Cost of Intelligence · flagship $/1M tokens on record
The cost of intelligence fell, then the newest frontier reset it up.
$25.00/1M · 2023 $4.06/1M · 2025 low $7.81/1M · 2026 ≈3.2× below 2023, but up 1.9× since the 2025 low
The price of flagship large-language-model inference, measured in dollars per million tokens at the blended 1:3 input:output basket. The floor fell sharply from $25.00 in 2023 to a low of $4.06 in 2025 (6.2× down), then the newest-generation flagship reset list prices back up to $7.81 in 2026 — 1.9× above the low. Net, today's floor sits about 3.2× below 2023, but the trajectory is not a clean decline: it bottomed in 2025 and the most recent move was up.
The headline tracks the flagship tier only and excludes reasoning/extended-thinking models (their visible-token list price understates true cost) and any reading below high confidence. Every figure is a dated reading with a cited source, not a live quote.
The floor is the cheapest still-listed flagship-class model, not the newest. The 2026 floor is Gemini 2.5 Pro ($7.81/1M); the very newest frontier sits higher — gpt-5.5 ($23.75/1M), Claude Opus 4.6 ($20.00/1M), Claude Opus 4.8 ($20.00/1M) — so the recent uptick is partly a current-vintage effect (a maturing prior-gen model held the 2025 low). All of these readings are in the ledger below.
Section · What that buys
Summarising a ~150,000-token book
One concrete task, costed at the flagship floor in each year — so the abstract $/1M figure lands as a number you can feel. The token assumption is stated; nothing is fudged to hit a dramatic figure.
Assumption: 150,000 input tokens (≈ a 150k-token book) + 3,000 output tokens = 153,000 tokens, priced at the blended flagship floor: $25.00/1M in 2023, $7.81/1M in 2026.
$3.82 → $1.20 for the same job, on the flagship floor. The drop is the index, not a discount: list prices, blended basket, the openai/google readings the ledger cites below.
Section · The reading ledger
Every reading on file
The data sheet behind the index: each row is one model at one observed date, with its blended and raw list prices and the receipt behind it — source, observed-at, freshness and confidence. Grouped by provider line, each opening with a mini-trend and the first→current drop on file; readings within a line run newest first. Reasoning models and demoted readings stay in the record, flagged.
- Summary row — a provider line's mini-trend and first→current drop, from confident readings only.
- Dagger — a reading demoted below confident (low confidence or unclean backfill). Kept on the record, excluded from the headline and chart.
- Amber left edge — a demoted reading's row.
- Oxblood left edge — a reasoning model, flagged and held out of the flagship basket.
- Confidence / freshness tags — blue = on file / high confidence or fresh; amber = pending / low confidence or aging; hollow dot = reported, not yet source-confirmed.
| Provider / model | Tier | Blended | Input | Output | Observed | Freshness | Source | Confidence |
|---|---|---|---|---|---|---|---|---|
| anthropic Claude 2.1 | flagship | one confident reading on file | ||||||
| anthropic Claude 2.1 reading | flagship | $20.00 † | $8.00 | $24.00 | Mar 2024 | stale · 2y old | Archived backfill | unclean backfill |
| anthropic Claude 3 Haiku | budget | one confident reading on file | ||||||
| anthropic Claude 3 Haiku reading | budget | $1.00 | $0.25 | $1.25 | Mar 2024 | stale · 2y old | Archived backfill | high conf. |
| anthropic Claude 3 Opus | flagship | one confident reading on file | ||||||
| anthropic Claude 3 Opus reading | flagship | $60.00 | $15.00 | $75.00 | Mar 2024 | stale · 2y old | Archived backfill | high conf. |
| anthropic Claude 3 Sonnet | mid | one confident reading on file | ||||||
| anthropic Claude 3 Sonnet reading | mid | $12.00 | $3.00 | $15.00 | Mar 2024 | stale · 2y old | Archived backfill | high conf. |
| anthropic Claude Haiku 4.5 | budget | one confident reading on file | ||||||
| anthropic Claude Haiku 4.5 reading | budget | $4.00 | $1.00 | $5.00 | Jun 2026 | 51d old | Official page fetch | high conf. |
