Batch API: Reduce LLM Costs With OpenAI, Claude and Gemini

Last verified: August 15, 2026

Short path: AI API pricing · Cheap LLM API · Models · OpenAI-compatible API

LumeAPI is an independent third-party gateway — not OpenAI, Anthropic, or Google. This page owns provider-native Batch versus LumeAPI realtime. LumeAPI has no Batch endpoint, no JSONL custom_id join, and no 24-hour completion SLA. If you need those semantics, stay on the vendor. If you only needed the token discount, compare vendor Batch rates to LumeAPI realtime catalog rows before you build a file pipeline.

GPT mini/Terra/Sol routing without Batch lives on the OpenAI bill guide. Cross-provider standard rates live on the 2026 pricing comparison.

Quick Answer

1. Confirm every record is independent and recoverable by ID.
2. Price vendor Batch against LumeAPI realtime on the same model family.
3. Build reconciliation before you submit the first production JSONL.

In short

Batch is an SLA and a join key. The ~50% vendor discount is real on eligible tokens (OpenAI Batch). LumeAPI does not implement that product. For GPT-5.6 Terra, LumeAPI realtime ($0.75/$4.50) is below OpenAI Batch ($1/$6). For Fable and several Gemini Flash rows, vendor Batch token rates match LumeAPI realtime — so Batch is only worth it if you need async files, not a lower $/1M.

What most guides get wrong

They tell you to “use Batch for 50% off” as if every host offered it. Gateways that speak Chat Completions are not Message Batches. They also add the 50% to a gateway discount and call it 75% off. You cannot stack OpenAI Batch with a LumeAPI Terra row: those are two different products.

They ignore expired, invalid, and quality-rejected records. The metric is cost per accepted result inside the documented window, not cost per uploaded line.

OpenAI Batch vs LumeAPI realtime (the table this page owns)

OpenAI documents 50% versus synchronous token prices and a 24-hour turnaround. After the 30 July 2026 Terra cut, standard Terra is $2 / $12, so Batch Terra is $1 / $6. Sol standard $5 / $30 → Batch $2.50 / $15. Sources: OpenAI pricing, Batch guide. LumeAPI billed rates from catalog.py (2026-07-22).

Same 10 million input + 2 million output tokens:

On this Terra shape, LumeAPI realtime undercuts OpenAI Batch by $5.50 ($22 − $16.50) if you do not need JSONL, custom_id order independence, or the Batch rate-limit pool. If you do need those, OpenAI Batch at $22 still beats OpenAI realtime at $44, and LumeAPI is the wrong product.

Luna Batch would be $0.10 / $0.60. Luna is not in the LumeAPI catalog — keep Luna Batch on OpenAI.

Claude and Gemini: Batch often equals LumeAPI realtime

Anthropic Message Batches are 50% of standard token prices. Most batches finish within an hour; incomplete batches expire after 24 hours (batch processing, pricing).

Gemini Batch is generally 50% of paid-tier token rates for supported models (Gemini pricing). Verify the current model row — Batch support varies by endpoint.

Read this table as: for Fable, Sonnet 4.6, and these Flash ids, switching to LumeAPI realtime does not beat vendor Batch on token rate. You choose Batch for the file/SLA workflow, or LumeAPI for one key + Chat Completions latency. GPT Terra/Sol is the family where LumeAPI realtime can beat vendor Batch on $/1M.

Batch fit matrix (architecture, not price)

If you cannot answer yes to the first four, the discount is irrelevant.

