Background
Market report · data through 2026-10-05

The state of AI pricing, measured.

Every list-price movement across 22 providers and 66 tracked tiers, counted from the nightly record. Each observation keeps its own verified or carried-forward status. No estimates and no forecasts: this page recomputes from the same append-only history as the rest of the site.

94
list-price movements recorded across 66 tiers since May 2026
21
price cuts
28
price raises
44
model swaps: a slot changing model, not a repricing

1 movement moved input and output in opposite directions and count as neither cut nor raise.

The last 30 days

6 movements in the window 2026-09-06 to 2026-10-05, across 66 tracked tiers.

DateProviderTierWas (in / out)Now (in / out)What
2026-10-01DeepInfraGemma 4 31B Turbo · Flagship$0.2 / $0.8$0.09 / $0.34model swap
2026-09-29OpenRouterDeepSeek V3 · Mid$0.24 / $0.9$0.25 / $1.00raise
in +4%, out +11%
2026-09-25OpenAIGPT-6 Luna · Small$0.2 / $1.20$0.1 / $0.5model swap
2026-09-25OpenAIGPT-6 Sol · Flagship$5.00 / $30.00$2.00 / $10.00model swap
2026-09-25AnthropicClaude Opus 5.5 · Flagship$10.00 / $50.00$4.00 / $20.00model swap
2026-09-10DeepSeekV4.1 Flash · Small · peak rate$0.44 / $1.32$0.3 / $1.20
off-peak $0.15 / $0.6
model swap

Scroll the table sideways to see every column.

A movement that coincides with a recorded model change in that slot is classed as a model swap, not a repricing: the rate moved because the model behind the tier changed. Input and output percentage changes are calculated separately on this page. The tier column shows the current occupant of the slot. The dated story behind each move is on Price Moves; retired models rest in the Model Graveyard.

Where the market sits

The median tracked rate fell in 1 of 3 tier classes over the last 30 days: Small input unchanged and output −9%. Mid and Flagship were unchanged.

Tier classMedian now (in / out)30 days ago (in / out)
Small · 22 providers$0.275 / $1.05$0.275 / $1.15
Mid · 22 providers$1.02 / $2.85$1.02 / $2.85
Flagship · 22 providers$1.70 / $4.45$1.70 / $4.45

Scroll the table sideways to see every column.

Rows compare Small input and output, Mid input and output, and Flagship input and output medians. Median of list prices across providers tracked on both dates. Medians move only when enough of the market moves: that is the point of quoting them.

The cheapest tracked rate is Llama 3.1 8B on DeepInfra at $0.02 per 1M input tokens. The priciest tracked flagship is Fugu Ultra at $5.00: a spread of about 250× for a million tokens of input. Capability differs as widely as price; the full ranking lists every provider.

In content terms: the cheapest tracked way to generate The Hobbit (95,356 words ≈ 127,141 tokens) as output is $0.0051 on Llama 3.1 8B (DeepInfra).

The ledger, month by month

MonthMovementsCutsRaisesMixedModel swaps
May 2026229904
June 2026155901
July 20263865027
August 2026131417
September 202650104
October 202610001

Scroll the table sideways to see every column.

Nightly tracking began mid-May 2026, so May is a partial month. A movement dated in May can also be the first verified reading catching up with an earlier change. A mixed movement moved input and output in opposite directions, so it counts as neither a cut nor a raise. Historical rows are never rewritten.

About these numbers. Provider pages remain the authority for current published prices. TokenScale checks those sources nightly and keeps an independent append-only record of what was verified, changed, corrected or carried forward (data through 2026-10-05). If a price cannot be reproduced, its last known value is carried and marked; a later successful check does not change that earlier status. Promotions stay outside the list-price record. A movement that coincides with a recorded model change is a swap, while TokenScale corrections are labelled separately on Price Moves. See the methodology and the reliability record.

More pricing pages