| anthropic Claude Opus 4.6 | flagship | one confident reading on file | ||||||
| anthropic Claude Opus 4.6 reading | flagship | $20.00 | $5.00 | $25.00 | Jun 2026 | 51d old | Official page fetch | high conf. |
| anthropic Claude Opus 4.8 | flagship | one confident reading on file | ||||||
| anthropic Claude Opus 4.8 reading | flagship | $20.00 | $5.00 | $25.00 | Jun 2026 | 51d old | Official page fetch | high conf. |
| anthropic Claude Sonnet 4.6 | mid | one confident reading on file | ||||||
| anthropic Claude Sonnet 4.6 reading | mid | $12.00 | $3.00 | $15.00 | Jun 2026 | 51d old | Official page fetch | high conf. |
| google Gemini 1.5 Flash | budget | $0.88 · 2024 $0.24 · 2025 3.6× on file | ||||||
| google Gemini 1.5 Flash reading | budget | $0.24 | $0.08 | $0.30 | Feb 2025 | stale · 1.5y old | Archived backfill | high conf. |
| google Gemini 1.5 Flash reading | budget | $0.88 | $0.35 | $1.05 | Jun 2024 | stale · 2y old | Archived backfill | high conf. |
| google Gemini 1.5 Pro | flagship | $8.75 · 2024 $4.06 · 2025 2.2× on file | ||||||
| google Gemini 1.5 Pro reading | flagship | $4.06 | $1.25 | $5.00 | Feb 2025 | stale · 1.5y old | Archived backfill | high conf. |
| google Gemini 1.5 Pro reading | flagship | $8.75 | $3.50 | $10.50 | Jun 2024 | stale · 2y old | Archived backfill | high conf. |
| google Gemini 2.0 Flash | budget | one confident reading on file | ||||||
| google Gemini 2.0 Flash reading | budget | $0.33 | $0.10 | $0.40 | Feb 2025 | stale · 1.5y old | Archived backfill | high conf. |
| google Gemini 2.5 Flash | budget | one confident reading on file | ||||||
| google Gemini 2.5 Flash reading | budget | $1.95 | $0.30 | $2.50 | Jun 2026 | 51d old | Official page fetch | high conf. |
| google Gemini 2.5 Flash-Lite | budget | one confident reading on file | ||||||
| google Gemini 2.5 Flash-Lite reading | budget | $0.33 | $0.10 | $0.40 | Jun 2026 | 51d old | Official page fetch | high conf. |
| google Gemini 2.5 Pro | flagship | one confident reading on file | ||||||
| google Gemini 2.5 Pro reading | flagship | $7.81 | $1.25 | $10.00 | Jun 2026 | 51d old | Official page fetch | high conf. |
| google Gemini 3.1 Pro Preview | flagship | one confident reading on file | ||||||
| google Gemini 3.1 Pro Preview reading | flagship | $9.50 | $2.00 | $12.00 | Jun 2026 | 51d old | Official page fetch | high conf. |
| mistral Codestral | budget | one confident reading on file | ||||||
| mistral Codestral reading | budget | $0.75 | $0.30 | $0.90 | Jun 2026 | 51d old | Official page fetch | high conf. |
| mistral Mistral Large 3 | mid | one confident reading on file | ||||||
| mistral Mistral Large 3 reading | mid | $1.25 | $0.50 | $1.50 | Jun 2026 | 51d old | Official page fetch | high conf. |
| mistral Mistral Medium | mid | one confident reading on file | ||||||
| mistral Mistral Medium reading | mid | — † | $2.50 | $7.50 | Jan 2024 | stale · 3y old | Archived backfill | unclean backfill |
| mistral Mistral Small | budget | one confident reading on file | ||||||
| mistral Mistral Small reading | budget | — † | $0.60 | $1.80 | Jan 2024 | stale · 3y old | Archived backfill | unclean backfill |
| mistral Mistral Tiny | budget | one confident reading on file | ||||||
| mistral Mistral Tiny reading | budget | — † | $0.14 | $0.42 | Jan 2024 | stale · 3y old | Archived backfill | unclean backfill |
| openai gpt-3.5-turbo | budget | $2.00 · 2023 $1.75 · 2023 1.1× on file | ||||||
| openai gpt-3.5-turbo reading | budget | $1.75 | $1.00 | $2.00 | Nov 2023 | stale · 3y old | Archived backfill | high conf. |
| openai gpt-3.5-turbo reading | budget | $2.00 | $2.00 | $2.00 | Mar 2023 | stale · 3y old | Archived backfill | high conf. |
| openai gpt-4 | flagship | one confident reading on file | ||||||
| openai gpt-4 reading | flagship | $52.50 | $30.00 | $60.00 | Mar 2023 | stale · 3y old | Archived backfill | high conf. |
| openai gpt-4-turbo | flagship | one confident reading on file | ||||||
| openai gpt-4-turbo reading | flagship | $25.00 | $10.00 | $30.00 | Nov 2023 | stale · 3y old | Archived backfill | high conf. |
