About LLMSunset

What is this?

LLMSunset is a commercial service that automatically tracks the lifecycle of Large Language Models across all major providers. Deprecation dates, pricing updates, context window sizes, capability flags, and sunset timelines — all indexed daily and consolidated in one place.

When a provider silently changes a model's status or price, LLMSunset detects it and surfaces it immediately. Never miss a deprecation that could break your production workflows.

Methodology

• Data is collected daily at 06:00 UTC via automated pipeline.
• Each provider has a dedicated collector: API-based where the provider exposes one, page parsing otherwise.
• Every ModelRecord includes a sources[] array with the exact URLs and fetch timestamps used.
• If a collector fails, the affected models are marked stale rather than removed — no silent data loss.
• Dates are never invented: if a date is not officially published, it is stored as null.

Covered providers

• Anthropic
• OpenAI
• Mistral
• Cohere
• xAI (Grok)
• DeepSeek
• Groq
• Together AI
• Amazon Bedrock
• Google Vertex AI
• Microsoft Azure OpenAI

Data schema

All model data is validated against a strict Zod schema (schema version 1.0.0).

Key fields per record:

• id, canonical_id, provider, display_name, family
• status: ga | preview | beta | deprecated | retired | unknown
• released_at, deprecated_at, sunset_at, retired_at (YYYY-MM-DD or null)
• context_window, max_output_tokens
• capabilities: tool_use, vision, json_mode, streaming, caching, reasoning
• pricing: input_per_mtok, output_per_mtok (USD per million tokens)
• sources[]: url, fetched_at, kind (api | scrape)
• freshness: fresh | stale

Capabilities explained

Each ModelRecord carries six capability flags. A value of null means "not yet collected" — never "not supported". See the compatibility matrix for a full cross-model view.

tool_use — Tool use / Function calling

The model can call external tools or functions as part of its output. This includes OpenAI-style function calling, Anthropic's tool use API, and similar mechanisms. Required for building agents and structured workflows.

vision — Vision / Image input

The model accepts image content (base64 or URL) alongside text in its input. Useful for document parsing, screenshot analysis, and multimodal tasks. Note: some providers restrict image inputs to specific API tiers.

json_mode — JSON mode / Structured output

The model guarantees that its output is valid JSON, either via a dedicated response_format parameter or constrained decoding. Distinct from instructing the model to "respond in JSON" without enforcement.

streaming — Streaming (SSE)

The model supports token-by-token streaming via Server-Sent Events (SSE). This allows applications to display responses progressively and reduces perceived latency. All major chat-oriented models support streaming today.

caching — Prompt caching

The provider offers server-side caching of prompt prefixes (e.g. Anthropic's prompt cache, OpenAI's cached inputs). Repeated identical prefixes are served at a discounted token rate, significantly reducing costs for long-context applications.

reasoning — Reasoning / Extended thinking

The model performs an explicit chain-of-thought reasoning step before producing its final answer (e.g. OpenAI's o-series, Anthropic's extended thinking on Claude 3.7+, DeepSeek-R1). Typically increases response quality for complex tasks at the cost of higher latency and token usage.