Anthropic launched Claude Sonnet 5 at the same $3/$15 per million token rate as Sonnet 4.6. Developers saw “stable pricing” and started planning migrations. But the rate card hides two compounding cost drivers: a new tokenizer that inflates token counts by 27–42% for identical input, and adaptive thinking enabled at high effort by default that pushes output token usage to roughly 3x what Sonnet 4.6 generates. The real per-task cost is $2.29 versus $1.20. That is a 90% increase dressed up as a flat rate.
The Rate Card vs. Reality
The nominal pricing looks straightforward. Sonnet 5 launched with introductory rates through August 31, 2026, then shifts to standard rates identical to Sonnet 4.6. On paper, nothing changed after the intro window. In production, everything changed.
| </> | Sonnet 4.6 | Sonnet 5 (intro, to Aug 31) | Sonnet 5 (standard, from Sep 1) | Opus 4.8 |
|---|---|---|---|---|
| Input rate per 1M tokens | $3 | $2 | $3 | $5 |
| Output rate per 1M tokens | $15 | $10 | $15 | $25 |
| Token inflation vs. Sonnet 4.6 | baseline | +27–42% | +27–42% | different tokenizer |
| Effective input cost vs. Sonnet 4.6 | baseline | -33% to -27% | 0% to +17% | N/A |
| Effective output cost vs. Sonnet 4.6 | baseline | -33% to -20% | 0% to +17% | N/A |
| Verbosity (output tokens per task) | ~87M | ~300M | ~300M | lower |
| Real cost per Artificial Analysis task | ~$1.20 | ~$1.53 | $2.29 | $1.80 |
That last row is what matters. Artificial Analysis’s Intelligence Index data puts Sonnet 5 at $2.29 per task under standard pricing, ranked 101st out of 161 models for cost efficiency. Sonnet 4.6 comes in at approximately $1.20. Even Opus 4.8, which costs $5/$25 per million tokens, lands at $1.80 per task because it is far less verbose. The “same price” model costs nearly twice as much to run in practice.
The Tokenizer Change: What It Actually Does
Sonnet 5 ships with a new tokenizer. The same input text that produced a given token count on Sonnet 4.6 now produces more tokens on Sonnet 5. Simon Willison’s analysis at simonwillison.net documents the inflation by language and content type.
- English text: +42% more tokens than Sonnet 4.6
- Spanish text: +33% more tokens than Sonnet 4.6
- Python code: +27% more tokens than Sonnet 4.6
- Simplified Mandarin: +1% more tokens than Sonnet 4.6 (minimal change)
This matters for two distinct reasons. First, your input costs rise even if your text does not change. Second, any token budgets you set in system prompts or application logic are now calibrated against an inflated count. A 4,000-token context window on Sonnet 4.6 maps to roughly 5,680 tokens of capacity on Sonnet 5 for the same English content. Prompt templates, retrieval chunk sizes, and context trimming logic all need recalibration.
The Mandarin exception is notable. Developers building applications primarily in Chinese are largely unaffected on the input side. English-heavy workflows, multilingual products with Latin-script languages, and Python-heavy coding assistants carry the full inflation burden.
The Verbosity Multiplier
The tokenizer inflation affects input. The verbosity problem hits output, and output pricing is five times higher per token than input pricing. Sonnet 5 generates approximately 300 million output tokens per Artificial Analysis Intelligence Index task. The median across all 161 models in the index is roughly 87 million. Sonnet 5 produces more than three times the output of an average model on the same task.
The driver is adaptive thinking. Sonnet 5 defaults to effort: "high" on both the API and in Claude Code. At high effort, the model thinks extensively before responding. That thinking process generates visible output tokens, and those tokens are billed at full output rates. This is not a bug. It is the mechanism that makes Sonnet 5 score well on reasoning benchmarks. But it is also why the output cost per task balloons from what you might expect based on the rate card alone.
The compound effect is significant. Tokenizer inflation raises your input count. Verbosity raises your output count. Both multiply against a rate card that looks identical to Sonnet 4.6’s, producing a real per-task cost that is 90% higher.
The Concrete Math: A Developer Scenario
Consider a developer sending 1,000 API calls per day with an average input of 2,000 English words, which is approximately 2,700 tokens on Sonnet 4.6.
