GPT-5 deprecation December 2026: gpt-5, mini and o3 end 11 Dec, 30-day plan

Summary. OpenAI stops serving six models in its API on 11 December 2026: gpt-5 (gpt-5-2025-08-07), gpt-5-mini (gpt-5-mini-2025-08-07), gpt-5-nano (gpt-5-nano-2025-08-07), gpt-5-pro (gpt-5-pro-2025-10-06), o3 (o3-2025-04-16) and o3-pro (o3-pro-2025-06-10). Fine-tuned versions go the same day. The notice, dated 11 June 2026, names GPT-5.6 replacements: gpt-5.6-sol for gpt-5 and o3, gpt-5.6-terra for gpt-5-mini, gpt-5.6-luna for gpt-5-nano, and gpt-5.6-sol in pro reasoning mode for both Pro models. Those picks cost more. gpt-5 lists at $1.25 input and $10 output per million tokens; GPT-5.6 Sol is $4 and $20 on a promotion that ends on 30 November 2026, eleven days before the cut-off, and OpenAI has not published the rate after that. gpt-5-mini at $0.25 and $2 maps to Terra at $2 and $12, eight times the input price. The GPT-6 family is often cheaper: GPT-6.1 Sol costs $2 and $10, GPT-6 Luna $0.10 and $0.50. On Vals AI's Terminal-Bench 4.0 run, GPT-6.1 Sol scored 55.05% at $1.72 per task against 37.88% at $7.98 for GPT-5.6 Sol. Test OpenAI's named pick and one GPT-6 model per workload, set reasoning effort on every call, check rate limits, and finish the rollout by 4 December.

About this article. For engineering leads and developers who call OpenAI models directly or through Azure. We checked every date, model ID, price, rate limit and parameter against OpenAI's deprecations page, model pages, pricing page, guides and changelog, Microsoft's Foundry retirement schedule and Vals AI's published leaderboard on 11 November 2026. Prices are US dollar list prices and change often, so re-check before you commit; no vendor paid for inclusion. Researched with AI assistance and reviewed by our editorial team.

What OpenAI shuts down on 11 December 2026

OpenAI's deprecations page lists the notice under 11 June 2026. It says OpenAI told developers using the older GPT-5 and o3 snapshots that day that the models leave the API on 11 December 2026, which meets the minimum six months' notice its policy sets for generally available models. All six model pages now show a Deprecated status. The table draws on those pages and the pricing page.

Four details matter for planning:

• Alias and snapshot go together. Code that calls gpt-5 resolves to gpt-5-2025-08-07, the only snapshot the page lists, so both strings fail on the same day.

• OpenAI's own pages point different ways. The deprecation table names gpt-5.6-sol for gpt-5, but the GPT-5 model page now recommends GPT-6 Astra, which lists at $10 input and $50 output. The gpt-5-mini and gpt-5-nano pages point to GPT-5.6 Terra and Luna, and the o3 page still names GPT-5, which retires the same day, as its successor. Treat the named replacement as a safe default, not the only option.

• You can check dates in code. The List models and Retrieve model endpoints return an optional shutdown_date field as a YYYY-MM-DD string, or null when no date is set. A nightly job that alerts on any model you call with a date set catches the next notice too.

• Dedicated capacity is a sales conversation. OpenAI says some developers may be able to provision dedicated capacity to keep using a model after its shutdown date, through its sales team. Treat that as an exception path for contract-bound workloads, not a plan.

Later removals follow: our guide to the GPT-4o and GPT-4.1 shutdown on 14 January 2027 covers the next large one, and the GPT-5.1 and GPT-5.3-Codex removal on 1 April 2027 the one after.

The replacement problem: GPT-5.6 is a generation old and costs more

OpenAI released the GPT-5.6 family on 9 July 2026. Since then it has released GPT-6 Astra (3 September), GPT-6 Sol and GPT-6 Luna (22 September), GPT-6.1 Sol (29 September) and GPT-6.1 Terra (26 October). Its later notices name GPT-6 targets: gpt-6-luna for GPT-4o and GPT-4.1, gpt-6-sol for GPT-5.1. Three facts change the decision.

First, GPT-5.6 Sol's price is promotional. Its model page lists $4 input, $0.40 cached and $20 output, describes them as 20% below the previous input price and 33% below the previous output price, and says the promotion runs until 30 November 2026. The changelog entry for the cut, dated 21 August 2026, gave an end of at least 21 November; the model and pricing pages now say 30 November. OpenAI has not published the rate from 1 December. Reversing the stated cuts gives $5 and $30 ($4 ÷ 0.8 and $20 ÷ 0.667), so budget for that until OpenAI says otherwise.

Second, the GPT-6 models are cheaper than their GPT-5.6 equivalents. The table compares each retiring model with OpenAI's pick and with the GPT-6 model we would test first, using standard prices per million tokens from OpenAI's pricing page.

