GPT-6 Luna Now Undercuts DeepSeek's Off-Peak Rates

Verdict: In 28 days OpenAI cut API prices twice and pushed its cheapest model onto free plans. The part almost nobody has written up: at $0.10/M input and $0.50/M output, GPT-6 Luna now undercuts DeepSeek's V4.1-Flash even during DeepSeek's own off-peak hours ($0.15/$0.60). For pay-as-you-go workloads, OpenAI's budget tier is cheaper than the discount king's discount price. For cached and long-context workloads, it isn't — and I'll show the exact split.

One honesty note up front: OpenAI's own announcement page blocks automated access from my network, so the Luna figures below come from OpenAI's community announcement and secondhand reporting, not from my own read of the pricing page. The DeepSeek numbers are pulled directly from DeepSeek's official pricing page today. I mark the difference.

The 28-day timeline

July 30: OpenAI cuts GPT-5.6 Luna's price by 80% — the first shot (official community post). September 22: GPT-6 Sol and Luna ship, with API prices set 50% below GPT-5.6's promotional rates (OpenAI; Yahoo Finance pins Luna at $0.10/$0.50). October 8: Luna rolls out to Free and $8 Go users in the ChatGPT apps, replacing GPT-5.6 as the default (Notebookcheck).

Read as one line, that's not a discount schedule. That's a siege. Cut 80%, then cut another 50% with a new generation, then give it away free — all inside one quarter. My read: OpenAI is not defending a price point, it's destroying the "frontier models are expensive" assumption that open-weight providers have been selling on.

The math against DeepSeek

DeepSeek's official pricing page (verified today): deepseek-flash, i.e. V4.1-Flash, costs $0.30/M input and $1.20/M output at peak, $0.15 and $0.60 off-peak, with cache hits as low as $0.003/M input. Off-peak windows are 01:00–04:00 and 06:00–10:00 UTC, weekdays — a real discount, but also a real constraint: your batch jobs must run at night, UTC time, or you schedule around it.

Now put Luna next to it. $0.10 input is 33% below DeepSeek's off-peak input price. $0.50 output is 17% below off-peak output. No time windows, no scheduling, no cache engineering required. For a workload doing 50M output tokens a month, you'd pay $25 on Luna versus $30 on DeepSeek at its cheapest hours — and $60 at peak.

Where DeepSeek still wins, honestly: cache-hit input at $0.003/M is roughly 33x cheaper than anything Luna advertises, so agent loops and RAG pipelines that re-send long context stay DeepSeek's home turf. DeepSeek also runs a 1M-token context window, and if you're processing very long documents, that matters more than $0.05 per million. And off-peak batch remains the right answer when you can actually schedule it — $0.60 output is still $0.60, just slower and at night.

What I'd actually do

If your workload is stateless short-context calls — classification, extraction, cheap drafts — try Luna this week. The pricing advantage is real and there's no scheduling tax. If your workload is cache-heavy or needs a million tokens of context, DeepSeek's economics still hold; this price cut doesn't touch that moat yet. What I'd watch next: DeepSeek's last big price moves came hard on the heels of competitor moves. A third response is a "when", not an "if" — and when it lands, I'd expect the off-peak gimmick to be the thing that gets killed, because time-window pricing looks increasingly like a liability in this fight.

Full disclosure: this site covers DeepSeek closely and we maintain our own DeepSeek cost calculator, so we're invested in the underdog narrative. The math above doesn't care.

Sources: DeepSeek official pricing (2026-10-14); OpenAI announcement; OpenAI community, 2026-07-30; Yahoo Finance, 2026-09-23; Notebookcheck, 2026-10-08. No affiliate links.