Model Deprecations Across Anthropic, OpenAI, and Google
Compare model lifecycle policies, deprecation notice periods, parameter sunsets, and migration timelines across Anthropic, OpenAI, and Google APIs.
- OpenAI provides at least 6 months of notice for GA models, 3 months for specialized variants, and 2 weeks for preview snapshots.
- Anthropic commits to at least 60 days of advance notice for publicly released models before retirement.
- Google routes deprecated models like gemini-3.7-flash and gemini-3.5-flash automatically to newer replacements in the API.
- Both Anthropic and Google have begun deprecating sampling controls like temperature, top_p, and top_k on newer model versions.
Lifecycle Stages and Retirement Terminology
Every major platform manages model lifecycles to decommission older systems and allocate infrastructure capacity for new releases. However, each provider defines the progression from launch to complete retirement differently. Anthropic categorizes models under four defined states: Active (fully supported and recommended), Legacy (still functioning but frozen without updates), Deprecated (functional but no longer advised, paired with a sunset schedule), and Retired (inaccessible, causing incoming API requests to fail). OpenAI marks models as either legacy (unsupported endpoints signaling upcoming migration) or deprecated (formally queued for shutdown with an announced termination date), using the words sunset and shut down interchangeably once access ceases.
Google operates through changelog announcements, transitioning models into deprecated states upon the launch of superior variants. While Anthropic and OpenAI leave deprecated endpoints operational until a hard terminal date, Google has utilized transparent backend routing for specific endpoints. For example, API calls targeting gemini-3.7-flash automatically route to gemini-3.8-flash, and requests directed to gemini-3.5-flash route directly to gemini-3.6-flash. For cost forecasting when switching between deprecated tiers and replacement architectures, consult the LLM API Pricing table and the cost tier selection guide.
Advance Notice Policies and Timeline Commitments
Notice windows differ considerably between vendors, directly impacting how fast development teams must test, validate, and cut over client software. OpenAI outlines formal notice tiers based on model stability. Anthropic establishes a single baseline window across public releases, whereas Google communicates specific retirement timelines alongside individual feature updates.
| Provider | Model Category | Minimum Advance Notice | Post-Sunset Invocations |
|---|---|---|---|
| OpenAI | Generally Available (GA) | At least 6 months | Inaccessible (Fails or contact sales for dedicated capacity) |
| OpenAI | Specialized Variants (Codex, Chat, Deep Research) | At least 3 months | Inaccessible (Fails) |
| OpenAI | Preview Models (*-preview) | As short as 2 weeks | Inaccessible (Fails) |
| Anthropic | Publicly Released Models | At least 60 days | Inaccessible (Returns failure error) |
| Flash / Production Models | Varies (Immediate rerouting or fixed sunset dates) | Routed to replacement or returns failure | |
| Experimental / Preview Managed Agents | Varies (Roughly 2 to 3 weeks in documented cases) | Inaccessible (Fails) |
OpenAI allows developers who need legacy access past an official shutdown date to explore provisioning dedicated capacity through their enterprise sales team. In contrast, Anthropic strictly terminates public endpoint accessibility on the retirement date to preserve cluster serving capacity, though it commits to the long-term preservation of underlying model weights. Anthropic dates apply strictly to its first-party Claude API, Claude Platform on AWS, and Microsoft Foundry; external partner platforms like Amazon Bedrock and Google Cloud Vertex AI publish their own distinct retirement schedules.
Tracking Deprecations Programmatically and via Audits
Preventing unexpected production outages requires automated detection or systematic administrative audits before models shut down completely. OpenAI supports direct programmatic querying via its model metadata endpoints. By issuing a GET request to /v1/models or retrieving a specific model object, developers can inspect the optional shutdown_date property. If a model has a scheduled termination, this key contains an ISO date string (YYYY-MM-DD); otherwise, it returns null.
{
"id": "gpt-5.3-codex",
"object": "model",
"shutdown_date": "2027-04-01"
}
Anthropic does not supply an API-level shutdown date property within the completions request schema. Instead, teams audit consumption through the Claude Console: navigate to the Usage section, click Export, and download a CSV file showing token activity broken down across API keys and individual model identifiers. For Google, developers monitor the official Gemini API documentation changelogs, breaking change announcements, or review activity via the AI Studio dashboard logs.
