BroadRouter

How often do LLM API prices change? 3 years of archived pricing pages

By the BroadNet Inference team · 7 min read · Published · Updated

Rarely, for most models. In 308 archived pricing pages from 11 providers (January 2023 to November 2026), 398 of 498 models seen in more than one month (80%) never changed price. We found 161 list-price changes: 100 cuts, 52 rises and 9 mixed. DeepInfra made 121 of them. OpenAI, Anthropic and Google repriced 8 of 126 models between them.

If you budget for an LLM API, you need to know two things: how often the price of a model changes, and in which direction. We answered that from primary evidence. BroadRouter read 308 Internet Archive captures of 11 providers’ own pricing pages, from January 2023 to November 2026, and compared every model’s price month by month. BroadRouter is BroadNet’s website. BroadNet is not in the dataset, because the archive has no captures of its pricing page.

The full data, with an archive link and a verbatim quote for each of the 3,332 prices, is our LLM API price history dataset. This post sums it up. For the biggest single moves, see LLM API price cuts 2023 to 2026 and LLM API price increases.

How often does an LLM API price change?

Almost never for a given model. We followed 498 models that appear in more than one calendar month of captures. 398 of them (80%) showed the same input and output price in every capture. Only 99 models had at least one list-price change; the 100th moved only into or out of a promotional price, which we don’t count.

In total we found 161 list-price changes on the same model:

Kind of change Count Median size (blended)
Cut: no price went up, at least one went down 100 −19%
Rise: no price went down, at least one went up 52 +40%
Mixed: one price up, one down 9 —

“Blended” means 3 input tokens per output token, as on the rest of BroadRouter. The median cut was 20% on input and 17% on output. 12 of the 100 cuts halved the blended price or more. 2 more moves were into or out of a promotional price; we show them in the dataset but do not count them.

Which providers change prices most often?

One host dominates. DeepInfra made 121 of the 161 changes (75%) and repriced 63 of its 126 long-listed models (50%). Everyone else changed much less often.

Provider Monthly captures used Models seen in 2+ months Models repriced Cuts Rises Mixed
DeepInfra 31 126 63 (50%) 75 40 6
Together AI 19 125 9 (7%) 3 7 0
OpenAI 48 77 5 (6.5%) 5 0 0
Mistral AI 25 38 5 (13%) 3 2 0
Groq 30 34 6 (18%) 6 0 0
Fireworks AI 23 28 3 (11%) 1 1 2
Anthropic 39 26 1 (4%) 1 1 0
Google (Gemini API) 32 23 2 (9%) 3 0 0
xAI 12 9 0 0 0 0
DeepSeek 29 6 3 (50%) 1 1 1
Cohere 20 6 2 (33%) 2 0 0

Two patterns stand out:

  • Model makers rarely touch a price once it is published. The seven makers (OpenAI, Anthropic, Google, DeepSeek, Mistral, xAI, Cohere) repriced 18 of 185 long-listed models, about 10%.
  • Hosts are split. All inference providers together repriced 81 of 313 models (26%), but without DeepInfra the figure falls to 18 of 187 (10%), the same as the makers.

DeepSeek and Cohere show 50% and 33% only because they list few models: 3 of 6 and 2 of 6.

Why do flagship prices fall if makers rarely reprice?

Because new versions arrive at new prices. OpenAI, Anthropic and Google made only 10 list-price changes across 126 long-listed models. Yet the first archived price of GPT-4 (8K context) was USD 30.00 input and USD 60.00 output per 1M tokens in April 2023, and GPT-6.1 Sol was first captured at USD 2.00 and USD 10.00 in October 2026. Almost all of that drop came from launching cheaper models, while the old ones kept their old prices on the page.

That matters for planning. If you stay on a model, do not expect its price to fall. To capture lower prices you usually have to migrate, test and switch. OpenAI vs Anthropic vs Google API price history traces each line model by model.

Are prices still falling?

Cuts led in 2024, but 2026 has more rises than cuts so far. Each change is dated by the later of its two captures:

Year of the later capture Cuts Rises Mixed
2024 45 2 0
2025 27 19 8
2026 (to 2 November) 28 31 1

Read the 2026 row with care. 23 of its 31 rises were at DeepInfra and 7 at Together AI, so it reflects two hosts, not the whole market. And two of the 2026 cuts, OpenAI’s gpt-3.5-turbo and o1-mini, compare a January 2026 capture with one from about a year or more earlier, so the cut itself may be older. Do LLM API prices only go down? goes through the rises one by one.

