Claude Sonnet 5: promotional price until August 31st changes the cost per task? Do the math before migrating

Claude Sonnet 5 costs less per token until August 31, 2026, but this is not enough to conclude that the migration will be cheaper. The model uses a new tokenizer, can consume more tokens to process the same text, and turns on adaptive thinking by default. If the change increases exits, tool calls, retries, or human review, the task may become more expensive even during promotion.

The safest decision is to test now, but budget with the price that takes effect on September 1, 2026. The promotion serves to reduce the cost of the pilot, not to justify a permanent migration on its own.

The promotional price reduces all rates by a third

Sonnet 5 was announced on June 30, 2026. On the Anthropic API, the introductory price is $2 per million tokens in and $10 per million tokens out until August 31. From September 1st, prices will be US$3 and US$15, respectively.

The cache follows the same proportional change:

MTok means one million tokens. A five-minute cache write costs 1.25 times the normal entry price; recording for an hour costs twice; a reading costs 10% of the normal admission price.

On September 1st, therefore, each of these rates increases by 50%. This does not mean that the total cost of the task will also rise by 50%, because salaries, human review, infrastructure and external tools do not automatically follow the token price.

The new tokenizer changes the comparison with Sonnet 4.6

The default price for Sonnet 5 will be the same as Sonnet 4.6: $3 per million tokens in and $15 per million out. The difference is that Sonnet 5 uses a new tokenizer.

According to Anthropic documentation, the same text generates approximately 30% more tokens in Sonnet 5 than in Sonnet 4.6. The launch ad has a range of approximately 1 to 1.35 times depending on the type of content. The company itself recommends recounting the prompts in the new model rather than reusing previous measurements.

In a simplification in which input, output and cache maintain the same composition:

• during the promotion, a 30% increase in the number of tokens combined with a one-third reduction in price produces a cost approximately 86.7% of the previous one;
• after the promotion, this same increase in tokens can increase the cost of an equivalent request by around 30%;
• To counteract this increase after August 31st, Sonnet 5 would have to reduce the weighted volume of tokens, calls or rework per approved task by approximately 23%.

This account does not prove that Sonnet 5 will be cheaper or more expensive in its flow. It only shows the balance point. A quality improvement can eliminate attempts, summarize steps or reduce revision. The opposite effect is also possible.

Adaptive thinking and effort need to be fixed in the test

A direct change of the model identifier can change consumption. In Sonnet 4.6, requests without thought configuration could be executed without this feature. In Sonnet 5, adaptive thinking is enabled by default, and the output limit includes both the thought and the response text.

The effort parameter also influences text, thought, function arguments and number of tool calls. Sonnet 5 uses high effort by default in the API and Claude Code. For a valid comparison, record and record the effort level of the two flows. Changing it in the middle of a cached conversation may also invalidate the stored prefix.

There are still two operational changes that can produce failures or rework during migration: Sonnet 5 does not accept non-standard values for temperature, top_p and top_k, and the manual extended thinking mode with fixed budget has been removed. Incompatible requests return error and must be adjusted before measuring cost.

Per-task cost calculator

Use five groups of tokens:

New entry: instructions and content that did not come from the cache.

Output: all response tokens recorded by the API, including the effects of effort level and tooling steps.

Five minute cache write: prefixes stored at default duration.

One-hour cache write: prefixes stored with extended duration.

Read cache: tokens retrieved from an already stored prefix.

The API response separates these values into input_tokens, cache_creation_input_tokens, cache_read_input_tokens, and output_tokens fields. The total input processed is the sum of new input plus cache writes and reads.

The basic formula is:

API cost in dollars =
(input × input price + output × output price + write 5 min × write price 5 min + write 1 h × write price 1 h + read × read price) ÷ 1,000,000

Rework cost = cost of additional executions required to get a task approved.

Human review cost = review minutes ÷ 60 × full reviewer hourly cost.

Total cost of the task in reais = API cost in dollars × effective exchange rate + human review + external tools + fees and taxes not included in the API price.

The best method is to add up the tokens actually consumed by retries. When this is not yet instrumented, use temporarily:

API with rework = first run cost × (1 + rework rate)

This approximation is conservative when the second attempt repeats only a part of the stream.

How to amortize the cache

Consider a shared prefix of 40,000 tokens used by six tasks within the cache lifetime. The first task writes the 40 thousand tokens; the next five do the reading.

On average per task:

• amortized recording: 40,000 ÷ 6 = 6,666.67 tokens;
• amortized reading: 40,000 × 5 ÷ 6 = 33,333.33 tokens.

