If your business is running Claude Sonnet 5 in production, you have 29 days to audit your AI costs before a compounding price change lands on September 1.

The mechanics are simple enough: Anthropic set introductory pricing of $2 per million input tokens and $10 per million output tokens when Sonnet 5 launched. On September 1, 2026, those rates revert to standard pricing of $3 per million input and $15 per million output. That is a 50% increase on both dimensions.

But there is a second factor that most teams have not fully priced in, and it makes the effective jump larger than the percentage suggests.

The Tokenizer Change Nobody Warned You About

When Anthropic shipped Claude Sonnet 5, they also shipped a new tokenizer. The new tokenizer produces approximately 30% more tokens for the same input text compared to Claude Sonnet 4.6. The exact increase depends on content type, but 30% is the documented figure for typical text.

This matters because per-token pricing is what you pay. If a request that generated 1,000 tokens on Sonnet 4.6 now generates 1,300 tokens on Sonnet 5 — same text, same task — you are already paying more than the token rate suggests, even at introductory pricing.

The compounding math looks like this:

• A workload that cost $1,000 per month on Claude Sonnet 4.6 might cost $1,300 on Sonnet 5 at intro pricing, due to the tokenizer increase alone.
• After September 1, that same workload runs at standard pricing on the inflated token count, pushing effective cost to roughly $1,950 per month or higher, depending on the actual tokenizer impact for your content.

Anthropic’s own documentation confirms the dynamic clearly: “The same input text produces approximately 30% more tokens than on Claude Sonnet 4.6. The cost of an equivalent request can differ from Claude Sonnet 4.6 even though per-token pricing is unchanged.”

What the September 1 Change Actually Affects

For businesses running Claude through the API directly, the rate change is automatic. From September 1, every request priced at $2/$10 moves to $3/$15. There is nothing to opt into or configure.

For businesses using Claude through AWS Bedrock, Google Cloud Vertex AI, or Microsoft Foundry, the same introductory period and deadline applies, though it is worth verifying your specific agreement terms with your cloud provider.

Prompt caching rates and batch processing rates also adjust, but the headline exposure is in standard synchronous requests.

What does not change on September 1 is the model itself. Sonnet 5 is an upgrade over Sonnet 4.6 in capability, particularly for coding and agentic tasks. The per-token price normalises to the same level as Sonnet 4.6’s standard rate. But the tokenizer increase means you are not paying the same effective cost per task, even at equal per-token rates.

Who Is Affected Most

Not every workload sees the same impact. The tokenizer change affects text-heavy inputs more than structured data, code, or tightly formatted prompts. If your Claude use is primarily conversational or document processing, expect the tokenizer inflation to land at the higher end of the 30% range.

High-volume, cost-sensitive applications are the ones to audit first. Customer-facing chatbots running millions of interactions, document processing pipelines handling large text volumes, and agentic workflows with long reasoning chains all sit in the impact zone.

If your team migrated from Sonnet 4.6 to Sonnet 5 and benchmarked costs at introductory pricing without accounting for the tokenizer, your September 1 bill will be different from what you expect.

What to Do Before August 31

The practical checklist is short.

Recount your tokens. Anthropic’s token counting API lets you measure prompts against Sonnet 5’s tokenizer directly. Any count taken against an earlier model is not accurate for Sonnet 5. For workloads that matter to your budget, run actual counts now rather than extrapolating from earlier benchmarks.

Revisit max_tokens budgets. If your application sets a hard limit on output length, a limit calibrated for Sonnet 4.6 may truncate equivalent output on Sonnet 5. Outputs that fit within a limit last week might not after the tokenizer math plays out.

Check your context window usage. Sonnet 5 has a 1M token context window, but each token covers less text than on earlier models. Prompts or documents that fit comfortably before may need adjustment.

Model your September 1 cost. Take your current monthly spend at introductory pricing, multiply by the standard rate ratio (1.5x on both input and output), then apply your measured tokenizer increase on top of that. The result is your September 1 baseline.

Decide on prompt caching. If your application sends repeated long system prompts or frequently reused context, prompt caching rates on Sonnet 5 can partially offset the tokenizer impact. If you are not already using caching, August is the right time to evaluate it.

What This Means for Business

AI running costs are a new operational category that most finance teams are still figuring out how to manage. The Claude Sonnet 5 situation is a useful example of why cost governance around AI infrastructure matters.

Two changes landed at the same time, each independently manageable, but easily missed in combination. The tokenizer change arrives with a model upgrade that most teams celebrate as a capability win. The pricing timeline is visible in the documentation but requires someone to be tracking it. In practice, neither change is unreasonable on its own. The issue is that both are happening at once, and the compounding effect is not immediately obvious from either announcement individually.

The businesses that will handle this well are the ones with someone responsible for AI cost operations, even informally. That does not require a dedicated FinOps function. It requires the same discipline that any well-run team applies to cloud spend: know what you are running, know what you are paying, and know when prices change.

If your organisation is deploying AI at scale and wants a clearer picture of costs, governance, and what comes next, book a discovery call with the Enterprise DNA team.