GPT-4o Is Being Retired: Migrating to GPT-6 Luna Cuts Your Bill ~95%

On October 14, 2026, OpenAI deprecated the entire GPT-4 series on the API: GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini. Shutdown date: January 14, 2027. If you still have traffic pointed at any of those model IDs, you have three months to move it — and the move OpenAI recommends is, for once, dramatically cheaper than staying put.

The direct answer: OpenAI's official recommended replacement for every one of those models is gpt-6-luna — $0.10 input / $0.50 output per million tokens. Against GPT-4o's last published $2.50/$10.00, that's a 25× input and 20× output price cut. A workload spending $325/month on GPT-4o costs about $15/month on Luna.

What's actually being shut down, and when

Two deprecation waves matter here, both straight from OpenAI's deprecations page:

The Oct 23 wave is the urgent one — reasoning models like o1 and o3-mini go dark in a week. But the interesting one is January: it's the first time OpenAI has deprecated a mainstream, high-traffic model family and pointed everyone at the cheapest tier they sell.

The migration math: $325 → $15

Take a realistic production workload: 50 million input tokens + 20 million output tokens per month, uncached, standard processing. No caching tricks, no Batch — plain synchronous calls.

Yes, you read the chart right — the Luna bar is a sliver. Migration this clean almost never happens. Usually deprecation means rewriting prompts and hoping the replacement is only slightly worse; here the replacement costs 4.6% of the old bill.

Prices verified Oct 16, 2026 — always confirm on the official pricing page. (GPT-4o rates are its last published prices, pre-deprecation; they're going away, not going up.)

Caching is where it gets silly

If your workload benefits from prompt caching — and if you're still on GPT-4o, your prompts are probably stable enough to cache — the gap widens. GPT-4o cached input was $1.25/M; Luna cached input is $0.01/M. That's a 125× difference on every cached token.

Same 50M-input workload, but with 80% of input cached:

• GPT-4o: 10M × $2.50 + 40M × $1.25 = $25 + $50 = $75.00 (input side)
• GPT-6 Luna: 10M × $0.10 + 40M × $0.01 = $1 + $0.40 = $1.40 (input side)

At this point the input bill is basically a rounding error. I've run cost comparisons for a year and this is the largest same-migration discount I've seen from any provider — even Haiku 5.5's price war only matched Luna, it didn't beat it by a factor of twenty.

What to watch during the migration

The money is the easy part. Three things actually bite:

• Output length drift. Luna is a smaller model; if your prompts were tuned on 4o, spot-check that you don't get longer, chattier completions. Output tokens are 5× the input rate, so verbosity shows up on the bill.
• The 272K-token tier. Luna holds $0.10/$0.50 until 272,000 input tokens per request, then steps to $0.20/$0.75. GPT-4o had a 128K context ceiling — if you're feeding longer prompts than 4o ever allowed, price the upper tier.
• Fine-tunes die too. The deprecation page is explicit: fine-tuned versions of GPT-4o/GPT-4o-mini follow the same January 14, 2027 shutdown. There's no Luna fine-tuning announced as a drop-in; plan a distillation or prompt-based fallback.

The October 23 deadline is the real fire drill

If your codebase still references o1, o3-mini, or gpt-4-turbo, those endpoints die October 23, 2026 — seven days from this post. OpenAI announced these back in April, but legacy strings have a way of surviving in production configs. Check GET /v1/models: any model with a shutdown_date field populated is on the clock.

Sources

• OpenAI API Deprecations — 2026-10-14 GPT-4 series deprecation, Oct 23 and Dec 11 shutdown schedules (official)
• GPT-6 Luna API docs — $0.10/$0.50 pricing, 272K tier threshold, cache rates (official)

Updated Oct 16, 2026. This is an independent analysis, not affiliated with OpenAI.