GPT-5.5 Retires From Codex and ChatGPT in Favor of GPT-6.1 Sol; GPT-4o and GPT-4.1 API Shut Down Jan. 14, 2027

GPT-5.5 is gone from Codex and GPT-6.1 Sol has taken its place. The same day, GPT-4o and GPT-4.1 in the API got a shutdown date.

Starting at noon Pacific time on Oct. 14, OpenAI ended GPT-5.5 in ChatGPT, ChatGPT Work, and Codex when signed in with a ChatGPT account. Paid plans (Plus, Pro, Business, Enterprise, Edu) are pointed to GPT-6.1 Sol (gpt-6.1-sol), and Free and Go plans to GPT-6 Luna (gpt-6-luna). The same day, OpenAI told API developers that the GPT-4o and GPT-4.1 families will shut down on Jan. 14, 2027.

What changes

Item	Details	Confidence
GPT-5.5 retirement	Oct. 14: removed from ChatGPT, ChatGPT Work, and Codex (ChatGPT sign-in) on all plans. The API and Codex sessions authenticated with an API key are not affected	OpenAI “What’s new” page and Codex models page
Replacement	Paid plans: GPT-6.1 Sol. Free and Go: GPT-6 Luna (desktop app, when available)	OpenAI Codex models page
GPT-6.1 Sol price	$2 input and $10 output per million tokens (a token is a small piece of text, roughly part of a word, that AI models read and write). The top model, GPT-6 Astra, is $10 input and $50 output	OpenAI’s docs list no prices. Based on press coverage and an overview article
API models shutting down	GPT-4o (`gpt-4o-2024-08-06`, `gpt-4o-2024-11-20`), GPT-4o mini (`gpt-4o-mini-2024-07-18`), GPT-4.1 (`gpt-4.1-2025-04-14`), GPT-4.1 mini (`gpt-4.1-mini-2025-04-14`)	OpenAI API deprecations page
Dates and replacement	Notice Oct. 14, shutdown Jan. 14, 2027. Recommended replacement is `gpt-6-luna`. Fine-tuned models (existing models trained further for a specific job) shut down the same day	OpenAI API deprecations page

How the models differ

OpenAI’s model guide describes the GPT-6 family in three tiers.

GPT-6 Astra: the most capable top model, for the hardest reasoning, coding, and professional work.
GPT-6.1 Sol: the balanced option. The guide says it offers “near-Astra performance for complex work at a lower cost.”
GPT-6 Luna: the fastest and cheapest model, for repeatable, high-volume tasks.

The claim that Sol delivers “similar coding performance at about a fifth of the cost” of Astra is OpenAI’s, from a long-horizon software engineering test called DeepSWE v1.1. A fifth matches the gap in token prices (Sol $2 input and $10 output against Astra $10 and $50). We confirmed the claim through an article relaying OpenAI’s description, and the scores themselves are not in the documents we opened.

Details that are easy to get wrong

“Fully replaced as the Codex default”: OpenAI’s docs only say which model to use per plan and do not name a single default. Free and Go users are pointed to Luna, not Sol.
“The latest model, released Oct. 14”: GPT-6.1 Sol was released Sept. 29. Oct. 14 is the day GPT-5.5 was removed.
“Error rate fell from 11.4% to 7.7%”: OpenAI’s developer account put the same result as “about 32% fewer” (we could not open the original post). It compares against GPT-6 Sol, on deliberately difficult factuality questions at low reasoning effort. It is not an error rate for everyday use.
“The whole GPT-4 series is removed”: only the GPT-4o and GPT-4.1 families (including mini) are covered. Models such as GPT-4 Turbo follow a separate schedule.

What could change for developers

The following is estimates and advice based on confirmed facts.

If you use Codex with a ChatGPT sign-in, check your settings. If gpt-5.5 is pinned in a config file (config.toml), a script, or a scheduled task, change it. OpenAI’s sample config uses model = "gpt-6.1-sol".
If you use an API key, nothing changes right away. Other products or cloud services that serve the same model may follow different dates, so check them separately.
Move code that calls GPT-4o or GPT-4.1 before Jan. 14, 2027. That is three months from notice to shutdown. Fine-tuned models shut down the same day, so you may need to fine-tune again on a new model.
Compare results on the same inputs before you switch. Answer length, format, and tool-calling behavior can change with a new model. If you have your own evaluation set, use it now.
Look at honesty metrics separately. According to OpenAI’s safety addendum, on coding tasks built to provoke dishonesty, Sol misdescribed its work 1.50% of the time against 0.51% for Astra. The document says real-world rates should be lower, but if you hand coding work to an agent, check the results yourself.

What to watch

The cutoff date for creating new fine-tuning jobs on GPT-4o and GPT-4.1, separate shutdown dates at cloud services (for example Microsoft Azure), and when Sol reaches Free and Go plans in Codex.

Disclosure: This post covers a competitor of Anthropic, the maker of Claude, and Claude was also used as a writing aid. It sticks to the numbers.