Frontier LLM prices broke a five-month freeze in August, with DeepSeek tripling one rate

A five-month freeze broke in three weeks

According to a report published on dev.to by the maintainer of ModelPriceWatch, an equal-weight index of ten flagship models — one per lab, blended at three parts input to one part output — went 40 daily readings without a single lab repricing an existing model. Every move since the index began on February 23 had come from new models replacing old ones. August ended that pattern twice over.

The month's first move was the familiar kind: on August 4, Alibaba's slot in the basket swapped Qwen3.7-Max for Qwen3.8-Max at $3.00 blended, down from $3.75, taking the index from $4.39 to $4.32 per million tokens. Three other August handovers — Muse Spark 1.1 to 1.2, Grok 4.5 to 4.6, and GLM-5.2 to 5.3 — left the index untouched because each successor kept its predecessor's list price.

The two pattern-breaking moves came from DeepSeek and OpenAI. The index closed the month at $4.14, down 5.7% for August and 9.4% since February.

DeepSeek turned a flat rate into a schedule

Until 16:00 UTC on August 16, DeepSeek V4 Pro billed a single flat rate of $0.435 input and $0.87 output. The pricing page then split into two tiers: peak hours (01:00–04:00 and 06:00–10:00 UTC) at $1.32/$3.96, and all other hours at exactly half, $0.66/$1.98.

The index reads the peak rate as the list price, which the author justifies on two grounds: DeepSeek defines off-peak as a discount from peak rather than the reverse, and a caller who cannot schedule around the clock needs a ceiling, not a floor. By that measure the blended price rose 264%. The report notes that even the off-peak blended rate of $0.99 sits 82% above the old flat price, making this a roughly 3x increase with a discount window attached rather than a discount layered onto the old price. The author adds that the vendor sent no email notifying customers of the change.

OpenAI's cut carries a clock

On August 21, GPT-5.6 Sol dropped from $5/$30 to $4/$20 — $11.25 to $8.00 blended, a 29% cut and the first list-price reduction by any model in the index's history. The catch is a sentence added to the pricing page the same day: the promotional pricing is "available at least through November 21, 2026." That is a floor, not a reversion date, and the report treats any budget built on Sol's new rates past that date as contingent on wording that can change.

Two of the ten flagships now publish a list price that effectively means "up to": one has a time-of-day schedule beneath it, the other a promotional clock over it.

The cheap end is priced by a reseller

The report also tracks a floor: the cheapest model clearing GPQA Diamond ≥ 70, roughly GPT-4-class capability. That price did not move in August, holding at $0.113 per million tokens on DeepSeek V4 Flash. Yet DeepSeek raised V4 Flash's own peak rate 3.8x on the same day it repriced V4 Pro. The floor held only because third-party host DeepInfra kept serving V4 Flash at $0.09/$0.18, or 5.9x below the lab's peak rate. As of September 1, GPT-4-class capability cost 37x less than the flagship ceiling — a gap that now depends partly on one host's margin rather than the lab's policy.

The flagship spread itself narrowed during August, from 21x on August 1 to 13x on September 1, with OpenAI's cut pulling the top down to Claude Opus 5's $10.00 and DeepSeek's peak pricing lifting the bottom to Mistral Large 3's $0.75.

September has already outdone August

The report is dated September 1, but the author flags the 72 hours that followed. On September 1, Anthropic launched Claude Fable 5.1 at $10/$50 with cached input at $0.25 — 2.5% of the input rate, the deepest cache discount on any card the author tracks. On September 3, OpenAI released GPT-6 Astra at the same $10/$50 and named it the default flagship; it took the OpenAI slot on September 4 and the index jumped from $4.14 to $5.34. That 29% single-day move is the largest in the index's history and the first time it has exceeded its February 23 starting level.

Why it matters

List prices have stopped being stable inputs for cost modelling. A DeepSeek pipeline running at 08:00 UTC now pays double what an off-peak run costs, though a 17-hour off-peak window exists for batchable workloads. Sol's rates may revert after November 21, so the report suggests modelling the reversion and being pleasantly surprised. And the cheapest GPT-4-class token is currently priced by a hosting reseller whose listing can change independently of the lab's own rates. The practical guidance from the report: write the tier next to the price — peak or promotional — and verify who is actually setting the price of your cheapest model before renewing a budget.