Pricing
Minimum API Price at ECI ≥125
Running minimum nominal API price per 1M tokens, using a 3:1 input/output blend, among exactly matched non-reasoning models with Epoch Capabilities Index ≥125. It tracks how the cost of obtaining a fixed broad-capability threshold changed, not price per ECI point or total task cost.
Minimum API Price at ECI ≥125
The minimum blended API price for ECI ≥ 125 fell from $37.50 to $0.1225 per million tokens across the observed frontier.
Methodology
Capability is defined with the Epoch Capabilities Index (ECI), a composite Item Response Theory measure estimated from overlapping results across many benchmarks. Epoch fits model capability, benchmark difficulty, and benchmark slope jointly, then rescales the index so Claude 3.5 Sonnet is 130 and GPT-5 is 150. This chart uses a fixed gate of ECI ≥ 125 based on each model's central estimate. It does not average raw benchmark percentages, compare percentages from different tests, or divide price by a benchmark score.
Token price follows the convention in Epoch AI's inference-price analysis. For a model m, blended price P(m) = [3 × input_price(m) + 1 × output_price(m)] / 4, in nominal USD per one million tokens. Epoch uses a first-party API price when one exists; otherwise it uses the median across API providers, and it uses the shortest-context tier where context length changes price. The Phi-4 observation is the only displayed row without a first-party API and therefore uses Epoch's median-provider price.
The plotted frontier is F(t) = min{P(m): release_or_price_date(m) ≤ t, ECI(m) ≥ 125, and m is a non-reasoning model}. A row is emitted only when F(t) reaches a new low. Model identities were matched explicitly between Epoch's ECI file and its published inference-price files: GPT-4-0314/GPT-4 (Mar 2023), GPT-4o (May 2024), Claude 3.5 Sonnet (Jun 2024), GPT-4o mini, Gemini 1.5 Flash (Sep 2024), and Phi-4. No fuzzy or proxy benchmark joins were used.
Reasoning models are excluded because hidden or visible reasoning tokens can make per-token list prices misleadingly cheap relative to total task cost. The line is a release-event frontier, not a monthly average: connecting points indicates the best observed price remained available until the next qualifying observation, but the CSV does not manufacture monthly carry-forward rows. The final point is dated 2024-12-13, when Epoch records the Phi-4 API-price observation; Epoch's ECI file dates that model to 2024-12-12.
Limitations are material. ECI is an interval-like, arbitrarily anchored index whose estimates and confidence intervals can change when Epoch refits the model; this is why the chart uses a threshold rather than price divided by ECI. Eligibility uses central estimates even when an ECI confidence interval crosses 125. List price per token does not capture token efficiency, latency, caching, batch discounts, context-tier differences, provider reliability, or total evaluation cost. The source pricing analysis was last updated November 20, 2025 and does not establish that no later model through August 27, 2026 beat the final frontier, so no values are extrapolated after the last matched observation.
Frequently asked questions
What is one unit of capability here?
It is a fixed capability bundle: an Epoch Capabilities Index point estimate of at least 125. The y-value is the cheapest blended API price that could buy access to that level, not dollars divided by 125.
Why does cost at a fixed capability threshold matter for AI?
It separates access-cost change from the moving capability frontier: readers can see how the minimum token price for models clearing the same broad ECI threshold changed. It does not show application-level cost, which also depends on token use, quality, latency, and other operating requirements.
Why not divide API price by a benchmark percentage?
A percentage point on GPQA is not comparable to a percentage point on MMLU or a coding benchmark, and benchmark saturation makes such ratios unstable. ECI statistically links overlapping benchmarks onto one scale, while the fixed threshold avoids treating its arbitrary zero as meaningful.
How is the blended token price calculated?
The chart uses 75% input-token price and 25% output-token price: (3 × input + output) / 4. This matches Epoch's published inference-price methodology.
Which sources and models enter the frontier?
The chart exactly joins named models in Epoch's ECI scores and inference-price data, keeps non-reasoning models with a central ECI estimate of at least 125, and plots only strict running price minima. Models without an exact identity match are excluded rather than joined by fuzzy names or proxies.
Does the last point mean the frontier stayed unchanged through August 2026?
No. It is only the last defensible exact join in the cited published price dataset. The chart does not carry that value forward or claim complete model coverage after the source cutoff.