On Sonnet 4.6, the input side looks like this:
- 2,700 tokens per call times 1,000 calls = 2.7 million tokens per day
- At $3 per million tokens: $8.10 per day in input costs
On Sonnet 5 at standard pricing, the same input produces more tokens:
- 2,700 tokens times 1.42 (English inflation) = 3,834 tokens per call
- 3,834 tokens times 1,000 calls = 3.834 million tokens per day
- At $3 per million tokens: $11.50 per day in input costs
- Input cost increase: 42% higher for identical text
The output side is more pronounced. If your Sonnet 4.6 calls generate an average of 500 output tokens per call, Sonnet 5 at high effort may generate 1,500 or more. At $15 per million tokens, that difference compounds quickly across 1,000 daily calls. A workflow that costs $7.50 per day in output on Sonnet 4.6 can reach $22.50 or higher on Sonnet 5 at default settings. Combined with the inflated input cost, daily spend can jump from under $16 to over $34 for the same task volume.
The Effort Parameter: Your Main Cost Control Lever
Sonnet 5 removes the temperature, top_p, and top_k sampling parameters. The primary control lever is now effort. It accepts three values: "low", "medium", and "high". The default is "high".
Lowering the effort level substantially reduces output token generation because the model performs less extended thinking. Here is how to match effort level to task type:
- effort: “high” — Use for multi-step reasoning, complex code generation, research synthesis, and agentic tasks where correctness matters more than cost.
- effort: “medium” — Use for summarization, moderate-complexity completions, and tasks where a balance between reasoning depth and cost is appropriate.
- effort: “low” — Use for classification, short-form completions, one-shot lookups, entity extraction, and any task where the input constraints the answer tightly.
Most production workloads contain a mix of task types. Audit your call patterns before migrating. Identify which calls genuinely benefit from extended thinking and set effort: "low" or effort: "medium" explicitly on everything else. Leaving the default in place on simple classification calls is the fastest way to overspend on Sonnet 5.
The Intro Pricing Window: When Does Sonnet 5 Actually Make Sense?
The introductory $2/$10 per million token pricing runs through August 31, 2026. At those rates with tokenizer inflation factored in, Sonnet 5 is still cheaper per input token than Sonnet 4.6 for most content types. After September 1, the effective input cost for English-heavy workflows rises 0–17% above Sonnet 4.6 before verbosity is even counted.
| Use case | Now (intro pricing, to Aug 31) | After Sep 1 (standard pricing) |
|---|---|---|
| Complex reasoning, coding agents, multi-step tasks | Use Sonnet 5 with effort: “high” | Use Sonnet 5 only if quality gains justify $2.29/task |
| Summarization, content generation, moderate tasks | Use Sonnet 5 with effort: “medium” | Benchmark Sonnet 4.6 vs. Sonnet 5 at medium effort |
| Classification, extraction, short completions | Use Sonnet 5 with effort: “low” | Stay on Sonnet 4.6 or use Sonnet 5 at effort: “low” |
| Mandarin-primary workflows | Use Sonnet 5 freely (minimal tokenizer impact) | Evaluate: tokenizer neutral, verbosity still applies |
| High-volume, cost-sensitive pipelines | Test now, measure actual token counts | Do not migrate without explicit effort controls in place |
The intro window is a genuine opportunity to benchmark Sonnet 5 at reduced cost. Use it to measure actual token counts on your production data, not synthetic benchmarks. Set explicit effort levels on every call. Record the output token counts per task type. That data will determine whether the model’s quality improvements justify the post-September cost structure for each workload independently.
The Decoder summarized the pattern accurately: Anthropic continues to hide effective price increases behind unchanged token rates. The mechanism this cycle is the tokenizer change combined with a verbose default effort setting. Developers who treat the rate card as the full picture will see budget overruns start September 1.
The Bottom Line
“Same price” is not the same cost. Claude Sonnet 5 carries a $3/$15 rate card identical to Sonnet 4.6, but the new tokenizer inflates English token counts by 42%, and adaptive thinking at high effort triples output token generation. The real per-task cost is $2.29 versus $1.20 on Sonnet 4.6. Before September 1, the introductory pricing partially offsets the tokenizer inflation and makes Sonnet 5 worth testing. After September 1, every production workload running at default settings will pay roughly 90% more per task. Audit your call patterns now. Set explicit effort levels. Measure actual token counts on your data. Do not migrate high-volume pipelines blindly based on a rate card that does not reflect what you will actually pay.
