit their prompt caching configuration before accepting the higher bills as baseline. Poorly cached prompts on Opus 4.7 are substantially more expensive than well-cached prompts on Opus 4.6.

For workloads that don't benefit from 4.7's specific improvements — high-volume batch processing, straightforward transformations, content generation at scale — Opus 4.6 remains available via the API and will likely offer better economics for the foreseeable future.

The Bottom Line

Claude Opus 4.7 is not the free upgrade Anthropic's pricing page suggests. If you're running meaningful code or documentation workloads, expect your bill to rise 20-47% despite the unchanged per-token price. Model the impact on your specific content before committing to migration. And keep in mind: per-token pricing only compares across models that share a tokenizer. When tokenization changes, the price changes — the sticker just doesn't show it.