GPT-5 deprecation on 11 December: gpt-5 and o3 to GPT-5.6

The GPT-5 deprecation removes six models from the OpenAI API on 11 December 2026, and every non-pro replacement costs more per token than the model it replaces. The steepest move, gpt-5-mini to gpt-5.6-terra, is 8x on input. GPT-5.6 Sol's current price is only promised through 21 November, three weeks before the cutoff. Here is the mapping, the price change on each row, and a plan for the 81 days left.

• Shutdown: 11 December 2026, announced to affected developers on 11 June 2026.
• Models: gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3 and o3-pro snapshots.
• Replacements: GPT-5.6 Sol, Terra and Luna, with reasoning.mode: pro for the two pro models.
• Price watch: Sol's $4 / $20 promotional pricing runs "at least through November 21, 2026".

The gpt-5 deprecation table, as OpenAI published it

Shutdown date	Model	Recommended replacement
11 Dec 2026	gpt-5-2025-08-07	gpt-5.6-sol
11 Dec 2026	gpt-5-mini-2025-08-07	gpt-5.6-terra
11 Dec 2026	gpt-5-nano-2025-08-07	gpt-5.6-luna
11 Dec 2026	gpt-5-pro-2025-10-06	gpt-5.6-sol (reasoning.mode: pro)
11 Dec 2026	o3-2025-04-16	gpt-5.6-sol
11 Dec 2026	o3-pro-2025-06-10	gpt-5.6-sol (reasoning.mode: pro)

The table names dated snapshots, but calling the bare alias does not protect you. The GPT-5 model page lists gpt-5-2025-08-07 as the default snapshot behind gpt-5, and the o3 page lists o3-2025-04-16 behind o3. Six months' notice is exactly OpenAI's stated minimum for generally available models, and the deprecations page says that, in some cases, dedicated capacity can be arranged through sales for access after shutdown.

Is GPT-5.6 more expensive than GPT-5? Per token, yes, except for pro

These are short-context Standard rates per 1M tokens from the pricing page, as of 21 September 2026:

Move	Old input / output	New input / output	Change
gpt-5 to Sol	$1.25 / $10.00	$4.00 / $20.00	3.2x / 2x
gpt-5-mini to Terra	$0.25 / $2.00	$2.00 / $12.00	8x / 6x
gpt-5-nano to Luna	$0.05 / $0.40	$0.20 / $1.20	4x / 3x
o3 to Sol	$2.00 / $8.00	$4.00 / $20.00	2x / 2.5x
gpt-5-pro to Sol pro mode	$15.00 / $120.00	$4.00 / $20.00 per token, more tokens	Measure
o3-pro to Sol pro mode	$20.00 / $80.00	$4.00 / $20.00 per token, more tokens	Measure

The pro rows run the other way: Sol's per-token price is far below gpt-5-pro and o3-pro. Per-token price is not per-task price, though. OpenAI's GPT-5.6 guide says the family "reaches flagship-level performance with fewer output tokens", and pro mode bills its aggregated work at the model's standard rates, so a pro-mode request can cost more or less than a gpt-5-pro call depending on how much work it does. That is OpenAI's claim about its own model; run your prompts and read usage before you believe it for your workload.

Two more costs are new. GPT-5.6 bills cache writes at 1.25x the uncached input rate, which GPT-5 did not, and prompts over 272K input tokens are priced at 2x input and 1.5x output for the whole request, per the Sol model page.

What replaces gpt-5-nano and gpt-5-mini, if not the recommended model?

The recommendation for gpt-5-mini is Terra, but Luna, at $0.20 / $1.20, is cheaper than gpt-5-mini on both input and output. OpenAI describes Luna as the nano-like tier and Terra as the mini-like tier, so Luna is a step down in class. For extraction, classification and routing calls it is still worth an eval before you accept a 6x output price. One caution: the Luna model page lists a 1,050,000-token context window, while OpenAI's Sol upgrade guide says Luna has 400K. Test the size you need.

What should I replace o3 with? Sol, and pro becomes a mode

OpenAI maps both o3 and gpt-5 to gpt-5.6-sol. For o3-pro and gpt-5-pro there is no separate model ID any more: you keep gpt-5.6-sol and set reasoning.mode to pro. The reasoning guide says the setting works in the Responses API, is independent of reasoning.effort, and defaults effort to medium in both modes. If your pro calls still go through Chat Completions, the endpoint move is part of this migration.

Five GPT-5.6 behaviour changes that break a string swap

OpenAI's own upgrade guide opens with "do not perform a blind model-string replacement". These are the changes we would check first, all from OpenAI's GPT-5.6 documentation:

1. Reasoning effort values. GPT-5 accepts minimal, low, medium and high. GPT-5.6 accepts none, low, medium, high, xhigh and max. Anything sending minimal needs a new value.
2. Chat Completions with function tools. On GPT-5.6 these work only with effective reasoning none. Because GPT-5.6 defaults to medium, a tool-calling Chat Completions route will break unless you set it or move to Responses.
3. Persisted reasoning. GPT-5.6 defaults reasoning.context to all_turns; earlier models default to current_turn. Multi-turn token usage can change.
4. Prompt caching options. Replace prompt_cache_retention with prompt_cache_options.ttl, and track cache_write_tokens now that writes are billed.
5. Safeguard pauses. GPT-5.6 runs cyber and biology classifiers during generation, and some requests pause "for several seconds mid-stream". Streaming timeouts tuned for GPT-5 may fire.

Also note that each GPT-5.6 model page lists a single snapshot, the model ID itself, with no dated version to pin. If your change control relies on dated snapshots, record response.model during testing, as the upgrade guide advises.

A test plan for the 81 days before the gpt-5 shutdown

1. This week: inventory every place a GPT-5 or o3 ID appears, including config, fallbacks, model pickers and pricing tables in your own code.
2. By late October: run your eval set against the recommended target and one alternative per route, such as Luna for a gpt-5-mini route, and compare cost per task, not per token.
3. Before 21 November: make the model decision while Sol's promotional price is still guaranteed, so a price change does not arrive mid-cutover.
4. Early December: shift traffic behind a flag, keeping the old model as fallback until it disappears on 11 December.

If you are also clearing the 23 October wave, our GPT-4 shutdown migration map covers it, and our GPT-6 Astra vs GPT-5.6 comparison helps if Sol may not be the right tier. Where caching is a large share of your bill, our note on prompt caching covers the cache-write change. Creuto's AI engineering team runs these migrations as a scoped piece of work. Whoever does it, start the eval runs in October, not in the first week of December.