Inference Token Prices
What hosted AI models cost to run and how fast they respond, provider by provider.
- 2,397token-price quotes
- 381models priced
Sample
2 recordsCurrent collected record (output price)
Representative. The study measures output price per million tokens on cheapest-available models; tiers are not pooled.
{
"observed_at": "2026-04-02T00:00:00Z",
"provider": "provider-a",
"model_family": "llama",
"model_listing": "llama-3.1-8b-instruct",
"output_price_per_mtok_usd": 0.02,
"input_price_per_mtok_usd": 0.01,
"model_tier": "small_open"
}Marketplace utilization snapshot
Real row shape from the reference sample. 1,366 snapshots, every one carrying a computed ratio.
provider,sku_canonical,instance_total,instance_available,instance_rented,utilization_ratio,snapshot_ts marketplace-a,RTX_3090_24GB,2368,1120,1248,0.527,2025-12-03T09:36:00Z
Specifications
| Formats | Parquet, JSON, CSV |
|---|---|
| Delivery | S3, REST API, Parquet |
| Cadence | Continuous polling, daily aggregates |
| Token-price quotes | 2,397 (six-month reference sample) |
| Models priced | 381 across 12 model families |
| Providers | 5 hosted-inference providers |
| Fields | Input and output USD per million tokens, context window, per snapshot |
| Marketplace utilization | 1,366 snapshots: instances total, available, rented, and utilization ratio |
| History | Token pricing from March 2024 onward |
| Provenance | Row-level source URL and collection method |
| Access | For internal research and model development. Sourced from public pricing surfaces; no PII or MNPI. |
What ships
Token Price Index
Input and output token rates per model per provider over time - the unit economics of inference, tracked continuously.
Latency & Throughput
Time-to-first-token, generation throughput, and latency percentiles - congestion signals that move before capacity announcements.
Reliability Signal
Request success and error rates across providers - a real-time read on where inference demand is outrunning supply.
Fields
11 fields
| Field | Description |
|---|---|
| observed_attimestamp · UTC | When the price was collected from the provider pricing surface.Example 2026-04-02T00:00:00Z |
| providerstring | One of four hosted-inference providers.Example provider-a |
| model_familystring | Open-weight lineage the listing is grouped under (Llama, Qwen, Mixtral, DeepSeek).Example llama |
| model_listingstring · nullable | Provider raw model product name before family grouping.Example llama-3.1-8b-instruct |
| output_price_per_mtok_usdfloat · USD / 1M output tokens | Normalized output-token price - the quantity the study actually measures.Example 0.02 |
| input_price_per_mtok_usdfloat · USD / 1M input tokens · nullable | Input-token price where listed.Example 0.01 |
| model_tierstring · nullable | Coarse tier flag; the study does not pool tiers.Example small_open |
| context_windowint · tokens · nullable | Context window listed for the model at that snapshot.Example 131072 |
| instance_totalint · nullable | Marketplace inventory: instances listed for the provider and SKU.Example 2368 |
| instance_rentedint · nullable | Marketplace inventory: instances currently rented.Example 1248 |
| utilization_ratiofloat · 0..1 · nullable | Rented over total for the provider, SKU, and region at that snapshot. The supply-tightness read.Example 0.527 |
Quality checks
2 checksCheapest-output price decline
How the price of the cheapest available output tokens moved over time.
Track the cheapest available output price per million tokens across the four providers over the window.
Descriptive, published as preliminary: fell from about $0.13 per million tokens in mid-2024 into the $0.01 to $0.03 range in 2025-2026. Direction robust; absolute level preliminary due to mixed tiers.
Marketplace utilization as a supply-tightness read
Whether rented-versus-available instance counts identify capacity tightening on a marketplace.
Poll marketplace inventory per provider, SKU, and region, recording instances total, available, and rented, and derive a utilization ratio per snapshot.
Real measured: 1,366 utilization snapshots across the reference window, every row carrying a computed ratio. Latency and success-rate telemetry is a separate collection surface and is quoted only where it is actually collected.
Ground truth
| What is correct | For the descriptive layer, the observed posted output-token prices across the four providers on a cheapest-available basis. For the proposed demand signal, there is no ground truth yet because the data is uncollected. |
|---|---|
| How it is established | Descriptive aggregation of the cheapest-available output price over time per provider and family. No predictive grader exists; the proposed demand-side validation is future work gated on API access. |
How it is built
- 01
Collection
Collect from the public pricing surfaces of four hosted-inference providers.
- 02
Normalization
Normalize to a per-million-token basis for output tokens.
- 03
Family grouping
Group heterogeneous model listings by the underlying open-weight family so comparisons stay within recognizable lineages.
- 04
Tier separation
Observed models span small open-weight checkpoints to frontier-scale hosted deployments priced very differently; tiers are not pooled into a single level.
- 05
Direction-over-level reporting
Report the direction - a steep decline at the cheap end - as robust, and treat absolute price levels as preliminary because the cheapest quote in any period may reflect a different tier than in another.
Request a sample
We send real records in the format you would receive, with the field list and where each record came from. Access for evaluation is limited and covered by a mutual NDA.
ProductInference Token Prices