12 Weeks of LLM Pricing Data. Here's What Actually Changed.

Twelve weeks ago, we started scraping LLM pricing pages every day. Not because we expected prices to move constantly, but because the only way to know whether they do is to check.

Here's what 84 days of data actually show. Some of it is reassuring. Some of it is not.

Finding 1: 22 of 33 providers never moved a single price

The names you'd expect to see: Anthropic, OpenAI, Google AI Studio, Meta, DeepSeek, Groq, Fireworks, Cerebras, Cloudflare, Cohere, Perplexity, Replicate, not one price change across any model in 84 days.

That's not laziness. It's a signal. Frontier model providers who set prices tend to hold them. Price changes at that layer are strategic, they require sales coordination, enterprise contract alignment, and careful messaging. You don't change them lightly.

Finding 2: OpenRouter accounts for 83% of all detected changes

Of the 218 price changes we detected, 182 came from OpenRouter. This deserves context before you conclude that OpenRouter is volatile.

OpenRouter is a proxy layer that routes to dozens of underlying providers. When an upstream provider changes a price, or when OpenRouter adjusts its own routing and margin, it shows up as a price change on our end. OpenRouter also tracks a much wider model catalog than most providers, including many niche and experimental models with thin trading volume.

The practical implication: if you're using OpenRouter to access a model, you may be paying a different price than the provider's own API charges and that difference can shift without announcement.

Finding 3: The 36 non-OpenRouter changes tell a sharper story

Strip out OpenRouter, and you get 36 price changes across 10 providers. That's what the actual LLM market did in 12 weeks. Here are the ones worth knowing about.

Google cut Gemini prices significantly (June 22)

The biggest single-provider move in the dataset. Google cut input prices on its Vertex AI Gemini models by 50–70% in one day, with no corresponding output price change on several models.

These cuts happened on Vertex AI specifically, not on Google AI Studio. If you're using Google models through different access layers, you may be on different pricing.

Mistral raised prices on mistral-medium-latest (June 27)

The only significant price hike from a major European provider. Mistral increased mistral-medium-latest by 275% on both input and output, quietly, without a public announcement we could find.

This is precisely the kind of change that slips past teams who aren't actively monitoring. The model didn't change. The capabilities didn't change. The bill did.

SambaNova cut MiniMax-M2.7 by 75–90% (July 1)

The largest single-model percentage cut in the dataset, outside of OpenRouter.

OCI raised xAI Grok prices 67–900% (July 2)

Oracle Cloud Infrastructure sharply repriced several Grok models in a single day, the largest price increases in the non-OpenRouter dataset.

Worth noting: xAI's own API pricing was flat at the time of these changes. This was OCI-specific, a reminder that cloud marketplace pricing for the same model can diverge significantly from the provider's native API.

Together AI cut Qwen3.7-Max 50% (May 27)

The earliest notable move in the dataset. Together AI halved the price of their Qwen3.7 Max model just over a month into our tracking window.

Finding 4: The price spread is 1,250,000x

Across all 1,040 models in our dataset, the cheapest input price is $0.00012 per million tokens (Cohere Embed v4 on Azure AI). The most expensive is $150 per million (OpenAI's o1 Pro via OpenRouter).

That's a 1.25 million times difference between the cheapest and most expensive model in the catalog for input tokens alone.

This isn't a useful comparison in isolation: embedding models and reasoning models serve different purposes. But the spread does illustrate why model selection matters as much as provider negotiation. Teams that land on the wrong model, one tier higher than they need, can easily spend 5–20x more than necessary.

Finding 5: Model releases are accelerating

We added 20 new models in the last 30 days alone. xAI released five new Grok variants in a single day (July 15). OpenAI released the GPT-5.6 family (Sol, Terra, Luna) on July 10. Google added a new Gemini Flash Lite variant.

This pace creates a compounding problem: the model you benchmarked against three months ago may already be two generations behind. Worse, its successor may be priced differently at different providers, on different context tiers, with different batch discounts.

The evaluation surface is growing faster than most teams can keep up with manually.

What this means for your infrastructure bill

Three takeaways from 84 days of data:

1. Frontier providers are stable; the edges are volatile. If your stack is on OpenAI, Anthropic, or Google AI Studio, your pricing is unlikely to change without warning. If you're routing through cloud marketplaces (OCI is the clearest example in our data) or aggregators (OpenRouter), you're more exposed.
2. Price changes don't come with announcements. The Mistral hike (+275%), the OCI Grok repricing (+400–900%), the SambaNova cut (−90%), none of these had prominent changelog entries we could find. They showed up in the daily diff.
3. The model landscape is moving too fast for manual tracking. Twenty new models in 30 days means the right model for your use case this quarter may not have existed last quarter.