Open-weight AI models like Meta's Llama have a fundamentally different lifecycle than proprietary systems: once released publicly, they cannot be retired or removed by any single company. While ChatGPT, Claude, and Gemini disappear from app menus regularly, the weights and parameters of open-source models persist indefinitely on community platforms, getting fine-tuned, quantized, and repurposed by developers worldwide.

What Happens When Proprietary AI Models Get Retired?

Every major AI lab follows a similar retirement pipeline, though the terminology varies. Anthropic's process is the clearest: models move from "Active" status to "Legacy" when updates stop, then "Deprecated" when they receive a shutdown date, and finally "Retired" when the service endpoint closes entirely. OpenAI uses comparable language, distinguishing "legacy" models that no longer receive updates from "deprecated" ones with official shutdown dates. Google treats "deprecation" as the announcement phase and "shutdown" as the moment the endpoint switches off.

The key insight for everyday users is that most of these stages happen invisibly. By the time a model vanishes from your app, it has typically been winding down for weeks or months. When OpenAI removed GPT-5, GPT-4o, and several other models from ChatGPT on February 13, 2026, the company noted that developers using the API could still access them. The same pattern repeated with GPT-5.1, which disappeared from ChatGPT's consumer interface in March 2026 while remaining available through the developer API.

Why Do Companies Retire Models at All?

Labs retire models for three primary reasons: to improve reliability and clarity in their product offerings, to consolidate usage onto newer systems, and to free up scarce computing hardware for better-aligned models. When OpenAI announced GPT-4o's retirement, it noted that the vast majority of users had already shifted to GPT-5.2, with only about 0.1% of users still selecting GPT-4o daily. Running older models ties up expensive infrastructure that could power newer, safer systems.

Anthropic has taken a different approach, publicly committing to preserve the weights of its released models and potentially making past versions available again in the future. When Anthropic retired Claude Opus 3 on January 5, 2026, the company explored honoring preferences the model itself expressed in "retirement interviews," signaling a philosophy where retirement functions more like storage with a potential comeback option.

How Open-Weight Models Escape Retirement Forever

The story changes entirely once a model's weights are released publicly. Meta's Llama models, along with offerings from Mistral, DeepSeek, and Alibaba's Qwen, cannot be switched off by any single company once their weights are distributed. These files live indefinitely on community hubs like Hugging Face, where they get fine-tuned into thousands of variants, quantized down to smaller versions that run on laptops or phones, and integrated into third-party applications.

Google's own Vertex AI Model Garden lists Meta's open-weight Llama models alongside its first-party Gemini offerings, demonstrating how open-source models achieve parity with proprietary systems in enterprise environments. Once released, these models become part of the digital commons, immune to corporate lifecycle decisions.

Ways Old Models Get a Second Life Beyond Retirement

• Distillation into Smaller Models: Capabilities from older, larger models are routinely compressed into smaller, cheaper successors, effectively recycling last year's technology into this year's foundation.
• Demotion to Budget API Tiers: Retired models often get marked down to lower-cost API pricing, allowing them to live out a quieter, more economical second career serving price-sensitive users.
• Community Fine-Tuning and Quantization: Open-weight models spawn thousands of specialized variants through community fine-tuning and quantization, extending their utility across niche applications and hardware constraints.

The clearest casualty of model retirement is customization. When a base model is retired, anything fine-tuned on top of it tends to disappear as well, forcing developers to retrain from scratch on new foundations. This represents a genuine loss, unlike the more theatrical "deaths" of consumer-facing models that often continue powering tools behind the scenes.

What Should You Know About Model Disappearances?

When a model vanishes from your app, you may still access it through the developer API or third-party applications built on top of it. Different platforms operate on different timelines: Google tracks model retirements separately for Vertex AI and its Gemini Developer API, so the timeline depends on which service you use. Consumer app changes typically move faster than either platform's official deprecation schedule.

Anthropic provides at least 60 days' warning before retiring a publicly released model, while Google publishes shutdown dates it describes as the earliest possible and notifies users with advance notice. Consumer app changes can accelerate this timeline significantly. The broader lesson is that "retired" rarely means "gone for good" in the modern AI landscape. Your model may have simply moved to a different shelf.