The all-in Batch formula

all-in batch cost = B + (f × R) + O + U
realized savings  = S − all-in batch cost

• S = standard realtime token cost for attempted records
• B = Batch token cost for those attempts
• f = fraction reprocessed
• R = reprocessing cost on the same full-job basis
• O = orchestration, storage, engineering
• U = urgent records diverted to realtime

A realistic production scenario

A nightly classify job would cost $1,000 on OpenAI Terra standard. Token-equivalent OpenAI Batch is $500. Five percent is reprocessed at standard, operations add $40, urgent fallbacks add $25:

all-in OpenAI Batch = $500 + (0.05 × $1,000) + $40 + $25 = $615
realized vs standard = $1,000 − $615 = $385 (38.5%, not 50%)

Same token counts on LumeAPI Terra realtime, no Batch file:

LumeAPI token cost = $1,000 × ($16.50 / $44.00) ≈ $375

If Chat Completions quality matches and you do not need custom_id reconciliation, LumeAPI realtime can undercut even the all-in OpenAI Batch number. If 8% of lines fail schema and you lack a join key, Batch’s operational cost explodes and LumeAPI still will not invent JSONL for you.

Figures are illustrative, not a customer invoice.

Copyable worksheet

Use cost per accepted result, not cost per submitted line.

Production reconciliation (vendor Batch)

OpenAI: each JSONL line needs a unique custom_id; do not assume output order matches input (Batch API). After completion:

1. Split succeeded, invalid, cancelled, expired, errored.
2. Validate schema, not only HTTP success.
3. Retry safe failures with an idempotency key.
4. Compute accepted-result cost and wall time against the 24-hour cap.
5. Keep lineage to reproduce an output.

That join layer is the engineering difference between “upload a file” and a dependable Batch pipeline. LumeAPI usage logs are per Chat Completions call — they are not a Batch result file.

A simple migration plan

Phase 1 — realtime baseline

Record tokens, success rate, latency, retries, cost per accepted result on the current host.

Phase 2 — shadow a small batch or a LumeAPI slice

If you need JSONL: duplicate a non-sensitive sample into vendor Batch. If you only needed a lower Terra rate: shadow Chat Completions on LumeAPI gpt-5.6-terra. Do not run both migrations in the same week unless you can attribute failures.

Phase 3 — bounded percentage

Route 10–20% of eligible records. Hard daily budget. Automatic realtime fallback for urgent items.

Phase 4 — scale when joins are boring

Increase volume when retries are idempotent and accepted-result cost stays below baseline.

Where LumeAPI fits

Use LumeAPI when the job is realtime Chat Completions (or documented image/video) and you want one key across catalog ids. Use vendor Batch when you need the async file contract. Do not assume /v1/batches exists at https://api.lumeapi.site/v1.

Expert take

Ask two questions in order: Do I need Batch semantics? If yes, pick OpenAI, Anthropic, or Gemini Batch and budget retries inside the window. If no, am I on GPT Terra/Sol? Then LumeAPI realtime can beat OpenAI Batch on listed token rates. On Fable and the Flash rows above, the token rates already match — Batch is a workflow choice, not a price cut versus LumeAPI.

Frequently asked questions

No. The published discount applies to eligible tokens. Retries, failures, storage, orchestration, and urgent realtime work reduce the realized saving. OpenAI Terra Batch vs LumeAPI Terra realtime can go the other way: $22 vs $16.50 on 10M+2M.

OpenAI documents a 24-hour turnaround. Anthropic says most finish within one hour; incomplete batches expire after 24 hours. Plan to the documented maximum.

Only for independent offline parts (evals, summarization, enrichment). Interactive tool loops stay realtime because later steps depend on earlier outputs.

No. There is no documented Batch product on the gateway. Compare vendor Batch to LumeAPI realtime, then pick one.

Only if the same vendor documents both on that model. Do not add a cache discount to a LumeAPI catalog row and a Batch discount as if they stacked.

Cost per accepted result, plus completion time and failure rate. Token price alone hides rework.

Sources

• OpenAI Batch API
• OpenAI pricing
• Anthropic Message Batches
• Anthropic pricing
• Gemini API pricing
• LumeAPI pricing
• LumeAPI models

Prices checked 15 August 2026. LumeAPI billed rates from catalog.py (2026-07-22). Worksheet numbers are planning examples, not customer savings.