New & deprecated LLM modelsMarch 2026

In March 2026, 6 new models were released and 10 models were deprecated or scheduled to sunset. Newest release: Qwen/Qwen3.6-Plus.

New models — March 2026

6 models with a public release date in March 2026, joined to live list prices and benchmark scores.

ModelProviderInput / 1MOutput / 1MBest scoreReleased
Qwen/Qwen3.6-PlusTogether AI$0.50$3.0057.9%31 Mar 2026
GPT-5.4 MiniOpenAI$0.75$4.5086.9%17 Mar 2026
gpt-5.4-nanoOpenAI$0.20$1.2578.5%17 Mar 2026
gpt-5.4-proOpenAI$30.00$180.0094.6%5 Mar 2026
GPT-5.4OpenAI$2.50$15.0076.9%5 Mar 2026
google/gemini-3.1-flash-liteDeepInfra$0.25$1.5081.8%3 Mar 2026

Deprecated & sunsetting — March 2026

10 models with an upstream retirement or sunset date in March 2026. Plan migrations before the date to avoid broken calls.

ModelProviderInput / 1MOutput / 1MBest scoreSunset date
meta-llama/llama-guard-4-12bGroq$0.20$0.20Retired5 Mar 2026
meta-llama/Meta-Llama-3.1-8B-Instruct-TurboTogether AI$0.18$0.18Retired6 Mar 2026
moonshotai/Kimi-K2-Instruct-0905Together AI$1.00$3.0059.1%Retired6 Mar 2026
gemini-3-pro-previewGoogle (Gemini)$2.00$12.0072.9%Retired9 Mar 2026
meta-llama/llama-4-maverick-17b-128e-instructGroq$0.20$0.6067.0%Retired9 Mar 2026
gpt-4-0125-previewOpenAI$10.00$30.0046.6%Retired26 Mar 2026
gpt-4-0314OpenAI$30.00$60.0046.6%Retired26 Mar 2026
gpt-4-turbo-previewOpenAI$10.00$30.0046.6%Retired26 Mar 2026
gemini-2.5-flash-lite-preview-09-2025Google (Gemini)$0.10$0.40Retired31 Mar 2026
labs-devstral-small-2512Mistral$0.10$0.30Retired31 Mar 2026

Prices from the LLM API Pricing Index, synced daily. Release & benchmark data via Epoch AI (CC BY); deprecation dates from provider metadata. New models are those with a public release date this month that appear in the tracked price catalogue.

Get price-change alerts

One email when a model’s price changes, a new model launches, or a model you might rely on gets a deprecation date. No schedule, no newsletter — it only sends on days something actually changed.

Unsubscribe with one click, any time.

A model you rely on just got deprecated. Would you know?

StackSpend tracks every model you actually call, flags deprecations before they break you, and surfaces cheaper equal-quality swaps from this same live dataset.