Discounted LLM APIs: 13 Models Below List Price

Thirteen models across five series are billing below their struck-through price on ofox right now. The best-value page lists them sorted by discount size, and it is worth two minutes because the interesting part is not the percentage.

The interesting part is what each percentage is measured from. Three of these cuts mean three different things, and only one of them is a gateway putting its own margin on the table.

Which Models Are Discounted?

Five series: GPT, Gemini, GLM, Doubao and Seedance. Text models bill per million tokens, Seedance bills per second of output.

The four Seedance models bill on a different meter: seedance-2.0-mini from $0.02 per second of output, seedance-2.0-fast at $0.042, seedance-2.0 at $0.063, and seedance-2.5 from $0.11 at 480p, $0.48 at 1080p text-to-video and $0.568 at 1080p video-to-video. That last model prices per resolution and mode rather than as a single number, so the discount page shows only its 1080p figure while the model page carries the full grid. Duration and resolution drive that bill, not prompt length, which is why a 30-second clip costs roughly five times a 6-second one at the same resolution.

Is a Discounted Gateway Rate Actually Cheaper Than the Vendor?

Sometimes, and the honest answer splits three ways. We checked every text rate above against the vendor’s own pricing page on 2026-08-21. They fall into three groups, and knowing which group a model is in is the difference between a real saving and a number that looks like one.

Group one: below the vendor’s list. GLM-5.2 bills $0.98 / $3.08 here against Z.ai’s own $1.40 / $4.40, which is 30% off the rate you would pay calling Z.ai directly. GLM-5.3 is $1.26 / $3.96 against the same $1.40 / $4.40 list, so 10% off. The two Doubao models belong here too: Seed 2.1 Pro bills $0.7072 / $3.536 and Turbo $0.3536 / $1.7696, against the ¥6 / ¥30 per million ByteDance published at launch, which converts to roughly $0.88 / $4.41 at the rate we recorded in our Seed 2.1 access guide. Volcano Engine prices in CNY, so treat that conversion as approximate rather than a cent-for-cent comparison. GPT-5.6 Sol is the largest of these: $2.50 / $15 against OpenAI’s standard-tier $5 / $30. That is half, on the tier you actually call synchronously.

Group two: matches the vendor exactly. GPT-5.6 Luna bills $0.20 / $1.20, which is OpenAI’s standard rate for that model to the cent. GPT-5.6 Terra bills $2 / $12, likewise identical. You save nothing per token against calling OpenAI directly, and that is the point: no markup, one key, and the fallback routing comes free rather than costing you a premium.

Group three: the vendor’s own promotion, passed through. Gemini 3.7 Flash and 3.6 Flash both bill $0.75 / $3.75. So does Google. That rate is Google’s introductory price, printed on its pricing page with an expiry of 2026-12-31, after which both models return to $1.50 / $7.50 on 2027-01-01. Nobody is discounting Gemini for you. The gateway is passing Google’s own limited-time number through unchanged, which is the correct behaviour, but it means the saving disappears on a date Google controls rather than one a gateway controls.

If you are budgeting past new year, that third group is the one to model twice. We wrote up what the Gemini 3.7 Flash rate is measured against in detail, including the archived pricing pages that show the comparison baseline moving inside the same 48 hours as the launch.

What Does GPT-5.6 Cost Here Against OpenAI?

One of the three tiers is genuinely half price; the other two match the source.

Worth knowing about the Sol row: $2.50 / $15 is also what OpenAI charges for Sol on its Batch and Flex tiers. Those tiers trade latency for price, with Batch returning results asynchronously within a window rather than on the call. Here you get that number on a normal synchronous request, which is the whole substance of that particular discount.

Cached input follows the same shape: $0.25 per million on Sol, $0.20 on Terra, $0.02 on Luna. On a long-running agent that replays the same system prompt on every turn, cached input is usually the line that decides the monthly bill, not the headline input rate.

Which Cuts Follow the Vendor’s Promotion?

Gemini and Seedance, both of them. Google’s Flash rate carries a printed expiry. The Seedance discounts track BytePlus promotional windows, which carry printed start and end dates rather than running open-ended, and the 2.5 window covers 1080p output only.

This matters for one specific decision: whether to hardcode a price into a cost model. A rate in group one is a commercial arrangement that tends to move slowly. A rate in group three moves when the vendor’s marketing calendar moves. Both are real money today. Only one of them is a number you can put in a spreadsheet cell labelled “next quarter”.

The best-value page states this in its own words, and it is the sentence on that page worth reading twice: promotions follow vendor changes and may end at any time.

How Do I Use the Discounted Rate?

You do not do anything. There is no coupon field, no plan to switch and no minimum top-up. The rate on the model page is what the request is billed at.

from openai import OpenAI

client = OpenAI(api_key="sk-...", base_url="https://api.ofox.io/v1")

r = client.chat.completions.create(
    model="z-ai/glm-5.2",              # $0.98 in / $3.08 out per 1M
    messages=[{"role": "user", "content": "Summarise this changelog"}],
)
print(r.usage)

Swapping models is a one-string change, which is the practical reason to care about a discount list at all: if GLM-5.2 at $0.98 clears your quality bar for a classification job currently running on a frontier model at $5, the migration is one line and the saving is 5x. The model finder ranks the catalogue by task type and budget if you want the comparison done for you, and our API pricing comparison covers the wider field including the models that are not on discount.

Two more worth reading if a specific series is your candidate: GLM 5.3’s pricing and its reasoning_effort trap, where the default effort setting bills 35x the cheapest one, and the Doubao Seed 2.1 access guide if you want the Volcengine models without a Volcengine signup.

Which Model Should You Pick From the List?

The rates above are what the console bills today. The percentages next to them on the page are context for how that number got there, and as this post argues, the context is worth more than the percentage.

References

• ofox: discounted models
• OpenAI: API pricing
• Google: Gemini API pricing
• ofox model page: GPT-5.6 Luna
• ofox model page: GPT-5.6 Sol
• ofox model page: GLM-5.2
• ofox model page: Doubao Seed 2.1 Pro
• Volcano Engine: ModelArk pricing
• ofox model finder