Batch APIs: Is It Really Half Price?

“Use the batch API, it’s half price” is repeated often enough that most people have stopped checking. It’s a good rule of thumb. It is also wrong about one batch offering in five in our data.

On 2026-09-06 our snapshot contained 69 models with a :batch variant, every one of which also had a standard counterpart we could compare it against directly — same model, same day, same source. Here’s what that comparison shows.

The rule of thumb holds — for 55 of 69

Fifty-five of the 69 pairs are priced at exactly 50% of standard, on both input and output. Not approximately half. Exactly half, to the cent, across every price tier we track:

The consistency here is itself a finding. Across OpenAI, Anthropic, Google and Mistral, batch pricing is not negotiated per model or tuned per tier — it is a flat 50% policy applied mechanically to the whole catalogue. The median batch-to-standard output ratio across all 69 pairs is exactly 50.0%.

If your model is from one of those four labs, you can stop reading and use the rule of thumb. It’s correct.

The exceptions, in order

Fourteen pairs are not at 50%:

Three separate things are happening in that table, and they need separating.

1. Grok 4.3: a real batch discount that isn’t 50%

x.AI offers batch on exactly one model in our data, at 80% of standard — a 20% discount. This is the cleanest counterexample to the rule of thumb: a first-party, closed-weight model where batch is genuinely cheaper, just not half. If you sized a migration on “batch will halve this,” you’d be over-promising by 30 points.

2. Five models where batch exists but saves nothing

Qwen3.8 2.4T A95B, both Thinking Machines Inkling models, Kimi K3, and MiniMax M3 all list a batch variant at identical pricing to standard. (Inkling Small is slightly worse: its batch input is priced 11% above standard.) For three of them this is a recent development — see the history section below, where the discount is removed on a specific date.

Batch endpoints normally trade latency for money — you accept a delayed response in exchange for a lower rate. These five ask for the latency and return nothing. Unless a batch endpoint is buying you something else — higher rate limits, a queue that doesn’t count against concurrent request caps — there is no pricing reason to use them.

3. The eight that cost more — and an important caveat

The bottom eight rows show batch priced above standard, up to 3.5× on gpt-oss-120b. Taken at face value that’s absurd, so it’s worth being careful rather than loud about it.

Here is the pattern we can see: every single one of those eight is an open-weight model — gpt-oss, Gemma, Qwen, DeepSeek, GLM, Muse Glimmer. None of the closed first-party models (GPT, Claude, Gemini, Mistral, Grok) appears in that group.

Here is our hypothesis, which we cannot confirm from our own data: open-weight models are served by many different hosts, and the non-batch price we record may be the cheapest host on the market, while the :batch listing may come from a specific — likely first-party — provider. If so, these rows aren’t “batch costs more than standard at the same vendor”; they’re “the cheapest third-party host undercuts the first-party batch rate,” which is a different and less surprising claim.

We are flagging this as a hypothesis because our collector records prices, not which endpoint served them. We would rather publish the uncertainty than a clean story we can’t stand behind.

What is true regardless of the explanation is the practical instruction: for open-weight models, do not assume the batch endpoint is the cheap option. Compare the two listed prices before you route traffic. In our snapshot, for these eight models, the batch route is the more expensive one on its face.

Input and output aren’t always discounted equally

For 61 of 69 pairs the input ratio and output ratio are identical — halve one, halve the other. Eight pairs discount them differently, and in every case the divergence appears in the group above: Inkling Small (100% output / 111% input), Muse Glimmer (136% / 117%), gpt-oss-20b (154% / 167%), DeepSeek V4 Flash 0731 (156% / 215%), Qwen3.5-9B (167% / 170%), Gemma 4 31B (285% / 433%), gpt-oss-120b (353% / 405%).

The lesson is narrow but real: if you’re evaluating batch on anything outside the big-four closed catalogues, compute your own blended rate using your actual input:output ratio. A single “% off” number can hide a split that moves your bill in the wrong direction — and since output tokens already cost several times what input does, the output side is usually the one that decides your bill.

Who even offers batch

Anthropic exposes batch pricing on 44% of its catalogue (12 of 27) — proportionally the widest coverage of any lab we track. OpenAI has the most in absolute terms. Google covers about 28%.

