5 Months of LLM Pricing Data. The Market Is Moving Now.

In July we published our 12-week data summary. The headline finding was surprising: 22 of 33 providers had not moved a single price in 84 days. The big names were frozen. The message was that if your stack ran on OpenAI, Anthropic, or Google, your pricing was stable, at least for now.

That was then. At 5 months of data, the picture has changed. The number of providers with zero price changes is now zero. Every one of the 34 providers we track has moved at least one price since April. The market that looked frozen in July has been repricing steadily since.

Finding 1: Every provider moved a price. The frozen era is over.

At 12 weeks, we found 22 of 33 providers with no changes at all. At 5 months, that number is 0. Every provider in our dataset has updated at least one price since April 24.

Some of this is model launches: a new model appears at a new price point, which registers as a change. But looking at the data at the model level, the repricing is real. Providers that held flat through June and July started moving in August and September as competitive pressure from cheaper models accumulated.

Finding 2: OpenRouter's share of changes dropped from 83% to 58%

At 12 weeks, OpenRouter accounted for 83% of all detected price changes, 182 of 218 events. At 5 months, it still leads but its share has fallen to 57.9%, or 1,344 of 2,322 total events. The rest of the market has caught up in activity.

OpenRouter is a routing layer over dozens of upstream providers. When any of those providers changes a price, it shows up in our OpenRouter diff. It will always be the noisiest signal in this dataset. But the narrowing gap tells a real story: the 33 other providers collectively generated 978 change events across the full 162-day window. At 12 weeks, the same group had produced only 36. The direct providers are now moving prices at a pace that did not exist at the start of our tracking period.

Finding 3: A clear volatility divide has emerged

Strip out OpenRouter and look at the 33 direct providers. A two-tier structure has formed.

High-activity providers (10+ change events)

These providers repriced frequently, driven by model launches, competitive adjustments, or upstream cost changes: OpenAI (94), Novita (90), Amazon Bedrock (79), DeepInfra (77), DigitalOcean (68), Azure OpenAI (67), Azure AI Foundry (45), SiliconFlow (43), Cloudflare Workers AI (39), Together AI (38), Replicate (36), Mistral (35), Voyage AI (34), Google AI Studio (31), Google Vertex AI (29), Anthropic (19).

Low-activity providers (1 to 10 change events)

These providers made few moves: xAI (8), Cohere (7), v0 (6), Groq (4), DeepSeek (3), Hyperbolic (2), Cerebras (2), AI21 (2), SambaNova (1), Perplexity (1).

Finding 4: The biggest moves since July

The July post covered the first wave of significant changes (the Vertex cuts, the Mistral medium hike, the SambaNova MiniMax cut). Here is what happened in the 10 weeks since.

OpenAI cut gpt-5.6-luna 80% (July 10)

The GPT-5.6 family launched at aggressive prices and then dropped further within weeks. gpt-5.6-luna went from $1.00 to $0.20 per million input tokens, an 80% cut that makes it one of the sharpest single-model repricing events in the dataset.

Mistral continued repricing both directions

Mistral is the most active major European provider in the dataset. Since April, it has raised prices on some models and cut on others, often without changelog announcements. mistral-medium-latest is up 275% from April. codestral-latest and magistral-small-latest are down 70% and 70% respectively.

The pattern suggests Mistral is actively managing its model portfolio: retiring or discounting older general-purpose models while raising prices on the ones it wants to position as premium. If you are using multiple Mistral models, the net effect on your bill depends heavily on which ones.

Google Vertex cut Gemini 2.5 Flash 70% (June 22)

Covered in the July post but worth repeating because it is the largest sustained cut from a major frontier provider in our dataset. Gemini 2.5 Flash went from $1.00 to $0.30 per million input tokens and has held there since.

DeepSeek cut v4-pro 65% (ongoing)

DeepSeek has been one of the quieter providers in the dataset in terms of change frequency (only 3 events), but the magnitude of its main move is notable. deepseek-v4-pro dropped from $0.435 to $0.15 per million input tokens since May, a 65.5% cut.

DeepSeek remains the price benchmark that most teams are comparing against. At $0.15/M input, it is one of the cheapest frontier-class models in the dataset with stable pricing history.

Finding 5: The price spread compressed from 1.25M to 8,824x

In the July post, the cheapest model in the dataset was Cohere Embed v4 on Azure AI at $0.00012 per million tokens, and the most expensive was o1 Pro at $150. That was a 1.25 million times difference.

As of October, the spread has compressed significantly. The cheapest input price is now $0.017 per million (IBM Granite 4.0 Micro via watsonx), and the most expensive remains OpenAI o1-pro at $150 per million. That is 8,824x, a meaningful compression from the July figure.

The compression is driven by two forces: small models getting cheaper faster than frontier models are getting more expensive, and the overall catalog shifting toward mid-tier general purpose models as new providers entered our tracking window.

Finding 6: The catalog grew 35x in 162 days

On April 24 we tracked 38 models. As of October 3 we track 1,319. That is a 35x increase in 162 days, most of it driven by adding new providers and by providers launching new model variants at a pace we have not seen before.

The jump from 38 to 718 between April and June reflects adding the Phase 2 provider set (DeepInfra, Together AI, Fireworks, Bedrock, Azure OpenAI, and others). The growth since June is organic: providers releasing new model variants faster than teams can evaluate them.

The practical implication for FinOps teams: the model you benchmarked against in Q2 may already have two newer versions at different price points. The evaluation surface is compounding.

What this means for your AI spend in Q4

Five months of data suggests four things worth acting on:

1. The frozen-provider assumption is gone. At 12 weeks, you could rely on Anthropic, OpenAI, and Google being stable. At 5 months, all three have moved prices. Budget variance is now a real risk even on the biggest providers.
2. Silent repricing is the norm, not the exception. The Mistral medium hike (+275%), the OpenAI luna cut (−80%), the DeepSeek v4-pro cut (−65%) did not come with prominent changelog entries. They showed up in the daily diff. If you are not watching daily, you will find out on your invoice.
3. Cuts are outpacing hikes. The majority of significant moves in our dataset are price reductions, driven by competition from DeepSeek and open-source models. If you signed enterprise contracts based on April pricing, there is a reasonable chance your models have gotten cheaper since then and you may not be capturing that.
4. Cloud marketplaces move independently. Bedrock, Azure OpenAI, and Vertex AI each set their own prices for the same underlying models, and those prices diverge from the provider's own API. If you are on a cloud marketplace route, your effective rate may differ from what the provider publishes directly.