Batch Saves 50%? Not Anymore: What Waiting Really Saves You on AI APIs

OpenAI is dropping Flex and bringing in deadline-based pricing. Google doubles Gemini 4 prices on December 1, and DeepSeek charges double during the German morning. We priced three typical workloads across four providers and every pricing tier, and went through the fine print. The upshot: time of day, workflow and cache decide what you pay, not the discount percentage.

Analysis Nov 20, 2026 10 min read Original analysis by our editors
Illustrative image, AI-generated

What this is about

For a long time, AI APIs came with a simple rule: if you don't need an answer right away, send your requests as a batch, a bulk job processed within 24 hours, and pay half. This week, that rule started to wobble. With GPT-6.2 Astra and Sol on November 17/18, OpenAI dropped the Flex tier for new models. Flex was a mode with a 50 percent discount where requests got answered normally, just more slowly. It's being replaced by completion windows: 30 percent off if the result only needs to be ready within 2 hours, 40 percent off for 8 hours. Google doubles prices for Gemini 4 Carbon and Argon on December 1, and that includes Batch and Flex. DeepSeek charges double during weekday peak hours, and those hours land squarely on the German morning.

Our thesis: the blanket claim that “batch saves 50 percent” no longer describes what you actually pay. To test it, we priced out three typical workloads and combed through the providers' documentation for traps.

Pricing at a glance (as of November 20)

For comparison, we picked upper-midrange models. GPT-6.2 Sol, Claude Sonnet 5.5 and Gemini 4 Argon have identical list prices until the end of November, so any differences come down to the tier alone. All prices are in dollars per million tokens (input / output / cache hit).

Provider & tier Input Output Cache read Wait time
GPT-6.2 Sol Standard 2.00 10.00 0.10 Seconds
GPT-6.2 Sol 2h window 1.40 7.00 0.07 Target ≤ 2 h
GPT-6.2 Sol 8h window 1.20 6.00 0.06 Target ≤ 8 h
GPT-6.2 Sol Batch 1.00 5.00 0.05 ≤ 24 h
Claude Sonnet 5.5 Standard 2.00 10.00 0.10 Seconds
Claude Sonnet 5.5 Batch 1.00 5.00 0.05 ≤ 24 h, usually < 1 h
Gemini 4 Argon Standard 2.00 → 4.00 10.00 → 20.00 0.10 → 0.20 Seconds
Gemini 4 Argon Batch / Flex 1.00 → 2.00 5.00 → 10.00 0.05 → 0.10 24 h / 1–15 min
DeepSeek V4.1 Flash off-peak 0.15 0.60 0.003 Seconds
DeepSeek V4.1 Flash peak 0.30 1.20 0.006 Seconds

The arrow for Gemini marks prices from December 1. OpenAI's window prices aren't listed as separate amounts on its pricing page; they're derived from the percentages. Be careful with copied tables: router provider OrcaRouter lists $1.60 and $1.40 for Sol. That doesn't match discounts of 30 and 40 percent.

For the flagship GPT-6.2 Astra ($8/$40), the 8-hour window comes to $4.80/$24 and Batch to $4/$20. According to Google's pricing page, Gemini Neon ($3/$15) and Hydrogen aren't getting more expensive.

The clock: what Central European Time means

According to DeepSeek's official pricing page, peak rates apply Monday to Friday from 01:00–04:00 and 06:00–10:00 UTC. Chinese public holidays are excluded, and weekends are entirely off-peak. In German winter time (CET), that's 2–5 a.m. and 7–11 a.m.; once daylight saving time kicks in, it's 3–6 a.m. and 8 a.m.–12 noon. When the scheme launched in August, InfoWorld wrote that Western customers would “mostly pay off-peak prices.” For teams in Germany, that's not true: their morning falls right into peak time. The Eesel blog calculates that a job at 08:00 UTC costs twice as much as the same job at 14:00 UTC.

A correction to our own starting point: peak pricing didn't arrive with V4.1. DeepSeek introduced it on August 16 alongside a price increase for V4 and kept it when V4.1 Flash launched on September 10. The announced redirect from V4-Pro to Flash is also contradictorily documented. DeepSeek's September announcement says it's coming. Codersera, on the other hand, cites a changelog saying it was rolled back. The current pricing page lists V4-Pro with its own price ($0.66/$1.98 off-peak).

With OpenAI, the window starts as soon as the API accepts the request. According to the documentation, though, the window is a target, not a guarantee. What happens if it's missed isn't documented anywhere.

Scenario 1: Processing 1,000 PDFs

Assumption: each PDF uses 3,000 input tokens and 500 output tokens, for a total of 3M input and 0.5M output tokens. In the “with cache” variant, 1,200 of those tokens are a fixed instruction block served from the cache after the first call (best case).

Tier without cache with cache
Sol Standard $11.00 $8.72
Sol 2h $7.70 $6.11
Sol 8h $6.60 $5.23
Sol Batch / Sonnet Batch $5.50 $4.36
Argon Batch/Flex until Nov 30 $5.50 $4.36
Argon Batch/Flex from Dec 1 $11.00 $8.72
Argon Standard from Dec 1 $22.00 $17.44
DeepSeek off-peak / peak $0.75 / $1.50 $0.57 / $1.15

Two things stand out. Caching doesn't buy you much here, because every PDF is different. And from December, Argon Batch costs as much as Argon Standard does today. The 50 percent discount stays, but the base price has doubled. For Sonnet, the value of caching depends on how often the cache actually hits in batch mode. A market comparison by Beri.net cites observed hit rates of 30 to 98 percent; that hasn't been officially confirmed. At 30 percent, costs would come to about $5.16 instead of $4.36.

