State of LLM Pricing — August 2026

1. Frontier intelligence cost $4.39 per million tokens on August 1, 2026. The LLM Price Index — the equal-weight blended price across 10 current flagship models, one per major lab — closed that day at $4.39/Mtok, down 3.9% from $4.57 at the first reading on February 23, 2026.
2. Frontier list prices are structurally sticky. Across 40 index readings over five-plus months, the level moved on exactly two dates — July 18 and July 25 — and both moves came from basket rebalances (new flagships entering), not from any lab repricing an existing model. As of August 1 the level had been flat for 7 days, since the last rebalance.
3. The real deflation is at the floor, not the frontier. The cheapest model clearing a fixed GPT-4-class capability bar (GPQA Diamond ≥ 70) fell from $1.93/Mtok on March 1 (o4-mini) to $0.11/Mtok on July 25 (DeepSeek V4 Flash) — a 17x collapse in five months, leaving GPT-4-class capability 41x cheaper than the flagship ceiling.
4. "Frontier flagship" spans a 21x price range. Basket constituents run from $0.54/Mtok blended (DeepSeek V4 Pro) to $11.25/Mtok (GPT-5.6 Sol). Closed-weights flagships averaged $4.47 against $4.08 for open-weights at publication — a 10% proprietary premium.
5. Price per unit of measured intelligence averaged $0.0637 at publication (blended $/Mtok ÷ benchmark score, over the seven of ten constituents with a published score on the August-1 basket; on August 18 the Z.AI slot passed from GLM-5.2 to the not-yet-scored GLM-5.3, dropping coverage to six; on August 27 Meta's Muse Spark 1.2 earned independently measured scores, restoring coverage to seven of ten; and on September 5 Meta's slot passed in turn to the not-yet-scored Muse Spark 1.3, leaving the 6 of 10 constituents with a published score that the live index reports today). The best value at publication was DeepSeek V4 Pro at $0.0076 per point — 18x better than the frontier average.

The index opened its record at $4.57/Mtok on February 23, 2026 and held that level unchanged until mid-July. On July 18 it stepped down to $4.25 when xAI's flagship slot passed from Grok 4 to the cheaper Grok 4.5 (OpenAI's slot also passed from GPT-5.5 to GPT-5.6 Sol at an unchanged list price). On July 25 it stepped up to $4.39 when Alibaba's slot passed from Qwen3-Max to the pricier Qwen3.7-Max; the same day Anthropic's slot passed from Claude Opus 4.8 to Claude Opus 5 at identical list prices, and Moonshot (Kimi K3) entered as a ninth, previously unrepresented lab — a composition change that is chain-linked out of the trend, so it moves what is measured, not the measured trend. On August 1 Meta entered as a tenth lab (Muse Spark 1.1, its first proprietary paid-API flagship), again as a composition change.

A note on the basis. Every level above is stated on the current ten-lab basket. The index is chain-linked and rescaled so its most recent point equals the published headline, so admitting a new lab restates the whole level history without changing the measured trend: the same series read on the nine-lab basket in force earlier on August 1 opened at $4.84 and stood at $4.66. The trend is the invariant — −3.9% since February 23 either way. Compare levels within one basis, not across two. Correction (August 4, 2026): the trend's rescale factor previously anchored on the cent-rounded headline, which let display rounding drift historical points; fixing it to anchor on the exact mean restated the February 23 first reading from $4.56 to $4.57 and the February-to-August trend from −3.7% to −3.9%. The as-published headline levels archived at each deploy are unchanged.

Every move in the series is a succession event: a lab shipping a differently-priced flagship. No basket constituent changed its own list price in the period. That is the sticky-ceiling finding: frontier pricing moves by generation, not by discounting.

Hold the capability bar fixed and the story inverts. The cheapest model with a verified GPQA Diamond score of 70 or better — roughly GPT-4-class general reasoning — stepped down five times between March and August:

Every figure above is computed from the ModelPriceWatch tracker: 181 models across 27 providers, 6,218 price snapshots spanning February 2024 to August 2026, each price traced to the provider's official pricing page. The full dataset is public: