Are AI API prices going up or down in August 2026?

Both, in opposite directions depending on the provider. OpenAI cut GPT-5.6 Sol from $5/$30 to $4/$20 on August 21, 2026, Anthropic cancelled a scheduled Claude Sonnet 5 increase and made $2/$10 permanent, and Google launched Gemini 3.7 Flash at an introductory $0.75/$3.75. Meanwhile DeepSeek raised V4 Pro and V4 Flash prices on August 16 and introduced peak/off-peak splitting. Western frontier labs are cutting to defend share while the cheapest tier repriced upward.

What is the cheapest frontier-class AI API in August 2026?

Among named frontier-tier models, GPT-5.6 Sol at $4 input and $20 output per million tokens is now the cheapest of the top tier after its August 21 cut, undercutting Claude Opus 5 at $5/$25. If you drop to the fast tier, Gemini 3.7 Flash at $0.75/$3.75 introductory and GPT-5.6 Luna at $0.20/$1.20 are dramatically cheaper — roughly $0.041 and $0.012 respectively on a 30K-in/5K-out task.

Which August 2026 prices are temporary?

Two are explicitly time-boxed. GPT-5.6 Sol's $4/$20 has been reported as a promotional rate running roughly three months, and Gemini 3.7 Flash's $0.75/$3.75 is a 50% introductory rate that reverts to $1.50/$7.50 on January 1, 2027. Claude Sonnet 5's $2/$10 is the opposite — Anthropic explicitly stated the previously scheduled increase will not occur, making it the standard price.

Why did DeepSeek raise prices when everyone else cut?

DeepSeek repriced V4 Pro and V4 Flash upward at 16:00 UTC on August 16, 2026 and split rates into peak and off-peak windows, roughly doubling peak pricing versus off-peak. The most plausible reading is capacity management now that it has real volume: the aggressive flat pricing that won it market share became expensive to serve, and time-of-day pricing shifts load away from congested hours rather than simply raising the headline rate.

Quick Answer

AI API Prices August 2026: Who Cut, Who Raised

Published: August 25, 2026

The Short Answer

August 2026 broke the “inference always gets cheaper” narrative. Four major providers moved prices within eleven days, in two different directions.

OpenAI cut its flagship 20–33%
Anthropic cancelled a planned increase
Google launched at a half-price introductory rate
DeepSeek raised prices and added peak pricing

The Full August 2026 Picture

Date	Provider	Move	Detail
Jul 30	OpenAI	⬇️ Cut	Terra −20%, Luna −80%
~Aug 10	Anthropic	⬇️ Cancelled hike	Sonnet 5 stays $2/$10, permanently
Aug 13	Google	⬇️ Intro rate	Gemini 3.7 Flash $0.75/$3.75 (50% off)
Aug 16	DeepSeek	⬆️ Raised	V4 Pro & Flash repriced, peak/off-peak split
Aug 21	OpenAI	⬇️ Cut	Sol $5/$30 → $4/$20

Current Rates, Verified August 25, 2026

Model	Input / MTok	Output / MTok	30K/5K task	Status
Claude Opus 5	$5.00	$25.00	~$0.275	Standard
GPT-5.6 Sol	$4.00	$20.00	~$0.22	⏳ Promotional
GPT-5.6 Terra	$2.00	$12.00	~$0.12	Standard
Claude Sonnet 5	$2.00	$10.00	~$0.11	✅ Permanent
Grok 4.6	$2.00	$6.00	~$0.09	Standard
GLM-5.3	$1.40	$4.40	~$0.064	Standard
Gemini 3.7 Flash	$0.75	$3.75	~$0.041	⏳ Intro to Dec 31, 2026
DeepSeek V4 Pro	$0.66	$1.98	~$0.030	⬆️ Off-peak; 2x at peak
DeepSeek V4 Flash	$0.22	$0.66	~$0.010	⬆️ Off-peak; 2x at peak
GPT-5.6 Luna	$0.20	$1.20	~$0.012	Standard

The Frontier Order Flipped

The single most consequential change: GPT-5.6 Sol is now cheaper than Claude Opus 5, having been more expensive before August 21.

Sol: ~$0.22 per 30K-in/5K-out task
Opus 5: ~$0.275 per equivalent task

Sol’s output cut (33%) was steeper than its input cut (20%), which is deliberate. Agentic workloads are output-heavy — an agent that reads a codebase once then writes across dozens of turns generates far more output than input. Cutting output pricing hardest is a targeted bid for exactly the workload class where Anthropic has been strongest.

DeepSeek’s Peak Pricing Is the Underrated Story

DeepSeek repriced at 16:00 UTC on August 16, 2026, moving from flat rates to a peak/off-peak split where peak costs exactly double off-peak.

Peak hours: 01:00–04:00 and 06:00–10:00 UTC (7 hours). The remaining 17 hours are off-peak.

This matters geographically:

US and EU business hours land mostly in off-peak — you get the cheap rate by default
Asia-Pacific business hours land substantially in peak — you pay double

Cache pricing moved much harder than the headline rates. V4 Pro cache-hit went from $0.003625 to $0.022 off-peak — up to a 12x increase. Any architecture built around aggressive prompt caching on DeepSeek needs its cost model rebuilt, not adjusted.

Never quote a single flat DeepSeek price. Since August 16 that number does not exist.

What Is Temporary vs Permanent

This is the distinction that will cost teams money in Q4 2026 and Q1 2027.

Temporary — will revert:

GPT-5.6 Sol $4/$20 — reported as roughly three months
Gemini 3.7 Flash $0.75/$3.75 — reverts to $1.50/$7.50 on January 1, 2027, a 100% increase

Permanent — safe to build on:

Claude Sonnet 5 $2/$10 — Anthropic explicitly stated the scheduled September 1 increase to $3/$15 will not occur
OpenAI’s July 30 Terra and Luna cuts

The rule: budget at the post-promotional rate, spend at the promotional rate, keep the difference as margin. A team that priced its product on Gemini 3.7 Flash’s introductory rate is facing a doubling of that line on January 1 — a date that is knowable today.

What To Actually Do

Re-run your cost model. If you moved workloads down-tier for budget reasons, the frontier tier may now be affordable for some of them.
Tag every rate as standard or promotional in your cost tracking, with the revert date.
Check your DeepSeek cache assumptions if you use it — the cache economics changed by an order of magnitude, not a percentage.
Do not migrate on price alone. A 20% rate difference is dwarfed by output-token efficiency differences between models. Measure cost per completed task on your own workload before moving anything.