How precise are these dates?

A change is only known to fall between two captures. The median gap between the two captures of a change was 32 days. For 6 of the 161 changes the gap was over 60 days, because a page was not captured or was skipped in between. So “the price changed in March” means “a capture in early March and one in early April show different prices”, not that the provider announced it then. We use capture dates throughout and make no claim about announcement dates.

What does this mean for your budget?

  • Plan on today’s price for the model you use. 80% of models never moved. A forecast that assumes yearly cuts on the same model is likely wrong.
  • Expect savings to come from switching models. Keep an evaluation set ready so you can test a cheaper successor quickly. How to evaluate an LLM before switching gives a method.
  • Watch hosts more closely than makers. Some hosts reprice often, in both directions. If you buy an open model from a host, check its pricing page monthly or compare hosts with the LLM API price index.
  • Leave headroom for rises. 52 rises, with a median of 40%, is enough to plan for. See budgeting for LLM API price rises.

Methodology

  • Source. Each provider’s own pricing page as captured by the Internet Archive, one capture per provider per month, 2023 onward. 417 captures were fetched; 308 gave prices and 109 were skipped (for example, empty app shells), each with its reason. No aggregator data. The method, all rows and the skipped captures are on the LLM API price history dataset page, and every row is in the CSV.
  • Which price. Standard pay-as-you-go text input and output per 1M tokens. Batch, flex and priority prices are excluded. Where a page splits by context length, the short-context price is used. Euro prices (Mistral, late 2023 to early 2024) are left out of every USD figure.
  • Exclusions. Moves into or out of a price the page marks as promotional or introductory are not counted (2 moves). One row with a zero price on one side only, a probable page error, is left out.
  • Models. A model is the provider’s API ID where the page shows one, otherwise its name. A renamed model starts a new series, so some cuts that came as a new snapshot (for example a new dated GPT-3.5 Turbo) count as a new model, not a change.
  • Reproduce it. Every number in this post comes from the dataset JSON with a short script (same rules as the dataset page). 80% is 398 of 498 models seen in more than one calendar month; medians are over the 100 cuts and 52 rises; repricing shares are models with a list-price change ÷ models seen in more than one month, per provider.

Where BroadNet Inference fits

BroadRouter is BroadNet’s own site. BroadNet Inference has no price history in this dataset, because the Internet Archive had no captures of its pricing pages when we checked on 15 November 2026. Its current rates are published as a machine-readable pricing.json, and BroadRouter syncs its BroadNet prices from it every hour. For a fixed monthly cost instead of a per-token rate, BroadNet also sells a flat plan for DeepSeek V4.1 Flash at $60 a month per session (one request at a time). See flat rate vs pay per token for when that is cheaper.

Questions

How often do LLM API providers change their prices?

For most models, never. In our archive of 11 providers' pricing pages from 2023 to 2026, 80% of models seen in more than one month kept the same price in every capture. Most changes came from one host, DeepInfra, which repriced half of its long-listed models at least once.

Do LLM API prices usually go down?

Changes to an existing model were cuts more often than rises: 100 cuts against 52 rises. But the balance has shifted. In 2024 we recorded 45 cuts and 2 rises; in 2026 up to early November, 28 cuts and 31 rises.

How big is a typical LLM API price cut?

The median cut was 19% on a blended price (3 input tokens per output token). Cuts by model makers were much larger, with a median of 58% across 15 cuts, against 14% for 85 cuts by hosts.

Why do OpenAI, Anthropic and Google prices seem to fall so fast if they rarely reprice?

Because each new model version arrives at its own price. OpenAI repriced 5 of 77 long-listed models, Anthropic 1 of 26 and Google 2 of 23. The drop from GPT-4 to later GPT models came almost entirely through new versions.

Is BroadNet Inference in this dataset?

No. The Internet Archive had no captures of BroadNet's pricing pages when we queried it on 15 November 2026, so BroadNet has no price history here. Its current prices are in its machine-readable pricing.json.

Read as MarkdownAsk ChatGPTAsk ClaudeAsk Perplexity