All promotions
GuidesDashboardMCP

How fast do LLM models become obsolete?

A model's real shelf life is set by its own maker's release cadence. Across 207 dated models from 15 labs, the busiest brands ship a new model every few weeks — and the global release rate keeps climbing. Three views: each lab's timeline, its median time between releases, and the market-wide acceleration. This measures release cadence, not forced retirement — old models often stay served, at falling prices.

Every release, brand by brand

One row per lab, one dot per model release. Rows sorted by cadence — the density tells the story. Hover a dot for the model; click for its page.

Jul 20242025Jul 20252026Jul 2026Alibaba · 50qwen-plus — 2024-01-25qwen-2.5-72b — 2024-09-19qwen-2.5-7b — 2024-10-16qwen-2.5-coder-32b — 2024-11-11qwen3-next-80b-a3b — 2024-12-01qwen2.5-vl-72b — 2025-02-01qwen3-30b-a3b — 2025-04-28qwen3-14b — 2025-04-28qwen3-32b — 2025-04-30qwen3-8b — 2025-04-30qwen3-235b-a22b — 2025-07-21qwen3-coder — 2025-07-22qwen3-coder-plus — 2025-07-23qwen3-235b-a22b-thinking — 2025-07-25qwen3-coder-flash — 2025-07-28qwen3-30b-a3b-thinking — 2025-07-30qwen3-coder-30b-a3b — 2025-07-31qwen3-next-80b-a3b-thinking — 2025-09-11qwen3-vl-235b-a22b — 2025-09-23qwen3-vl-235b-a22b-thinking — 2025-09-23qwen3-max — 2025-09-23qwen3-omni-30b-a3b — 2025-09-24qwen3-omni-30b-a3b-thinking — 2025-09-24qwen3-vl-30b-a3b — 2025-10-04qwen3-vl-30b-a3b-thinking — 2025-10-06qwen3-vl-8b — 2025-10-14qwen3-vl-32b — 2025-10-21qwen3-max-thinking — 2026-01-23qwen3-coder-next — 2026-02-03qwen3.5-397b-a17b — 2026-02-15qwen3.5-plus — 2026-02-16qwen3.5-35b-a3b — 2026-02-23qwen3.5-27b — 2026-02-23qwen3.5-9b — 2026-02-23qwen3.5-122b-a10b — 2026-02-23qwen3.5-flash — 2026-02-23qwen3.6-plus — 2026-04-02qwen3.6-35b-a3b — 2026-04-17qwen3.6-max-preview — 2026-04-20qwen3.6-27b — 2026-04-22qwen3.6-flash — 2026-04-27qwen3.7-max — 2026-05-21qwen3.7-plus — 2026-06-02qwen3.7-flash — 2026-07-15qwen3.8-max — 2026-08-03qwen3.8-2.4t-a95b — 2026-08-12qwen3.8-27b — 2026-08-14qwen3.8-flash — 2026-08-26qwen3.8-omni-flash — 2026-09-17qwen3.8-max-prime — 2026-09-23Other · 30hermes-3-llama-3.1-405b — 2024-08-16hermes-3-llama-3.1-70b — 2024-08-18phi-4 — 2024-12-11reka-flash-3 — 2025-03-12ernie-4.5-vl-424b-a47b — 2025-06-30ernie-4.5-21b-a3b — 2025-06-30dolphin-mistral-24b-venice-edition — 2025-07-09baichuan-m2-32b — 2025-08-13hermes-4-405b — 2025-08-26devstral — 2025-12-09reka-edge — 2026-03-20mimo-v2.5 — 2026-04-22mimo-v2.5-pro — 2026-04-22step-3.7-flash — 2026-05-29hy3 — 2026-07-06inkling — 2026-07-15ling-3.0-flash — 2026-07-23inkling-small — 2026-07-30muse-glimmer-30b — 2026-08-10hy4-preview — 2026-08-28muse-spark-1.3 — 2026-09-02ling-3.0-flash-fin — 2026-09-03ling-3.0-flash-vl — 2026-09-04ling-3.0-flash-sante — 2026-09-04step-5-preview — 2026-09-16mimo-v2.6-pro — 2026-09-21mimo-v2.6-flash — 2026-09-22ling-3.1-flash — 2026-09-29muse-wonder — 2026-10-12muse-spark-open — 2026-10-15Anthropic · 10claude-opus-4-1 — 2025-08-05claude-sonnet-4-5 — 2025-09-29claude-haiku-4-5 — 2025-10-15claude-fable-5 — 2026-06-09claude-sonnet-5 — 2026-06-30claude-opus-5 — 2026-07-24claude-fable-5-1 — 2026-09-01claude-opus-5-5 — 2026-09-22claude-sonnet-5-5 — 2026-09-28claude-haiku-5-5 — 2026-10-07Z.ai · 16glm-4-32b — 2024-01-01glm-4.5-air — 2025-07-28glm-4.5 — 2025-07-28glm-4.5v — 2025-08-11glm-4.6 — 2025-09-30glm-4.6v — 2025-12-08glm-4.7 — 2025-12-22glm-4.7-flash — 2026-01-19glm-5 — 2026-02-12glm-5-turbo — 2026-03-16glm-5v-turbo — 2026-04-01glm-5.1 — 2026-04-07glm-5.2 — 2026-06-13glm-5.3 — 2026-08-14glm-5.3-flash — 2026-08-26glm-5.3-flashx — 2026-09-18Meta · 8llama-3.1-8b — 2024-07-23llama-3.1-70b — 2024-07-23llama-3.2-3b — 2024-09-25llama-3.2-1b — 2024-09-25llama-3.3-70b — 2024-12-06llama-4-scout — 2025-04-05llama-4-maverick — 2025-04-05llama-guard-4-12b — 2025-04-30Google · 15gemma-2-27b — 2024-07-13gemma-3-27b — 2025-03-12gemma-3-4b — 2025-03-12gemma-3-12b — 2025-03-13gemini-2.5-pro — 2025-06-17gemini-2.5-flash — 2025-06-17gemini-3-flash-preview — 2025-12-17gemini-3.1-pro-preview — 2026-02-19gemma-4-31b — 2026-04-02gemma-4-26b-a4b — 2026-04-02gemini-3.6-flash — 2026-07-21gemini-3.5-flash-lite — 2026-07-21gemini-3.7-flash — 2026-08-13gemini-3.8-flash — 2026-09-02gemini-4-argon — 2026-09-30Mistral · 18mixtral-8x22b — 2024-04-17mistral-nemo — 2024-07-01mistral-small-24b — 2025-01-30mistral-saba — 2025-02-17mistral-small-3.1-24b — 2025-03-17mistral-medium — 2025-05-07mistral-medium-3 — 2025-05-07mistral-small-3.2-24b — 2025-06-20voxtral-small-24b — 2025-07-15codestral — 2025-07-30mistral-medium-3.1 — 2025-08-13mistral-large — 2025-12-02ministral-14b — 2025-12-02ministral-8b — 2025-12-02ministral-3b — 2025-12-02mistral-small-4 — 2026-03-16mistral-medium-3-5 — 2026-04-30mistral-large-4 — 2026-10-06DeepSeek · 15deepseek-chat — 2024-12-26deepseek-v3 — 2024-12-26deepseek-chat-v3 — 2025-03-24deepseek-r1 — 2025-05-28deepseek-chat-v3.1 — 2025-08-21deepseek-v3.1-terminus — 2025-09-22deepseek-v3.2-exp — 2025-09-29deepseek-v3.2 — 2025-12-01deepseek-v4-pro-0423 — 2026-04-24deepseek-v4-flash-0423 — 2026-04-24deepseek-v4-flash — 2026-07-31deepseek-v4 — 2026-08-13deepseek-v4-flash-vision-exp — 2026-08-21deepseek-v4.1-flash — 2026-09-10deepseek-v4.1-flash-1017 — 2026-10-17OpenAI · 15gpt-4o — 2024-05-13gpt-4o-mini — 2024-07-18gpt-4.1-mini — 2025-04-14gpt-4.1 — 2025-04-14gpt-oss-20b — 2025-08-05gpt-oss-120b — 2025-08-05gpt-oss-safeguard-20b — 2025-10-29gpt-5.3-codex — 2026-02-24gpt-5.6-sol — 2026-07-09gpt-5.6-terra — 2026-07-09gpt-5.6-luna — 2026-07-09gpt-6-astra — 2026-09-03gpt-6-sol — 2026-09-22gpt-6-luna — 2026-09-22gpt-6.1-sol — 2026-09-29xAI · 3grok-4.5 — 2026-07-08grok-4.6 — 2026-08-12grok-4.7 — 2026-09-21MiniMax · 9minimax-01 — 2025-01-15minimax-m1 — 2025-06-17minimax-m2 — 2025-10-27minimax-m2.1 — 2025-12-23minimax-m2-her — 2026-01-23minimax-m2.5 — 2026-02-12minimax-m2.7 — 2026-03-18minimax-m3 — 2026-06-01minimax-m3.1-flash — 2026-10-19Moonshot · 5kimi-k2-thinking — 2025-11-06kimi-k2.5 — 2026-01-27kimi-k2.6 — 2026-04-21kimi-k2.7-code — 2026-06-12kimi-k3 — 2026-07-16NVIDIA · 6nemotron-3-nano-30b-a3b — 2025-12-15nemotron-3-super-120b-a12b — 2026-03-11nemotron-3-ultra-550b-a55b — 2026-06-04nemotron-3.5-content-safety — 2026-06-04nemotron-3.5-lightning — 2026-08-11nemotron-3.5-super-vl — 2026-10-15Cohere · 5command-r — 2024-08-30command-r-plus — 2024-08-30command-r7b — 2024-12-02command-a — 2025-03-13command-a-plus — 2026-09-22IBM · 2granite-4.0-h-micro — 2025-10-02granite-4.2-8b — 2026-08-31

