GPT-4 (legacy)→GPT-5 Mini
OpenAI · retired 2025
Successor: $0.250 in / $2.00 out
GPT-5 mini matches GPT-4 on chat quality with first-class tool calling and a fraction of the cost. Use GPT-5 for tasks where GPT-4 was a quality ceiling.
Migration guide · 2026-05-22
Deprecated AI Models and Their Recommended Replacements
Vendors retire models on aggressive schedules. This page tracks the most-migrated models and points each one at a current successor — chosen for migration ease, not raw benchmark performance.
GPT-3.5 Turbo→GPT-5 Mini
OpenAI · retired 2025-01
Successor: $0.250 in / $2.00 out
GPT-5 mini is cheaper than late GPT-3.5 Turbo pricing while being meaningfully more capable. Migration is usually a 1-line swap.
Claude 2.x / Claude Instant→Claude Haiku 4.5 (latest)
Anthropic · retired 2024-07
Successor: $1.00 in / $5.00 out
Claude Haiku 4.5 is the direct successor on the cheap-and-fast tier, with dramatically improved tool calling and structured output.
Claude 3 Opus→Claude Opus 4.5 (latest)
Anthropic · retired 2025
Successor: $5.00 in / $25.00 out
Claude Opus 4.5 is the modern Opus with the same long-form-reasoning sweet spot at lower per-token cost.
Claude 3 Sonnet→Claude Sonnet 4.5 (latest)
Anthropic · retired 2025
Successor: $3.00 in / $15.00 out
Direct in-line successor — same balanced position in Anthropic's lineup, with improved tool calling and stronger output limits.
Gemini 1.5 Pro→Gemini 2.5 Pro
Google · retired 2025
Successor: $1.25 in / $10.00 out
Gemini 2.5 Pro keeps the 1M-token context window while improving reasoning quality and tool-use reliability.
Gemini 1.5 Flash→Gemini 2.5 Flash
Google · retired 2025
Successor: $0.300 in / $2.50 out
Same fast-and-cheap niche; 2.5 Flash's response quality is markedly better at the same price point.
DeepSeek V2 / V2.5→DeepSeek-V3.2
DeepSeek · retired 2025
Successor: $0.252 in / $0.378 out
V3.2 supersedes the V2 line on every metric and is the current default open-weight pick from DeepSeek.
Llama 2→Llama-3.3-70B-Instruct
Meta · retired 2024
Successor: $0.050 in / $0.230 out
Llama 3.3 70B Instruct is the modern flagship — better tool calling, larger context, equivalent licensing.
Qwen 2.x series→Qwen3 235B A22B Instruct 2507
Alibaba · retired 2025
Successor: $0.100 in / $0.100 out
Qwen3 235B is the current flagship and the natural migration target for any Qwen 2.x deployment.
Frequently asked questions
How do you decide a model is deprecated?
We rely on the vendor's own deprecation announcements (retirement schedules, replaced-by notices, removal from official pricing pages). Each entry below was added by a human editor — we deliberately avoid automatic 'old release date = deprecated' heuristics because they produce wrong answers (Llama 3 is old but very much active).
What if the recommended replacement isn't available in my region?
Open the linked /providers/[vendor] page on each successor — it lists every provider currently hosting that model, including regional clouds. If you cannot use the recommended successor, the brand's /alternatives page surfaces non-vendor swaps.
How fresh is this list?
We update the list when a vendor announces a retirement. Each entry has a 'retired on' date that captures when the deprecation took effect or was announced, not when we noticed it.
Why isn't my deprecated model in the list?
We only catalogue widely-used models that show up in production migrations. Niche or short-lived models are usually superseded silently. If you think a model belongs here, the canonical catalogue is updated daily and replacements typically map to that vendor's flagship.
See also
最終更新：
Prices in USD per 1M tokens. Unknown means the provider does not publish per-token pricing.
Pricing and capabilities are refreshed daily and reconciled against each provider's official documentation. Always verify critical production decisions with the provider directly.