Version 1.3 · Updated November 12, 2026 · Base: March 2023 = 100
Token Price Index Methodology
1. Universe
The basket covers general-purpose hosted API models from OpenAI, Anthropic, Google, DeepSeek, Mistral, xAI, and Moonshot AI. Each priced SKU from these providers needs a reviewed tier or an explicit exclusion. Audio, dedicated coding models, restricted-access variants, shared effort profiles, and historical listings without confirmed current rates are excluded. New constituents enter at their first dated price observation; current rates are never applied to months before that observation.
2. Blended price
Each constituent's price is blended as (3 × input + output) / 4 per 1M tokens, assuming a 3:1 input:output token ratio. This matches the blending used across BenchLM's pricing statistics.
3. Activity rules
A constituent is active in month M if it has a price observation on or before M, has not exited under a sourced supersession or retirement record, and its most recent price observation is no more than 12 months old. The current month also requires an available numeric rate for the same canonical SKU. A new catalog date alone does not renew an unchanged price; a dated provider price check does. A successor launch does not retire an older API model by itself.
4. Index formula
Each tier's sub-index for month M is the median blended price of its active constituents, normalized so the tier's first month with data = 100. The headline index is the frontier tier, based at March 2023 = 100 (GPT-4 launch). Medians, not means, so a single extreme price cannot move the index.
5. Tiers
Every constituent carries a published tier: frontier (flagship, top-capability line), mid-tier (workhorse models), or budget (small/cheap models). Assignments are versioned in code and listed below; a model's tier does not change after entry unless the methodology version is bumped.
6. Revisions
Each site build first records new constituents and changed prices from the pricing catalog, then recomputes the monthly series. Price changes are dated when observed, including corrections recorded on the same day. Older monthly observations remain in the history. There are no frozen vintages: basket and rule changes can revise history and require a methodology version bump. Validation rejects unreviewed priced SKUs, unrecorded observations, and drift in the generated series.
7. Revision log
v1.3 (September 30, 2026): missing general-purpose models, including GPT-6, GPT-6.1 Sol, Claude Sonnet 5.5, Gemini 3.7 and 3.8 Flash, and the current DeepSeek APIs, entered at their first index observation. Build-time reconciliation now preserves price changes automatically and rejects unreviewed SKUs from the seven covered providers. DeepSeek V3 and R1 exit when their hosted aliases upgraded to V3.1 on August 21, 2025. Unavailable current rates and stale observations no longer gain freshness from an unrelated catalog update. v1.2 (September 22, 2026): Claude Opus 5.5 entered the frontier constituent tier at its confirmed launch list price. v1.1 (September 2, 2026): the constituent universe was extended to the 2026 flagship, workhorse, and small lines (GPT-5.5 and 5.6, Claude Opus 4.8 and 5, Claude Fable 5 and 5.1, Claude Sonnet 5, Grok 4.6, Kimi K3, Gemini 3.6 Flash). Each enters at its confirmed release date at the list price recorded in the pricing table; introductory rates are recorded where the provider published them. Provider retirements recorded in the model lifecycle registry now exit constituents on their effective date (Claude 3 Sonnet, Claude 3.5 Sonnet, Claude 3 Haiku), and Gemini 1.5 Pro and Flash are superseded by their 2.5 successors at the successors' registry date, matching the existing GPT-4 and Claude 3 Opus convention. The live-price join keeps the first (standard-rate) row when a model is listed more than once. History was recomputed under these rules. v1.0 (initial release) tracked 34 constituents through mid-2026.
8. Caveats
Early history is thin: March 2023 opens with a single frontier constituent (GPT-4), and sub-indexes begin when their tiers first have data. Prices use standard uncached API tiers and active introductory rates; DeepSeek uses peak rates. Caching, Batch, Flex, volume, and off-peak discounts are excluded. The median measures the changing basket, so entry and exit effects are part of the monthly change. It does not measure a fixed workload or quality-adjusted cost.