| openai gpt-5.4 | flagship | one confident reading on file | ||||||
| openai gpt-5.4 reading | flagship | $11.88 | $2.50 | $15.00 | Jun 2026 | 51d old | Official page fetch | high conf. |
| openai gpt-5.5 | flagship | one confident reading on file | ||||||
| openai gpt-5.5 reading | flagship | $23.75 | $5.00 | $30.00 | Jun 2026 | 51d old | Official page fetch | high conf. |
Blended = (input + 3×output) ÷ 4, the basket the index tracks; input and output are the raw list prices. The first→current drop on each summary row uses confident readings only; the mini-trend line connects confident points (a hollow node marks a demoted reading, off the line). A † marks a reading demoted below confident — kept on the record, excluded from the headline. Source links open the archived snapshot or provider announcement of record; aggregators are cross-reference only.
Section · How this index is built
Method, provenance and caveats
Raw vs blended vs benchmark-adjusted, the reasoning caveat, and where each reading comes from.
The metric
blended-basket (1:3 input:output) is the only derivation in v1; benchmark-adjusted (quality_adjusted_per_1m) is a deferred phase 2 deliverable and is null
Raw vs blended vs adjusted
input_per_1m/output_per_1m are RAW-LIST; blended_per_1m is BLENDED-BASKET (a fixed-ratio modeling convenience, not a market measurement). quality_adjusted_per_1m is the BENCHMARK-ADJUSTED view and is NULL in v1 — no BENCHMARK-ADJUSTED numbers are emitted until a provenance-clean benchmark anchor lands (phase 2).
Reasoning caveat
o-series reasoning models burn invisible "thinking tokens" billed as output — actual API cost can be 2-5x the visible output. reasoning-tier blended prices therefore understate real cost and are kept in their own tier, never blended into the flagship/mid/budget headline.
Source of record
Provenance of record is a Wayback snapshot or an official provider announcement. Aggregators are cross-reference only.
- Archived backfill — an immutable Wayback snapshot of the provider's own pricing page on the reading's date.
- Official page fetch — a dated reading taken from the official provider page on its observed-at date. The receipt link is live and may now differ; it is a dated fetch, not an immutable snapshot.
- Recorded static — harvested from a provider announcement, pending live re-verification.
- Live pipeline — re-checked by the running price pipeline.
Freshness
Recorded through Jun 2026 is the date of the most recent reading on file — it is the newest observed-at date in the ledger below, never the date this page was published. It moves only when a new reading lands, so a deploy that changes nothing about this index does not re-date it.
That is a statement about the index, not about any one price. Each row carries its own observed-at date and its own age, measured against the dataset's build date; a reading older than 120 days is labelled stale in the ledger — never shown as fresh — and the count sits in the stat strip above.
Confidence & fail-closed
A reading is treated as confident only when it is high-confidence and its backfill is clean. Anything below that is demoted, flagged with a †, and kept out of the headline basket. If the series fails to load, this page shows an honest "data unavailable" state rather than a stale or invented number.
Figures are dated $/1M-token readings, each carrying its source and observed-at date — archived backfill, recorded static, or live pipeline. The headline tracks the flagship tier only and excludes reasoning models and sub-confident readings. Blended = (input + 3×output)/4. This index is a record of published list prices, not a live quote and not a promise of accuracy; confirm current pricing with the provider.
Section · For people and agents
Cite this index
The series is machine-readable and openly licensed (CC BY 4.0). Answer engines and agents can read, cite and link the underlying JSON directly.
This page also embeds schema.org Dataset markup so search and answer engines can index the series with its units, license and temporal coverage.