Introducing the PALM SOUQ LLM Price Index

Today we are publishing the PALM SOUQ LLM Price Index: list prices for 36 models across 10 providers and hosts, covering input, output, cached input and cache write rates per 1M tokens. It is free to use and licensed CC BY 4.0, so you can put it in a spreadsheet, a pricing deck or a blog post as long as you credit PALM SOUQ.

Why we built it

PALM SOUQ prices every request it meters. To do that correctly we maintain a price table for every model our customers route through the gateway, including cache rates and the multipliers that change cost in ways a single headline price hides. Customers kept asking for that table in a form they could read and cite, so we cleaned it up and published it.

We also wanted one place where cache economics are visible next to headline prices. A model with a higher input price and a deep cache discount can be cheaper for a chat product with a long system prompt than a model that looks cheaper on the pricing page.

What is in the table

• Input and output price per 1M tokens at the standard tier and the shortest context band.
• Cached input price where the provider offers prompt caching, and the cache write price for providers that charge a write premium.
• Whether a batch discount is available, the context window, and a tier label: frontier, mid, small or reasoning.
• A link to a model guide for the most used models, with cost drivers, circuit breaker settings and a sample PALM SOUQ policy.

How to read blended cost

Headline input and output prices are hard to compare because workloads use them in different proportions. The index shows a blended price that assumes three input tokens for every output token, which is close to what we see for retrieval-heavy chat. On that basis the least expensive frontier model in the table today is Mistral Large 2 at $3.00 per 1M blended tokens and the most expensive is Claude Opus 4.1 at $30.00.

Your own ratio may be very different. Agents that read large files and write short tool calls can run at twenty input tokens per output token, while drafting tools often run closer to one to one. The calculator on the index page lets you change the ratio and the cache hit rate.

How we keep it current

Prices are taken from each provider's public pricing documentation. The table carries the date of the last full review, currently June 2026, and individual rows are corrected as soon as a provider announces a change, because the same table prices live traffic in the gateway. If you spot a price that has moved, email support@palmsouqltd.com and we will confirm and update it.

We do not publish latency or throughput benchmarks in the index. Those numbers depend on region, account tier, prompt size and time of day, and a single figure would mislead more than it helps. PALM SOUQ measures time to first token and throughput per tenant for the traffic you actually send, which is the number that matters for your product.

Citing the index

If you use the data, a link to the index page with the review date is enough. Model guides can be cited individually. We would love to hear what you build with it.