Current constituents and tier assignments
Frontier (39)
| Model | Creator | Blended $/1M |
|---|---|---|
| Grok 4.3 | xAI | $1.56 |
| Kimi 2.6 | Moonshot AI | $1.71 |
| DeepSeek V4 Pro 0813 | DeepSeek | $1.98 |
| Grok 4.6 | xAI | $3.00 |
| Grok 4.7 | xAI | $3.00 |
| Grok 4.5 | xAI | $3.00 |
| GPT-5.1 | OpenAI | $3.44 |
| Gemini 2.5 Pro | $3.44 | |
| GPT-5 (high) | OpenAI | $3.44 |
| o3 | OpenAI | $3.50 |
| GPT-6 Sol | OpenAI | $4.00 |
| GPT-6.1 Sol | OpenAI | $4.00 |
| Gemini 4 Argon | $4.00 | |
| Kimi K3.1 Pro | Moonshot AI | $4.00 |
| Gemini 4 Carbon | $4.00 | |
| GPT-4o | OpenAI | $4.38 |
| Gemini 3.1 Pro | $4.50 | |
| GPT-5.2 | OpenAI | $4.81 |
| GPT-5.3 Instant | OpenAI | $4.81 |
| GPT-5.4 | OpenAI | $5.63 |
| Kimi K3 | Moonshot AI | $6.00 |
| Claude Opus 5.5 | Anthropic | $8.00 |
| GPT-5.6 Sol | OpenAI | $8.00 |
| Claude Opus 4.6 | Anthropic | $10.00 |
| Claude Opus 4.8 | Anthropic | $10.00 |
| Claude Opus 5 | Anthropic | $10.00 |
| Claude Opus 4.7 | Anthropic | $10.00 |
| Claude Opus 4.5 | Anthropic | $10.00 |
| GPT-5.5 | OpenAI | $11.25 |
| Claude Fable 5.2 | Anthropic | $16.00 |
| Claude Fable 5 | Anthropic | $20.00 |
| Claude Fable 5.1 | Anthropic | $20.00 |
| GPT-6 Astra | OpenAI | $20.00 |
| o1 | OpenAI | $26.25 |
| o3-pro | OpenAI | $35.00 |
| GPT-5.2 Pro | OpenAI | $57.75 |
| GPT-5.4 Pro | OpenAI | $67.50 |
| GPT-5.5 Pro | OpenAI | $67.50 |
| o1-pro | OpenAI | $262.50 |
Mid-tier (20)
| Model | Creator | Blended $/1M |
|---|---|---|
| DeepSeek V4.1 Flash | DeepSeek | $0.53 |
| Mistral Large 3 | Mistral | $0.75 |
| Mistral Medium 3 | Mistral | $0.80 |
| Mistral Large 4 | Mistral | $1.05 |
| Gemini 3 Flash | $1.13 | |
| Gemini 3.6 Flash | $1.50 | |
| Gemini 3.8 Flash | $1.50 | |
| Gemini 3.7 Flash | $1.50 | |
| GPT-5.4 mini | OpenAI | $1.69 |
| o3-mini | OpenAI | $1.93 |
| GPT-6.1 Terra | OpenAI | $2.00 |
| Mistral Medium 3.5 128B | Mistral | $3.00 |
| Gemini 4 Hydrogen | $3.00 | |
| Gemini 3.5 Flash | $3.38 | |
| GPT-4.1 | OpenAI | $3.50 |
| Claude Sonnet 5 | Anthropic | $4.00 |
| Claude Sonnet 5.5 | Anthropic | $4.00 |
| GPT-5.6 Terra | OpenAI | $4.50 |
| Claude Sonnet 4.5 | Anthropic | $6.00 |
| Claude Sonnet 4.6 | Anthropic | $6.00 |
Budget (20)
| Model | Creator | Blended $/1M |
|---|---|---|
| Ministral 3 3B | Mistral | $0.10 |
| GPT-5 nano | OpenAI | $0.14 |
| Ministral 3 8B | Mistral | $0.15 |
| GPT-4.1 nano | OpenAI | $0.17 |
| Gemini 2.5 Flash-Lite | $0.17 | |
| GPT-6 Luna | OpenAI | $0.20 |
| Ministral 3 14B | Mistral | $0.20 |
| Claude Haiku 5.5 | Anthropic | $0.20 |
| GPT-4o mini | OpenAI | $0.26 |
| Mistral Small 4 | Mistral | $0.26 |
| Grok 3 Mini | xAI | $0.35 |
| GPT-5.6 Luna | OpenAI | $0.45 |
| GPT-5.4 nano | OpenAI | $0.46 |
| Gemini 3.1 Flash-Lite | $0.56 | |
| GPT-5 mini | OpenAI | $0.69 |
| GPT-4.1 mini | OpenAI | $0.70 |
| Gemini 2.5 Flash | $0.85 | |
| Gemini 3.5 Flash-Lite | $0.85 | |
| o4-mini | OpenAI | $1.93 |
| Claude Haiku 4.5 | Anthropic | $2.00 |
Questions about a rule or a constituent? The computation lives in scripts/lib/compute-price-index.js in the site's repository and is validated on every build. See the index itself at /token-price-index or download the series at /data/price-index.json.