AI Model Deprecation Dates 2026 — The Complete EOL Timeline

GPT-4o is deprecated. Gemini 2.0 Flash shuts down June 1, 2026. Llama 3's 70B model was deprecated on major platforms in May 2025. The models powering production systems are disappearing faster than the engineering teams that depend on them. This is the complete, primary-source deprecation timeline for every major AI model — with exact shutdown dates, EOL Risk Scores, and migration paths.

Why AI model EOL is different from software EOL

When Node.js 18 reached end of life, your servers kept running. The code didn't break on April 30, 2025 — you simply stopped receiving security patches. You had time to plan, test, and migrate.

AI model deprecation is not like that. When an AI model is deprecated, the API endpoint stops responding. Every application that calls the model — production inference, RAG pipelines, fine-tuned deployments, automated workflows — returns an error simultaneously. There is no grace period after shutdown. The model is simply gone.

This creates a qualitatively different risk profile from traditional software EOL. The blast radius is immediate, the failure mode is hard (not graceful degradation), and the typical scanner that would catch a vulnerable Node.js version will never flag a deprecated model ID in a config file.

Three forces are making this worse in 2026:

Lifecycle compression. GPT-4's initial release-to-deprecation window was roughly 18 months. Gemini 1.0 Pro lasted about 16 months. Many 2025 and 2026 model snapshots are being deprecated in 6–9 months. The window to catch and respond is shrinking.

Version explosion. The number of model versions in production has multiplied. Each provider now releases multiple snapshots, previews, and fine-tuned variants — all with independent deprecation clocks. What looks like one integration is often five or six model IDs that each need to be monitored.

No standard tooling. Snyk, Dependabot, SBOM scanners, and CVE feeds don't track AI model lifecycle. There is no aimodel-advisories feed equivalent to the NVD. Engineers find out about deprecations the way they find out about most things: a forum post, a Slack message from someone else's production outage, or — worst — a 404 in the logs.

OpenAI API — Deprecation Timeline

OpenAI publishes an official API deprecations page. Unlike many providers, OpenAI typically gives 3–6 months' notice before removal. However, the ChatGPT product and the API are managed on separate schedules — a model removed from ChatGPT may remain available in the API for months, or vice versa.

→ Full OpenAI API deprecation schedule with EOL Risk Scores for all 20 tracked models

Google Gemini — Shutdown Schedule

Google publishes an official Gemini API deprecations page with exact shutdown dates. Unlike OpenAI's model-level versioning, Google uses snapshot dates as version identifiers (e.g. gemini-2.5-flash-preview-05-20). Preview models are shut down aggressively — often within 3–6 months of release. Stable (GA) models get longer windows but are still time-limited.

The critical near-term date: Gemini 2.0 Flash and Gemini 2.0 Flash Lite both shut down June 1, 2026 — any production deployment using these model IDs will begin failing that day.

→ Full Google Gemini deprecation schedule with EOL Risk Scores for all 13 tracked models

Meta Llama — Platform Deprecation Timeline

Meta Llama is open-source — Meta does not deprecate models in the same way API providers do. The models themselves remain available for download indefinitely. However, hosted inference providers deprecate Llama model endpoints on their own schedules, and those dates are what production teams need to track.

Groq deprecated Llama 3 8B and 70B endpoints on May 31, 2025. Llama 3.1 70B was deprecated on January 24, 2025. AWS Bedrock, Azure AI Studio, and Google Vertex all have independent retirement schedules for each Llama variant they host.

The strategic shift: Meta has pivoted from the Llama open-source model to a proprietary program. Llama 4 (Scout and Maverick) represents the current generation, but long-term Llama availability on hosted platforms depends on each provider's individual policies, not Meta's.

→ Full Meta Llama platform deprecation schedule with EOL Risk Scores for all 15 tracked models

Mistral AI — Retirement Dates

Mistral AI publishes model availability information but does not maintain a single consolidated deprecation page. Retirement dates vary by platform — the same model may be active on the Mistral API while already retired on Azure AI Studio or Google Vertex AI.

Key dates: Mistral Large 2402 was retired on Vertex AI in January 2025. Mistral Large 2407 was retired on Azure in April 2025. Codestral 2405 shuts down on Vertex AI in January 2026. Mistral Medium, the original commercial offering, was discontinued in mid-2025.

→ Full Mistral AI deprecation schedule with EOL Risk Scores for all 14 tracked models

Claude (Anthropic) — Deprecation Schedule

Anthropic publishes model deprecation dates on their official model deprecations page. Claude models follow a structured lifecycle: each snapshot version (e.g. claude-3-7-sonnet-20250219) has an announced EOL date, after which API calls must use a newer model ID.

Claude 3 Sonnet was EOL July 21, 2025. Claude 3.5 Sonnet (20240620) and Claude 3.5 Sonnet (20241022) both reached end of life October 28, 2025. Claude 3.7 Sonnet reached end of life February 19, 2026. The pattern is clear: Anthropic deprecates model snapshots on roughly 12-month cycles.

→ Full Claude deprecation schedule — 25 model versions tracked with EOL dates and Risk Scores

What to do when your model is deprecated

Audit before the shutdown date

Search your codebase for hardcoded model IDs. This includes config files, environment variables, prompt templates, and any string that looks like gpt-4o-2024-08-06 or gemini-2.0-flash. Grep is your friend. Set a calendar alert 90 days before the shutdown date, not the day of.

Use stable aliases where available

Most providers offer alias endpoints (e.g. gpt-4o without a snapshot date) that automatically point to the current recommended version. These auto-update on the provider's schedule. The trade-off: behavior can change unexpectedly when the alias is updated. For production inference where determinism matters, pin to a snapshot and manage the migration yourself. For prototypes, aliases are fine.

Test your replacement before the deadline

Replacement models are not drop-in compatible. Output format, token limits, pricing, latency, and behavior all differ. Run your evaluation suite against the replacement model 30–60 days before the shutdown. Do not migrate on the shutdown date.

Treat model deprecation as a dependency change

Add model IDs to your software bill of materials (SBOM). Track them in the same system you use for library dependencies. The endoflife.ai API provides programmatic access to deprecation dates for all tracked models — wire it into your CI pipeline so a failing check surfaces before your deployment does.