Cached input also moves. GPT-6.1 Sol reads cache at $0.10, below gpt-5's $0.125 and o3's $0.50. But GPT-5.6 and later models bill cache writes at 1.25 times the input rate for the default 30-minute lifetime, or 2.5 times for a 12-hour lifetime in the Responses API, where gpt-5 and o3 charged nothing extra for writes. GPT-6 Sol, at $2 input, $0.20 cached and $10 output, sits beside 6.1 Sol and still supports none effort.

Third, independent results favour the GPT-6 models. Vals AI's Terminal-Bench 4.0 leaderboard, updated 9 November 2026, runs every model in the same minimal harness that gives the model a single bash tool. It shows GPT-6.1 Sol at 55.05% for $1.72 per task, GPT-6 Sol at 44.44% for $5.79 and GPT-5.6 Sol at 37.88% for $7.98. GPT-6.1 Terra scored 49.49% ($2.16) against 22.73% ($5.60) for GPT-5.6 Terra, and GPT-6 Luna 13.64% ($0.35) against 11.62% ($0.73) for GPT-5.6 Luna. Terminal-Bench measures long agentic terminal work, so it says little about extraction or classification. It does suggest that following the deprecation table to the letter buys an older model at a higher price.

Behaviour differences that break migrations

Changing the model string is the easy part. These differences cause silent cost increases, 400 errors or changed outputs. Sources: the model pages, OpenAI's reasoning, GPT-6 and prompt caching guides.

Three of these deserve a test case each. The rate-limit drop is easy to miss: gpt-5-mini allows 180 million tokens per minute at the Grow tier, while GPT-5.6 Terra and GPT-6.1 Terra allow 40 million. A batch pipeline that ran comfortably on mini can start returning 429 errors on Terra with no code change.

The pro-mode billing changes shape. OpenAI's reasoning guide says pro mode adds up the model work behind the final answer, bills those tokens at the chosen model's standard rates and does more work than standard mode, while the old Pro model IDs keep their own pricing. The per-token rate falls from $15 and $120 (gpt-5-pro) to $4 and $20 or $2 and $10, but the token count rises by an amount OpenAI does not publish.

The tool list and instruction handling change. GPT-5.6 and GPT-6 models add hosted shell, apply_patch, skills, computer use and tool search to the Responses API, so rerun tool-selection tests with the full production tool list. OpenAI's GPT-6 guide says GPT-6 Astra is more likely to stop and ask a clarifying question and can be more sensitive to instructions in skills and files such as AGENTS.md, and asks you to test its prompts on whichever GPT-6 model you choose. It also says to remove temperature, top_p and top_logprobs when effort is not none. An unattended job that pauses for a question is a failed job.

A minimal request change for a gpt-5 workload moving to GPT-6.1 Sol looks like this:

            from openai import OpenAI
client = OpenAI()

resp = client.responses.create(
    model="gpt-6.1-sol",            # was "gpt-5"
    reasoning={"effort": "low"},    # 6.1 Sol has no "none" or "minimal"
    input=prompt,                   # drop prompt_cache_retention; 30-minute caching is the default
)
          

For gpt-5-pro or o3-pro, keep the Responses API and add mode set to pro inside reasoning on the target model.

Decision table: pick a target per workload

Worked cost examples

These scenarios use list prices and hold token counts constant. Newer models often use different numbers of reasoning tokens for the same job, so replace our volumes with your own measured ones.

Scenario A: gpt-5 support agent. 1,000 million input tokens a month, 60% served from cache (600 million), 400 million uncached, 100 million output tokens. For the newer models we show a range: uncached tokens billed at the input rate (low) or all as 30-minute cache writes (high). Rates come from OpenAI's pricing page.

Following OpenAI's mapping more than doubles this bill even at the promotional price ($3,840 ÷ $1,575 = 2.4x). GPT-6.1 Sol adds 18% to 31%, and its lower token use per task on the Vals agent run (2.86 million against 10.49 million for GPT-5.6 Sol) suggests the real gap may be smaller.

Scenario B: gpt-5-mini extraction pipeline. 3,000 million input and 300 million output tokens a month, no caching.

The named replacement turns a $1,350 line into $9,600. If Luna at none effort passes your extraction tests, the bill falls by two thirds instead.

Scenario C: gpt-5-nano classifier. 5,000 million input and 250 million output tokens. Today: 5,000 × $0.05 + 250 × $0.40 = $350. GPT-5.6 Luna: 5,000 × $0.20 + 250 × $1.20 = $1,300. GPT-6 Luna: 5,000 × $0.10 + 250 × $0.50 = $625. OpenAI prices Batch at 50% of Standard for GPT-6 Luna, so a job that can wait for asynchronous results costs $312.50, just under today's Standard bill (gpt-5-nano on Batch would be $175).

