Anthropic Batch API cost estimation

See also / canonical cluster. Pillar for token & cost estimation: How to estimate LLM token costs before you ship. Interactive: token & cost estimator. This page stays on a different intent — see the intro.

Intent here: Batch API economics (discount assumptions, job latency, partial failure / resubmit) — not sync feature estimation. For the general pre-ship formula, use the estimation pillar; sanity-check tokens in the token estimator.

What Batch changes (conceptually)

Compared with synchronous Messages API traffic:

Factor	Sync	Batch
Latency	Seconds	Hours-scale completion window (check current docs)
Price	List rates	Often discounted vs sync — confirm on Anthropic’s pricing page the day you plan
UX fit	Chat, agents, live UI	Offline jobs only
Failure handling	Per-request retries	Job-level partial success / resubmit design

Do not hardcode a discount percentage in your finance model without a dated source link. Paste the current sync and batch rates into your sheet.

Estimation worksheet

Fix a job profile:

1. Model id (e.g. a Haiku- or Sonnet-class SKU you actually use)
2. Requests in the batch
3. p50 / p90 input tokens and average output tokens per request
4. Expected share of empty / truncated / policy-filtered outputs
5. Resubmit rate (bad inputs, schema fails, partial job failures)

Then:

token_$ ≈ requests × (in_tok × in_$/M + out_tok × out_$/M) / 1e6
all_in_$ ≈ token_$ × (1 + resubmit_rate) + eng_ops_amortized

Use token estimator for the token_$ line; keep rates in a dated “assumptions” tab.

Where people under-estimate

• Output variance — evals and long summaries blow p90; plan with p90 not p50 for budget caps
• Schema repair loops — “just one more retry” turns Batch into sync-priced chaos if you fall back incorrectly
• Duplicate jobs — idempotency keys / content hashes prevent paying twice for the same corpus
• Storage & egress — results landing in your bucket still cost you; not Anthropic’s token line
• Human QA — batch content pipelines still need review (AI content ops)

Sync vs Batch decision heuristic

Choose Batch when:

• Users are not waiting on the response
• You can tolerate multi-hour turnaround
• Volume is large enough that discount + queueing complexity pays for itself

Stay on sync when:

• Interactive agents / tool loops (tool-calling overhead)
• Tight SLO on first token
• Tiny volume where ops overhead exceeds discount

Operational checklist before first big job

• Dated pricing snapshot linked in the sheet
• Cap on max output tokens per request
• Dead-letter path for failed items
• Cost alert at 50% and 80% of monthly batch budget
• Clear ownership for “who may enqueue production corpora”
• PII / retention review for the dataset

Tie-in to unit economics

Batch is a cost lever, not a strategy. Fold discounted rates into LLM feature unit economics so product managers see contribution margin with the offline path, not only the chat path.

Disclaimer

Vendor pricing, SLAs, and Batch semantics change. Treat this page as a worksheet method, not a live price list. Always verify on Anthropic’s official pricing and Batch docs before committing budget.