What is the cheapest AI API in August 2026 after DeepSeek's price increase?

DeepSeek V4-Flash off-peak remains cheapest at about $0.0099 per 30K-input/5K-output task, narrowly beating GPT-5.6 Luna at $0.0120. At DeepSeek's peak hours (01:00-04:00 and 06:00-10:00 UTC) V4-Flash costs $0.0198 and Luna wins. Gemini 3.7 Flash is roughly 4x either at $0.0410 but buys a 1M context window and a stronger benchmark profile.

Is GPT-5.6 Luna cheaper than DeepSeek now?

Only during DeepSeek's peak window. GPT-5.6 Luna costs a flat $0.20 per million input and $1.20 per million output tokens after OpenAI's 80% cut on July 30, 2026. DeepSeek V4-Flash is $0.22/$0.66 off-peak — cheaper on output, marginally more expensive on input — and $0.44/$1.32 at peak, where Luna wins outright.

Should I switch away from DeepSeek after the August 2026 price increase?

Usually no. Off-peak V4-Flash is still the cheapest option and 17 of every 24 hours are off-peak, with North American working hours falling entirely outside DeepSeek's peak window. Switch only if your workload is latency-bound to the 01:00-10:00 UTC window, or if you depended on DeepSeek's near-free cache-hit pricing, which rose up to 400% on Flash.

Quick Answer

V4-Flash vs Luna vs Gemini 3.7 Flash After the Hike

Published: August 17, 2026

The Short Answer

DeepSeek V4-Flash is still the cheapest — but only off-peak, and only barely.

Model	Input / Output (per MTok)	30K-in / 5K-out task	Context
DeepSeek V4-Flash (off-peak)	$0.22 / $0.66	$0.0099	1M
GPT-5.6 Luna	$0.20 / $1.20	$0.0120	Long
DeepSeek V4-Flash (peak)	$0.44 / $1.32	$0.0198	1M
Gemini 3.7 Flash	$0.75 / $3.75 (intro)	$0.0410	1M

Prices verified August 17, 2026. DeepSeek’s new rates took effect 16:00 UTC August 16, 2026. Gemini’s intro rate expires December 31, 2026 → $1.50/$7.50.

The margin between the top two is now 19%. Before August 16 it was roughly 2x. That is the real headline.

What Changed on August 16

DeepSeek V4-Flash went from a flat $0.14 / $0.28 to a peak/off-peak schedule of $0.22 / $0.66 and $0.44 / $1.32. Output pricing rose 136% to 371%. Cache-hit input — the mechanism that made DeepSeek structurally, not just nominally, cheap — went from $0.0028 to $0.007 off-peak and $0.014 at peak.

Peak is 01:00-04:00 and 06:00-10:00 UTC. Seven hours. The other seventeen are half price.

Where Each One Wins

DeepSeek V4-Flash wins on output-heavy work off-peak. Its $0.66 output is 45% cheaper than Luna’s $1.20, which matters enormously when your workload generates long responses — code generation, document drafting, agent traces. Input is marginally more expensive than Luna’s $0.20. It carries a 1M context and a 384K max output, both far beyond what a budget tier normally offers, plus native OpenAI Responses API and Anthropic-format endpoints.

GPT-5.6 Luna wins on predictability and peak hours. A flat $0.20/$1.20 with no clock to watch is worth real money in engineering time. After OpenAI’s 80% cut on July 30, 2026, Luna is the cheapest frontier-lab option and the one you pick when your traffic is bursty, your team is distributed across timezones, or your finance team wants a bill they can forecast. It also carries no China-hosting question, which is a hard blocker at many enterprises.

Gemini 3.7 Flash wins on capability per dollar, not raw price. At $0.75/$3.75 it costs roughly 4x the other two, but it scored 43.6% on FrontierCode 1.1 (up from 34.4% for 3.6 Flash) and 30.4% on AutomationBench (up from 17.0%) at its August 13, 2026 release. It carries a 1M context and a 65,536-token max output. If cheap models keep failing your task and you’re paying for retries, the “expensive” option is often cheaper.

The Hour-By-Hour Reality

This is the decision most teams get wrong. Peak hours in local time:

Your timezone	DeepSeek peak window
US Eastern (UTC-4)	21:00-00:00, 02:00-06:00
US Pacific (UTC-7)	18:00-21:00, 23:00-03:00
London (UTC+1)	02:00-05:00, 07:00-11:00
Central Europe (UTC+2)	03:00-06:00, 08:00-12:00
India (UTC+5:30)	06:30-09:30, 11:30-15:30

If you are in North America, the entire business day is off-peak. DeepSeek’s price increase, for you, is a much smaller event than the headlines suggest — you pay $0.22/$0.66, never $0.44/$1.32.

If you are in India or East Asia, most of your workday is peak. Your effective increase is the full 3-4x, and Luna becomes the better default.

If you are in Europe, you are split. Mornings are peak, afternoons are free. Schedule batch work after 12:00 local and you keep the off-peak rate.

The Cache Trap

If your application relies on a large stable system prompt or document set, cache-hit pricing dominates your bill, not headline rates.

Model	Cache-hit input	Discount vs base
DeepSeek V4-Flash (off-peak)	$0.007	~97%
DeepSeek V4-Flash (peak)	$0.014	~97%
DeepSeek V4-Flash (before Aug 16)	$0.0028	~98%
Gemini 3.7 Flash	Provider-dependent	—

The percentage discount barely moved. The absolute cost rose 150-400% because the base moved. Architectures tuned to DeepSeek’s near-free cache — long shared context, high hit rate, thin outputs — absorbed the largest increase of anyone. If that’s you, re-run the numbers before assuming you’re still ahead.

The Decision Framework

North American team, cost-sensitive, output-heavy → DeepSeek V4-Flash. Your workday is off-peak; you keep the cheapest rate on the board.
Distributed team, bursty traffic, want a forecastable bill → GPT-5.6 Luna. The 19% premium off-peak buys you a flat rate and no clock management.
Asia-Pacific team on business hours → GPT-5.6 Luna. You’d be paying DeepSeek’s peak rate most of the day.
Cheap models keep failing the task → Gemini 3.7 Flash. Retries at $0.01 cost more than one success at $0.04.
Enterprise with data-residency constraints → Luna or Gemini. Price is not the binding constraint here.
Heavy cache reuse → recompute before choosing. This is where the August 16 change actually bit.

Last verified: August 17, 2026. Prices from official vendor pricing pages.