A new model every N days

Median time between two releases of the same lab — the shorter the bar, the faster yesterday's model stops being the newest.

AlibabaAlibaba: a new model every 7 days (median over 50 releases) · last gap 6 daysevery 7 daysOtherOther: a new model every 9 days (median over 30 releases) · last gap 3 daysevery 9 daysAnthropicAnthropic: a new model every 21 days (median over 10 releases) · last gap 9 daysevery 21 daysZ.aiZ.ai: a new model every 24 days (median over 16 releases) · last gap 23 daysevery 24 daysMetaMeta: a new model every 25 days (median over 8 releases) · last gap 25 daysevery 25 daysGoogleGoogle: a new model every 26 days (median over 15 releases) · last gap 28 daysevery 26 daysMistralMistral: a new model every 28 days (median over 18 releases) · last gap 159 daysevery 28 daysDeepSeekDeepSeek: a new model every 34 days (median over 15 releases) · last gap 37 daysevery 34 daysOpenAIOpenAI: a new model every 38 days (median over 15 releases) · last gap 7 daysevery 38 daysxAIxAI: a new model every 38 days (median over 3 releases) · last gap 40 daysevery 38 daysMiniMaxMiniMax: a new model every 66 days (median over 9 releases) · last gap 140 daysevery 66 daysMoonshotMoonshot: a new model every 67 days (median over 5 releases) · last gap 34 daysevery 67 daysNVIDIANVIDIA: a new model every 68 days (median over 6 releases) · last gap 65 daysevery 68 daysCohereCohere: a new model every 98 days (median over 5 releases) · last gap 558 daysevery 98 daysIBMIBM: a new model every 333 days (median over 2 releases) · last gap 333 daysevery 333 days

Median of the gaps between consecutive dated releases of each brand (15 brands with 2+ dated models).

The release rate is accelerating

Model releases per quarter across all brands in the index — the market ships faster every year, so shelf lives shrink for everyone.

26512024-Q1: 2 releases220242024-Q2: 2 releases2Q22024-Q3: 12 releases12Q32024-Q4: 8 releases8Q42025-Q1: 11 releases1120252025-Q2: 18 releases18Q22025-Q3: 30 releases30Q32025-Q4: 20 releases20Q42026-Q1: 21 releases2120262026-Q2: 25 releases25Q22026-Q3: 51 releases51Q32026-Q4: 7 releases7Q4

Only models with a full, sourced release date and a live paid price are counted (207 of the index). The current quarter is still filling up. Verified 2026-10-23. See also price by model age, per brand and the falling price of intelligence ·methodology.

FAQ

How quickly do LLM API models get replaced?

Fast, and accelerating. Across 207 dated models, the fastest-shipping lab (Alibaba) puts out a new model every 7 days on median. Most major labs ship a new model every one to two months — a model is rarely the newest of its brand for more than a quarter.

Is the pace of LLM releases accelerating?

Yes. Counting the models in this index by release quarter, 2024-Q1 saw 2 releases while 2026-Q4 saw 7 — the release rate has climbed steadily, so the effective shelf life of any given model keeps shrinking.

Does a new release make the previous model obsolete?

Not immediately — older models often stay served and get price cuts, which can make them the better value. This page measures release cadence, not forced retirement: it tells you how long a model stays the newest of its line, not when it disappears.