Two $10/$50 Flagships in 72 Hours: LLM Price Index Up 29%

GPT-6 Astra entered a ten-model price index on September 4 at $10/$50 and pushed it up 29% in a day — the largest move on record.

The number comes from ModelPriceWatch's dated report of September 1, 2026, whose author discloses maintaining the index: "Disclosure: I maintain ModelPriceWatch." The index is an equal-weight average of ten flagship models, one per lab, blended three parts input to one part output at each vendor's printed list price.

The move matters because of what the same report said a month earlier. Across 40 daily readings, the author wrote, "not one lab had ever changed the price of an existing model. Every move in the index had come from a new model replacing an old one." Three weeks of August ended that.

What actually moved in August

Date	Event	Index ($/Mtok)
Aug 16	DeepSeek V4 Pro repriced, flat $0.435/$0.87 to peak $1.32/$3.96 (+264% blended)	$4.46
Aug 21	GPT-5.6 Sol repriced $5/$30 to $4/$20, labelled promotional	$4.14
Sep 1	Closing level	$4.14

Net for the month: −5.7%, and −9.4% since the index's first reading on February 23. Three other flagship handovers — Muse Spark 1.1 to 1.2, Grok 4.5 to 4.6, GLM-5.2 to 5.3 — moved the index nothing, because each successor kept its predecessor's list price. That is the old pattern. The two repricings are the pattern breaking.

DeepSeek turned a list price into a schedule

Until 16:00 UTC on August 16, DeepSeek V4 Pro billed one flat rate: $0.435 input, $0.87 output per million tokens. After that the pricing page carried two tiers — peak at $1.32/$3.96 for the windows 01:00–04:00 and 06:00–10:00 UTC, and off-peak at exactly half, $0.66/$1.98, for every other hour.

Read the off-peak rate carefully. The report puts it at $0.99 blended, 82% above the old flat price, and characterises the change as "a 3x increase with a discount window attached" rather than a discount layered on the old rate. Its blunter warning: if you cannot batch, "your V4 Pro bill roughly tripled in mid-August and the vendor did not send you an email about it."

For teams that can schedule, the off-peak window runs 17 hours a day. The report's advice is direct: "If your pipeline runs at 08:00 UTC, you're paying double for no reason."

OpenAI's cut came with a clock on it

On August 21, GPT-5.6 Sol went from $5/$30 to $4/$20 — $11.25 to $8.00 blended, down 29%, and per the report the first list-price cut by any index constituent. A sentence appeared on the pricing page the same day: "GPT-5.6 Sol's promotional pricing is available at least through November 21, 2026."

That is a floor on the promotion, not a date the price returns. The report notes OpenAI publishes no reversion date, and recommends treating $4/$20 as expiring on November 21 until OpenAI says otherwise.

A reseller now sets the cheap end

The report also tracks a floor: the cheapest model clearing a fixed capability bar. It did not move in August, holding at $0.113 per million tokens, set by DeepSeek V4 Flash since July 25 — even though DeepSeek raised V4 Flash's own peak rate from $0.14/$0.28 to $0.44/$1.32, a 3.8x jump in blended terms.

The reason is that the floor reads the cheapest listed price across every host serving the model, and DeepInfra kept listing V4 Flash at $0.09/$0.18. The report's summary of the risk is worth quoting: "A lab's list price is a policy; a reseller's list price is a margin." On September 1, the report puts GPT-4-class capability at 37x cheaper than the flagship ceiling — a gap that now depends partly on one host's pricing decision.

What Astra costs, and what you get

Astra lists at $10 per million input tokens and $50 per million output, with a context window up to 1,050,000 tokens, maximum output up to 128,000 tokens, and the API model name gpt-6-astra. A fast mode doubles speed at a higher price. Rollout is staged: a limited group of organizations first, then paid ChatGPT users and developers across the OpenAI API, Microsoft Azure and AWS Bedrock.

OpenAI frames the launch around advanced computer use, software engineering, browsing, cybersecurity tasks and professional knowledge work, and says the model reaches a Critical level of cybersecurity capability under its Preparedness Framework. On the developer-facing side, the announcement quoted on Simon Willison's blog claims Astra "has more attention to detail, better understanding of the user's prompt, and can build more sophisticated outputs. In particular, it excels at building 3D models." That is the vendor's characterisation, not an independent evaluation.

On cost-per-capability the report is less flattering: at a published benchmark score of 96, Astra works out to $0.208 per point of measured intelligence, 3.5x the September 1 frontier average. Two days before Astra, Anthropic launched Claude Fable 5.1 at the same $10/$50 headline — but with cached input at $0.25, which the report calls 2.5% of the input rate and the deepest cache discount on any flagship card it tracks.

Two flagships at identical list prices, then, with very different economics for context-heavy work. If your workload resends the same corpus, codebase or long system prompt on every call, the cache rate decides the invoice, not the headline pair.

The spread reopened

Frontier flagship pricing spanned 21x on August 1 ($0.544 to $11.25 blended) and had narrowed to 13x by September 1 ($0.75 to $10.00), tightening from both ends at once. Astra reopened it to 27x in a day, and lifted the index above its February 23 first reading for the first time, up 16.8%.

The report declines to call a turn on one reading, and the caution is appropriate: "the frontier price never moves" was August's finding, and September's first four days delivered a reprice up, a reprice down, and the biggest step up on record.

What to check this week

• Your tier, not just your rate. Two of ten flagships now publish a price that means "up to" — one with a time-of-day schedule under it, one with a promo clock over it. The report's rule: "Write the tier next to the price, not just the date."
• Your cron's UTC hour. Batchable DeepSeek work scheduled inside the peak windows is paying double for nothing.
• November 21. Model Sol's reversion to $5/$30 after that date rather than assuming the promotion holds.
• Who lists your cheapest model. If it is a host rather than the lab, confirm the listing before renewing a budget against it.

The next reading to watch is whether a second lab adopts time-of-day pricing. One lab doing it is a pricing experiment; two would make it a format every buyer has to model.