Do not count the 40 thousand tokens as normal input in every call. Also don't assume there will be a cache hit: the prefix must remain identical, the call must occur within the duration, and certain configuration changes may invalidate it.

Three reproducible scenarios

The scenarios below are not Sonnet 5 benchmarks. They are examples of filling out the calculator.

The sales PTAX of R$ 5.1154 per dollar, published by the Central Bank for August 5, 2026, was used. The actual conversion of the invoice may be different due to the settlement day, card, marketplace, taxes and contract.

Human review was calculated with an editable assumption of R$60 per hour.

Scenario 1 — Short batch screening

New input of 2,500 tokens, output of 600 tokens and shared prefix of 12 thousand tokens. 20 tasks are considered within the five-minute cache, 5% rework and one minute of review per task.

Scenario 2 — Report with documents

New entry of 18 thousand tokens, exit of 3,500 tokens and shared prefix of 40 thousand tokens. Six tasks are considered within the one-hour cache, 15% rework and five minutes of review.

Scenario 3 — Multi-step agent

New entry of 60 thousand tokens, exit of 12 thousand tokens and shared prefix of 80 thousand tokens. Four tasks are considered within the cache of one hour, 35% rework and 12 minutes of review.

API cost breakdown during promotion

The “cache” field brings together the amortized writes and reads. The rework was applied to the entire run just to keep the examples reproducible. In production, replace this percentage with the actual tokens from additional attempts.

Complete cost per task

In all three examples, the API becomes 50% more expensive in September, but the complete cost of the task increases by about 3.7%, 5.5%, and 8%, respectively. Human labor absorbs most of the budget.

Visual reading of the promotional cost

Short screening: API ██ 7% | human review ████████████████████ 93%

Report: API ██ 11% | human review ██████████████████ 89%

Multi-step agent: API ███ 16% | human review █████████████████ 84%

The graph makes visible the risk of comparison per million tokens: a small change in review time can overcome the entire promotion economy.

In the reporting scenario, for example, reducing the review from five to four minutes would save $1 per task. This is more than the full cost of the API even after the promotion, estimated at around R$0.92.

When Sonnet 5 is worth it

Migrate now tends to make sense when testing shows that the model maintains or improves first-time pass rate, reduces agent steps, or materially decreases review time. The calculation should remain favorable with September prices.

Waiting or limiting the migration to a pilot is more prudent when:

• the savings appear only with the promotional price;
• the flow does not yet record input, output and cache separately;
• the new tokenization is truncating responses or blowing up budgets;
• adaptive thinking or high effort increases calls without measured gain;
• compatibility errors are generating new attempts;
• review time grows, even with a drop in API spending.

The promotional price offers room for experimentation. It does not reduce the risk of being stuck with a flow that fails to close the account on September 1st.

How to measure for seven days

Record a line for each task, not just for each request:

1. task identifier and type;
2. model and level of effort;
3. uncached entry tokens;
4. tokens written to the five-minute cache;
5. tokens written to the one-hour cache;
6. tokens read from cache;
7. output tokens;
8. number of attempts;
9. approval on the first attempt;
10. minutes of human review;
11. costs of external tools;
12. exchange rate and rates used in conversion;
13. result approved, rejected or sent for correction.

The main metric is:

Cost per approved task = sum of API, rework, review and tools ÷ number of approved tasks.

Do not divide by the number of calls. A cheap flow that takes three calls to produce a valid delivery may cost more than one with a more expensive request.

Compare Sonnet 5 to the current model using the same tasks, instructions, tools, limits, and passing criteria. Project the results with the September price. The promotion must appear in a separate column.

For processes that support asynchronous execution, the Batch API offers a 50% discount on input and output. This flow must be analyzed separately, because it changes deadlines, architecture and operational experience; do not mix your results with interactive tasks. Tools running on the server may also incur additional charges in addition to tokens.

The decision rule

Migrate when the cost per approved task remains competitive with the standard price, quality reaches the defined threshold, and operational errors are resolved.

Keep a pilot when the advantage still depends on the promotional price or when the history is too short to measure rework and revision.

Do not migrate the entire flow when the account improves per request, but worsens per approved delivery.

Copy the calculator, replace the assumptions with actual usage fields, and record the cost for seven days. The token can become cheaper; the task only becomes cheaper when automation delivers it with less rework and less human intervention.

To relate this calculation to other model and method comparisons, see the AI in Test page.