Pro workloads. An o3-pro job using 20 million input and 10 million output tokens a month costs 20 × $20 + 10 × $80 = $1,200. On GPT-6.1 Sol in pro mode, the same token count would cost 20 × $2 + 10 × $10 = $140. Pro mode will not use the same token count, so measure the multiplier on 50 real prompts before you commit to a number.

The 30-day migration plan: 11 November to 11 December

There are 30 days from 11 November to the shutdown. Leave the last week as buffer: the end of the GPT-5.6 Sol promotion on 30 November and the doubling of Gemini 4 Carbon's price on 1 December, from $2 and $10 to $4 and $20, land in the same window and may prompt late re-routing. Our note on the Gemini 4 Carbon price increase has the detail.

If you planned to run the golden sets on OpenAI's hosted Evals platform, it became read-only on 31 October and shuts down on 30 November 2026, the same day as Agent Builder and reusable prompt objects; OpenAI points to Promptfoo.

Test plan: what to measure

A migration test should answer one question: does the new model produce acceptable output at an acceptable cost per successful result? Measure these on every golden set.

1. Task success. Exact-match or rubric score for extraction and classification; tests passed for code; human-graded samples for open text. Set a regression threshold in advance, for example no more than two points below baseline.

1. Schema and tool accuracy. Structured-output validity, correct tool chosen, valid arguments and no unrequested tool calls, run with the full production tool list.

1. Tokens per request, split four ways. Uncached input, cached input, cache writes and output including reasoning. OpenAI reports reasoning tokens under output_tokens_details in the usage object.

1. Latency. p50 and p95 time to first token and to completion, at the effort level you plan to ship.

1. Cost per successful task. Total cost divided by passing outputs. A cheaper model that fails more often can cost more.

1. Behaviour checks. Clarifying-question rate in unattended jobs, refusal rate, and answers that depend on dates, since knowledge cutoffs move by up to two years.

1. Load. A burst test at peak traffic against the target model's tokens-per-minute limit for your tier.

Keep the golden sets after the move. The GPT-5.1 removal on 1 April 2027 and any future GPT-5.6 notice will need the same tests.

Azure, Bedrock and data residency notes

Azure dates differ, and o3 goes first. Microsoft Foundry's model retirement schedule lists o3 (2025-04-16) and o3-pro (2025-06-10) retiring on 19 November 2026, 22 days before OpenAI's date, with gpt-5.6-sol as the replacement. gpt-5 and gpt-5-nano (2025-08-07) retire on 9 February 2027, gpt-5-pro on 7 April 2027 and gpt-5-mini on 13 July 2027, all with no replacement named. GPT-5.6 Sol, Terra and Luna run on Azure until 11 January 2028. Teams that fail over between OpenAI and Azure need both calendars; our Azure OpenAI GPT-5 retirement checklist covers the Azure side.

Bedrock. OpenAI's model pages list Bedrock IDs such as openai.gpt-5.6-sol and openai.gpt-6.1-sol. OpenAI's deprecations page covers its own API, so confirm Bedrock dates in AWS's documentation rather than assuming OpenAI's.

Data residency. OpenAI adds 10% to regional processing (data residency) endpoints for models released on or after 5 March 2026, which covers every GPT-5.6 and GPT-6 target and none of the retiring models. A team that used gpt-5 through a regional endpoint pays the new rate plus 10%. Fast mode supports EU data residency for GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna but not for GPT-6 Sol. For Indian teams handling personal data, check where inference runs, not only where data is stored.

How eCorpIT can help

eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. For a model migration, our AI evals and observability service builds the golden sets, runs candidate models side by side and reports cost per successful task before you switch. Talk to us at /contact-us/.

Last updated: 11 November 2026.

References

1. Deprecations — OpenAI API

1. GPT-5.6 Sol model page — OpenAI API

1. Pricing — OpenAI API

1. Terminal-Bench 4.0 Leaderboard and Methodology, updated 9 November 2026 — Vals AI

1. GPT-5 model page — OpenAI API

1. GPT-5 Mini model page — OpenAI API

1. GPT-5 nano model page — OpenAI API

1. GPT-5 Pro model page — OpenAI API

1. o3 model page — OpenAI API

1. o3-pro model page — OpenAI API

1. GPT-6 Astra model page — OpenAI API

1. Changelog — OpenAI API

1. GPT-6.1 Sol model page — OpenAI API

1. Prompt caching — OpenAI API

1. GPT-6 Sol model page — OpenAI API

1. Reasoning models — OpenAI API

1. Using GPT-6 — OpenAI API

1. GPT-6.1 Terra model page — OpenAI API

1. GPT-6 Luna model page — OpenAI API

1. GPT-5.6 Terra model page — OpenAI API

1. Gemini Developer API pricing — Google AI Studio

1. Model retirement schedule, Microsoft Foundry — Microsoft Learn