Scenario 2: A 12-step overnight agent run

Assumption: the agent starts with 8,000 tokens (instructions and tool descriptions). Each step adds 3,000 tokens of tool output and 1,000 output tokens. Over 12 steps, that's 360,000 input and 12,000 output tokens. With caching, 308,000 tokens come from the cache and 52,000 are newly written.

Tier without cache with cache Worst-case duration
Sol Standard (30-min cache) $0.84 $0.28 Minutes
Sol 2h (12-h cache required) $0.59 $0.29–0.37 24 hours
Sol 8h (12-h cache required) $0.50 $0.25–0.35 96 hours
Sonnet Standard (5-min cache) $0.84 $0.28 Minutes
Sonnet, each step as batch (1-h cache) $0.42 $0.18 12 × 24 hours
Argon Standard until Nov 30 / from Dec 1 $0.84 / $1.68 $0.25 / $0.51 Minutes
Argon Flex from Dec 1 $0.84 $0.25 approx. 3 hours
DeepSeek off-peak / peak $0.06 / $0.12 $0.02 / $0.03 Minutes

This is where the math flips. According to OpenAI, a window only applies to the first response; every follow-up step restarts the clock. If hours pass between two steps, the cheap 30-minute cache has expired. You then need the 12-hour cache, whose write price is $5 instead of $2.50 per million tokens. OpenAI doesn't say whether the window discount also applies to cache writes. The Apiyi blog says they're billed separately. That's where the range in the table comes from. Our take: for agents, the Standard tier with caching is almost as cheap as the 8-hour window, often cheaper, and it's days faster. On Gemini, Flex remains the only real middle ground: half price, but according to Google only 1 to 15 minutes of wait time per request.

Scenario 3: The daily newsletter summarizer

Assumption: 20 newsletters a day, each with 4,000 input tokens and a 300-token summary. Per month, that's 2.4M input and 0.18M output tokens.

Tier per month
Sol / Sonnet / Argon Standard (Argon until Nov 30) $6.60
Sol 2h / 8h $4.62 / $3.96
Batch (Sol, Sonnet, Argon until Nov 30) $3.30
Argon Standard / Batch from Dec 1 $13.20 / $6.60
DeepSeek off-peak / peak $0.47 / $0.94

The difference between Sol Standard and the 8-hour window is about $32 a year. The schedule matters more: if you want the summary ready by 7 a.m., an 8-hour job has to go out by 11 p.m. Anything that arrives overnight gets left out. DeepSeek has a handy gap: from 5 to 7 a.m. CET (04:00–06:00 UTC), off-peak pricing applies.

The fine print: what holds up and what doesn't

  • Separate caches – true. According to OpenAI's docs, regular requests can't use cache entries from 2h or 8h jobs, even with an identical prompt. Eesel spells out the consequence: you can't fill the cache cheaply at night and read from it cheaply during the day. Another pitfall: with GPT-6.2, instructions in the instructions field aren't cached. In Eesel's test, that pushed the cost of 1,000 support tickets from $5.45 to $29.20.
  • Windows add up – true. 12 steps at 8 hours each means up to 96 hours. The Batch API isn't suited to multi-step agents anyway, since it only processes independent, standalone requests.
  • Storage despite Zero Data Retention – true, but not how you'd think. Zero Data Retention (ZDR) is a commitment that no content gets stored. Background mode still stores data temporarily and encrypted, for about ten minutes in the standard window. For 2h and 8h, OpenAI doesn't specify a fixed duration. ZDR projects are explicitly allowed to use the windows, though. The Batch API, by contrast, isn't ZDR-eligible at OpenAI at all, nor at Anthropic (29-day retention). We weren't able to get a specialist law firm to assess before publication whether the temporary storage is permissible under a GDPR data processing agreement. The question remains open. Also unclear: OpenAI offers EU data residency for Batch on GPT-6.2, but it isn't documented for the completion windows. As for DeepSeek, Eesel writes that there's neither a zero-retention option nor a data processing agreement, and processing takes place in China.
  • Expired batches – only half true. According to OpenAI's docs, completed requests are billed even in an expired batch. Whether unfinished requests are free isn't stated. Anthropic explicitly says expired requests aren't billed. A case on the OpenAI forum shows how this can go wrong: $212.40 for batches without a single result. The money was only refunded six weeks later.
  • Gemini after 48 hours – true. According to Google, an expired job returns no results. On the developer forum, users report jobs that sat in the queue for more than 24 hours and then flipped to “EXPIRED.” Whether you get charged for those isn't documented.

Our verdict

The 50 percent still exists, but only with Batch and only for independent, standalone tasks like the PDF processing. There, Sol Batch and Sonnet Batch are neck and neck. Argon falls behind from December, and Anthropic's rules on expired requests are more customer-friendly.

For agents, OpenAI's new completion windows are rarely worth it. The Standard tier with a well-built cache lands in the same price range and is days faster. If you're coming from the Flex tier, Eesel says you'll pay 20 to 100 percent more.

DeepSeek remains an order of magnitude cheaper. German teams, though, shouldn't use it between 7 and 11 a.m., and the data protection question needs to be settled before you deploy it.

For small tasks like the newsletter, we're talking cents. That rarely justifies the overhead of windows, polling and retries. Our take: when it comes to cost, three questions now matter more than the discount percentage: How many steps does the task have, how long does the cache live, and what time is it in Beijing?

Sources

#api-pricing#openai#gemini#deepseek#anthropic#gdpr

Ask the AI assistant

Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.

Connecting …

The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.

Comments

What do you think? Already tried it? Write it here, no sign-up needed.

  1. Loading comments …

Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.