Sampling Parameter Sunsets: Temperature, Top-p, and Top-k
Model deprecation extends beyond complete checkpoints to the underlying generation arguments accepted by HTTP endpoints. Both Anthropic and Google have initiated breaking changes that retire traditional stochastic sampling controls on frontier generations, shifting orchestration toward steering via system prompts.
- Anthropic deprecated
temperature,top_p, andtop_kfor Claude 4.7, later models, and Claude Mythos Preview. Setting these arguments to non-default values triggers an HTTP 400 bad request error. Additionally, Python SDK releases (v1.0 and later) delete these properties from client types, causing invocations to raise a client-side TypeError. - Google deprecated
temperature,top_p, andtop_kstarting with the release of Gemini 3.6 Flash on July 21, 2026, advising developers to omit them when structuring requests. - OpenAI continues serving legacy parameter surfaces across standard inference endpoints, although specialized agentic tools and reasoning variants handle reasoning modes internally (such as
reasoning.mode: prowhen routinggpt-5-pro-2025-10-06workloads togpt-5.6-sol).
For engineering teams migrating legacy pipelines, verify parameters against structured outputs across APIs and read the guide on extended thinking and reasoning to avoid runtime 400 validation rejections.
Recent Decommission Timelines and Target Replacements
Each provider publishes specific transition paths when queuing legacy systems for shutdown. The following table tracks major deprecations and their recommended destination models across Anthropic, OpenAI, and Google.
| Provider | Deprecated Identifier | Retirement Date | Recommended Replacement |
|---|---|---|---|
| Anthropic | claude-sonnet-4-5-20250929 | November 30, 2026 | claude-sonnet-5-5 |
| Anthropic | claude-opus-4-1-20250805 | August 5, 2026 | claude-opus-4-8 |
| Anthropic | claude-3-7-sonnet-20250219 | February 19, 2026 | claude-sonnet-5-5 |
| OpenAI | gpt-4o-2024-08-06 | January 14, 2027 | gpt-6-luna |
| OpenAI | gpt-5.3-codex | April 1, 2027 | gpt-6-sol |
| OpenAI | gpt-5.4-nano | April 1, 2027 | gpt-6-luna |
| OpenAI | o3-2025-04-16 | December 11, 2026 | gpt-5.6-sol |
| gemini-3.7-flash | Immediate reroute (Oct 8, 2026) | gemini-3.8-flash | |
| gemini-3.5-flash | Immediate reroute (Oct 8, 2026) | gemini-3.6-flash | |
| deep-research-pro-preview-12-2025 | October 23, 2026 | deep-research-preview-04-2026 |
Switching to current generation endpoints changes token processing costs. For instance, claude-sonnet-5-5 costs $2 per 1M input tokens, $0.2 per 1M cached input tokens, and $10 per 1M output tokens. Migrating from GPT-4 variants to gpt-6-luna reduces rates to $0.1 per 1M input tokens, $0.01 per 1M cached tokens, and $0.5 per 1M output tokens. Moving to gemini-3.8-flash sets pricing at $0.75 per 1M input tokens, $0.075 cached input, and $3.75 output. Calculate project impact with the LLM API cost calculator.
Migration and Sunset Preparation Checklist
To prevent service interruptions across your production environments, establish a standardized migration workflow before models reach terminal shutdown dates:
- Poll the OpenAI GET /v1/models endpoint in your CI/CD pipeline or weekly monitoring jobs to capture changes to the shutdown_date field.
- Export Anthropic usage data monthly via the Claude Console Usage tab to locate active client keys invoking deprecated identifiers like claude-sonnet-4-5-20250929.
- Update Gemini endpoints to explicit current models (such as gemini-3.8-flash or gemini-3.5-flash-lite) rather than relying passively on backend routing.
- Scrub deprecated generation hyperparameters (temperature, top_p, top_k) from payloads routed to Claude 4.7+ and Gemini 3.6+ models to prevent HTTP 400 errors.
- Audit custom fine-tuning workflows at OpenAI: new job creation halts on January 6, 2027, with serving disabled once base models sunset.
- Verify agent-specific tool interfaces when upgrading, such as updating parameter naming from snake_case to PascalCase when moving between Google Antigravity Agent releases.