And several major providers in our data show no batch variants at all: Meta (Llama), Amazon, Cohere, Perplexity, NVIDIA, and Microsoft. If your architecture assumes a batch tier will be available wherever you go, that assumption doesn’t survive a provider switch.

The 50% is a policy, not a promise — and it has broken

The table above is one day’s snapshot. Because we record every day, we can also ask a question a snapshot can’t: does the batch price stay at half?

Across the 82 batch/standard pairs where our history covers both, 55 sat at exactly 50% every single day we observed them. Nineteen were never at 50%. And eight pairs held 50% for a while and then stopped — the relationship broke, on a specific date, and we caught it.

2026-08-19: three batch discounts disappeared on the same day

Three unrelated providers raised their batch output price to match standard on the same date. In each case the standard price didn’t move at all — only the batch rate did. If you had checked any of these in July, you’d have recorded a 50% batch discount that no longer exists.

This also corrects the snapshot reading above: MiniMax M3 and the Inkling models don’t have a batch endpoint that “never had a discount.” They had a 50% discount and lost it on 2026-08-19.

NVIDIA Nemotron 3 Ultra kept going. Its standard price then fell to $2.20 on 08-28 and $2.40 on 09-03 while the batch rate sat at $3.60 — leaving batch at 164% and then 150% of standard. Its batch listing then dropped out of our snapshot entirely; the last reading we have is 2026-09-03.

Batch can also lag a standard increase — in your favour

Google Gemini 3.7 Flash shows the opposite:

The standard price doubled on 08-29. The batch price didn’t follow until 09-03. For five days, batch ran at a quarter of standard instead of half — an unadvertised discount that existed only in the gap between two repricings.

And sometimes the discount appears because standard went up

OpenAI’s GPT-5.6 Luna and Terra lines are the strangest case. When their batch listings first appear in our record on 2026-08-07, they are priced identically to standard — Luna at $0.60/$0.60, Terra at $6/$6, no discount at all. Then on 2026-08-18 the standard price doubled (Luna $0.60 → $1.20, Terra $6 → $12) while the batch price stayed put.

The 50% relationship that holds today wasn’t created by discounting batch. It was created by raising standard. A snapshot taken today shows a textbook 50% batch discount; the history shows how it got there.

Why this matters more than the headline number

If batch pricing were genuinely a formula — half of standard, maintained automatically — you could check it once and forget it. The record says otherwise: 8 of 82 pairs saw the relationship break, the breaks lasted between 5 and 19 days, and they went in both directions.

A discount that can silently vanish on a Tuesday is not something to hard-code into a cost model. If batch savings are load-bearing in your budget, they need to be checked on the same cadence as everything else.

The bottom line

• “Batch is half price” is true 80% of the time — 55 of 69 pairs are exactly 50% off, on both input and output.
• It is a flat policy, not a per-model deal, at OpenAI, Anthropic, Google and Mistral.
• x.AI’s batch is 20% off, not 50%. One data point, but a real one.
• Five models offer a batch endpoint with zero price benefit. You pay in latency and get nothing back.
• For open-weight models, check before you assume. Eight of them list batch above the standard price in our snapshot — most likely a first-party-vs-cheapest-host comparison rather than a vendor charging a batch premium, but either way the cheap route is not automatically the batch one.
• Verify input and output separately. Eight pairs discount them at different rates.
• The discount can be withdrawn. Three providers ended their 50% batch rate on 2026-08-19; one model ran at 25% of standard for five days because batch lagged a price rise.

Batch pricing is one of the few genuinely free discounts in this market, and for the big closed labs it is exactly as good as advertised — 55 pairs held at exactly half, every day we looked. It just isn’t universal, it isn’t guaranteed to stay, and the exceptions cluster precisely where people assume the savings are biggest.

Based on computetrail’s snapshot of 2026-09-06 (69 batch/standard pairs) and our daily history from 2026-06-29 to 2026-09-06 (82 pairs with overlapping records). Prices are as published to our data source on that date and exclude taxes, committed-use discounts, and rate-limit tiers; see our methodology. Where a model is served by multiple hosts, the recorded standard price may not come from the same provider as the batch listing — this is stated as a limitation, not a finding.