Cheaper models | Ollama Cloud, Unleashed

How per-token pricing works

Each model has three prices per million tokens: fresh input, cached input and output. Each call is charged token by token at those rates. Some models list no cached price at all. At the last check that was mistral-large-3 and nemotron-3-nano.

Worked example on glm-5.3-flash, priced at $0.15 fresh input, $0.03 cached input and $0.50 output per million tokens. An agent call with 10,000 input tokens, 8,000 of them cached, and 1,000 output tokens:

2,000 fresh input   x $0.15 / 1,000,000 = $0.00030
8,000 cached input  x $0.03 / 1,000,000 = $0.00024
1,000 output        x $0.50 / 1,000,000 = $0.00050
                                   total = $0.00104

The model atlas uses this same illustrative call for every model, so the columns compare directly.

Verified 12 Oct 2026

Prices per million tokens, as listed. Source: ollama.com/pricing

Cached input against fresh input

Cached input is the part of a prompt that repeats from an earlier call. It is priced well below fresh input.
On glm-5.3-flash, fresh input costs five times cached input: $0.15 against $0.03 per million.
Where a cached price is listed, it is at or below the fresh input price. On nemotron-3-super and nemotron-3-ultra the two prices are equal, so caching saves nothing on those two.
Our agent-heavy account had about 93% of its input tokens cached over 30 days. Caching and tokens explains why.

Off-peak hours

Off-peak pricing applies outside 12:00 to 18:00 UTC on weekdays, and all day at weekends. Only models with an off-peak row get it. At the last check that was two models, both from DeepSeek.

Per million tokens: standard, then off-peak
Model	Fresh input	Cached input	Output
deepseek-v4.1-flash	$0.30, then $0.15	$0.006, then $0.003	$1.20, then $0.60
deepseek-v4-pro	$1.32, then $0.66	$0.044, then $0.022	$3.96, then $1.98

Off-peak is exactly half the standard price in every column. On weekdays the window covers 00:00 to 12:00 UTC and 18:00 to 24:00 UTC.

Verified 12 Oct 2026

Off-peak rows and window, as listed. Source: ollama.com/pricing

What our per-call tests showed

Our tests. Old-plan session cost, not dollars.

We sent identical prompts of about 400 to 500 tokens and measured how much of a five-hour session each call used. One session point is one percentage point of that session.

Session points per call, identical prompts
Round	glm-5.3-flash	DeepSeek model
Round 1, 27 Aug 2026, 20 calls per model	0.030	0.015 (deepseek-v4-flash:0731)
Round 2, 30 Sep 2026, 15 calls per model	0.093	0.140 (deepseek-v4.1-flash)

In round 1 the DeepSeek model used about half the session points of GLM. In round 2 DeepSeek v4.1 Flash used about 1.5 times as many.

The lesson

The cheapest model changed within one model generation. Re-test before you assume, and re-test when a model is updated.

These are small samples on one prompt shape. Your prompts, caching and output length will change the answer.

Where to compare

The model atlas lists every cloud model on the live list, with its prices, off-peak rows and sources. You can sort it by input price, output price or context. Open the model atlas.