Cost Watch · October 14, 2026
The model price is only the first line of the bill.
A lower rate is a reason to investigate. The useful number is what you pay for an answer you can actually use.
What changed, and who should care.
Anthropic lists Sonnet 5.5 at $2 input / $10 output per million tokens, versus $3 / $15 for Sonnet 4.6. Equal billed token counts would cost one-third less. Provider pricing. Its API behavior also changes, so test your integration before switching. Migration details.
Haiku 5.5 has a lower starting rate, but prompts above 100,000 tokens use a higher pricing tier. A large context window does not mean a flat price. Provider conditions.
Scruuge's decision: test. Support teams should check extraction failures; builders should count cache hits; operations teams can examine batch pricing without changing a working model.
Three worked costs, not three model tests.
Illustrative token budgets; models were not run. Counts are assumed billed tokens including any billed reasoning. Excludes retries, tools, taxes and human review. Direct API prices are not subscription allowances or guarantees about OpenRouter routes. No claim of equal quality or realized savings.
All rates below are USD per million tokens on the Anthropic direct API, checked October 14. These examples do not replace the calculator's separately sourced OpenRouter baseline prices.
TEST · illustrative calculation
Support extraction: test the cheaper candidate.
10,000 tickets; 2,000 uncached input and 200 billed output tokens per ticket. Each prompt stays below 100,000 tokens.
Exact model IDs: claude-sonnet-4-6 / claude-haiku-5-5
Before: $90
- Input: 20M × $3
- Output: 2M × $15
Alternative: $3
- Input: 20M × $0.100
- Output: 2M × $0.500
The $87 difference assumes identical billed token counts. Check omissions, required fields and escalation behavior before routing real tickets. Above the prompt threshold, Haiku's higher rate applies; cached input also contributes to the threshold.
Suggested prompt template (not executed)
Extract order_id, issue_type and requested_action from {{ticket}} as JSON. Use null for missing fields; flag cases requiring a human.
Token counts above are assumptions, not measurements of this template.
Evidence: Provider pricing · Model specifications
TEST · illustrative calculation
Repeated coding context: count the cache hits.
100 calls share a 10,000-token prefix; each adds 1,000 uncached input and 500 billed output tokens. One five-minute cache write, then 99 hits before expiration.
Exact model IDs: claude-sonnet-5-5
Before: $2.70
- All input: 1.1M × $2
- Output: 0.05M × $10
Alternative: $0.824
- Cache write: 0.01M × $2.50
- Cache reads: 0.99M × $0.100
- Uncached input: 0.1M × $2
- Output: 0.05M × $10
Input categories are disjoint. The prefix must match, meet eligibility rules and stay within cache lifetime. Confirm hits in usage records. Misses, changed context and extra reasoning/output increase the bill.
Suggested prompt template (not executed)
Use {{repository_context}} to propose a minimal patch for {{task}} and explain how to verify it. Do not claim tests were run.
Token counts above are assumptions, not measurements of this template.
Evidence: Model specifications · Cache rules
STAY / TEST BATCH · illustrative calculation
Draft summaries: keep the model, change the billing mode.
1,000 draft summaries; 3,000 uncached input and 1,000 billed output tokens each. Same older model; compare standard requests with asynchronous Batch pricing.
Exact model IDs: claude-sonnet-4-6
Before: $24
- Input: 3M × $3
- Output: 1M × $15
Alternative: $12
- Batch input: 3M × $1.50
- Batch output: 1M × $7.50
The $12 difference is billing arithmetic, not an older-model quality win. Use batching only when delayed completion fits the job; check results and failures. Retaining a working older model can avoid an unnecessary migration.
Suggested prompt template (not executed)
Summarize {{operations_notes}} for an internal weekly update. Preserve decisions, owners and deadlines; flag missing information.
Token counts above are assumptions, not measurements of this template.
Evidence: Provider pricing
Keep the older model when the evidence says to.
The batch example changes the billing mode while keeping Sonnet 4.6. It does not prove that an older model outperforms a newer one. That would require the same workload, saved outputs and a reviewed rubric. We have not run that comparison for this edition.
Before moving a workload, compare accepted results, retries, review time and the bill. Keep counterexamples. A cheaper token can still produce a more expensive task.
Evidence edition: 2026-10-14.cost-watch-v1. Paid evaluations: off. Historical model outputs retain their original dates and labels. Recommendation changes remain pending workload evidence.