# Off-peak pricing: half price for agents that run at night
> Since Monday, every chat model on sference bills 50% off for requests that arrive between 22:00 and 06:00 Berlin time. It works on the normal realtime endpoints, so agent runs, evals and backfills scheduled overnight get the discount without being rewritten as batch jobs.
- Author: Jernej Štrasner (Co-founder & CEO)
- Published: 2026-11-18
- Tags: pricing, agents, inference
- Source: https://sference.com/blog/off-peak-pricing

---

Since Monday, every chat model on sference bills 50% off for requests that arrive between 22:00 and 06:00 Berlin time (21:00 to 05:00 UTC in winter). There's nothing to turn on. You call the same endpoints and models at the same latency, and the price depends only on when the request arrives.

## Why batch doesn't fit agents

The usual way to get cheap tokens is a batch API: you write your requests to a JSONL file, submit it and come back within 24 hours. We have one too, at half the list price, and for a static backfill it's the right tool.

For an agent it doesn't work. Each call depends on the answer to the previous one, so you can't write the whole run into a file up front, and if every step can wait up to 24 hours a run that takes 20 minutes now could take days. So teams running agents in the background (evals, research agents, coding agents working through a backlog, nightly data jobs) end up paying the realtime price for work nobody is actively waiting on.

That's the work off-peak pricing is meant for. Schedule it for the night and every call it makes is half price, at the same latency you get during the day.

## Why we can offer it

Our GPUs run in Europe and most of our traffic follows European working hours. We wrote about what that does to speed in [Rush hour on the GPU](/blog/rush-hour-on-the-gpu): the fleet that is busy in the afternoon has spare capacity at night. A GPU costs us the same whether it's serving tokens or sitting idle, so selling those hours cheaper still beats leaving them empty, and the discount is roughly what that spare capacity is worth to us.

## What it costs

On GLM-5.3, $1.20 in / $4.20 out per million tokens becomes $0.60 / $2.10. On DeepSeek V4.1 Flash, $0.35 / $1.32 becomes $0.175 / $0.66. Cached input gets the same discount. The full table is on the [pricing page](/docs/pricing#off-peak-pricing).

Some details that matter if you plan around it:

- Arrival time decides the price. A request that arrives at 05:59 bills off-peak even if it finishes after 06:00. Each call of an agent is priced on its own, so a run that crosses 06:00 pays the normal rate for the calls after that.
- The window follows Berlin daylight saving, so in summer it's 20:00 to 04:00 UTC.
- Flex doesn't stack with it. Flex is 0.7× list during the day; at night the lower multiplier wins, so flex and realtime both cost 0.5×. Priority keeps its 1.3× premium on top of the discounted rate.
- Batch API jobs stay at 0.5× around the clock, and decisions models aren't discounted.
- `GET /v1/models` returns the window per model as `pricing.off_peak`, so a scheduler can read it instead of hardcoding hours. Off-peak requests are also flagged in the activity view of the console.

If you have work that can wait for the night, the simplest version is to kick off the eval suite or the agent backlog before you log off and read the results in the morning. The window is fixed to Berlin time, so if your team sits elsewhere check the UTC hours above.

This is a permanent part of our pricing, not a launch promo. I'm curious what people end up moving into the window, and if you have a workload that would move to the night at half price (or a reason it can't), I'd like to hear about it at hello@sference.com.
