Skip to content
modeldeprecations.dev
Calendar Changelog API GitHub

Providers

Amazon Anthropic Cohere DeepSeek Google MiniMax Mistral Moonshot AI OpenAI Qwen xAI Xiaomi MiMo Z.ai

Providers

Amazon 8 Anthropic 31 Cohere 35 DeepSeek 5 Google 34 MiniMax 17 Mistral 53 Moonshot AI 18 OpenAI 106 Qwen 26 xAI 26 Xiaomi MiMo 4 Z.ai 22
Amazon nova-sonic

Amazon Nova Sonic (Amazon)

Deprecated Shuts down in 42 days

Yes — Amazon deprecated Amazon Nova Sonic on March 13, 2026 and shuts it down on September 14, 2026 (in 42 days). The recommended replacement is OpenAI GPT Realtime 2.1.

Amazon's speech-to-speech model, taking audio in and returning audio plus a transcript over a bidirectional stream, in five languages. Legacy since 13 March 2026 with an end of life on 14 September, and Bedrock's table names no replacement - there is currently no other Amazon-trained speech model on Bedrock. The Bedrock invoke id is amazon.nova-sonic-v1:0.

The facts

Provider Amazon
Model id nova-sonic
Released —
Deprecated March 13, 2026
Shutdown September 14, 2026 · in 42 days
Status Deprecated
Last verified August 3, 2026

What to use instead

OpenAI GPT Realtime 2.1 gpt-realtime-2.1 Active

Bedrock's lifecycle table names no successor and Nova 2 Sonic is documented by name only, with no Bedrock model id published in the pages cited here. OpenAI's realtime model is the nearest live speech-to-speech equivalent this catalogue can point at; it is our suggestion, not Amazon's.

What parameters gpt-realtime-2.1 accepts on modelparams.dev

Change the id

In most cases the migration is one string. Check the parameter differences before you ship — newer models drop older sampling knobs more often than they add them. 

- "model": "nova-sonic"
+ "model": "gpt-realtime-2.1"

Frequently asked questions

When will Amazon Nova Sonic shut down?
Amazon Nova Sonic is scheduled to shut down on September 14, 2026, in 42 days.
What replaces Amazon Nova Sonic?
Amazon recommends OpenAI GPT Realtime 2.1 (`gpt-realtime-2.1`) as the replacement for Amazon Nova Sonic. Bedrock's lifecycle table names no successor and Nova 2 Sonic is documented by name only, with no Bedrock model id published in the pages cited here. OpenAI's realtime model is the nearest live speech-to-speech equivalent this catalogue can point at; it is our suggestion, not Amazon's.
Is Amazon Nova Sonic still available?
Yes, for now. Amazon Nova Sonic is deprecated but still answers requests until September 14, 2026. Deprecated models get no further updates and are the first to lose capacity.
How do I migrate from Amazon Nova Sonic to gpt-realtime-2.1?
Swap the model id in your API call: replace "nova-sonic" with "gpt-realtime-2.1". Bedrock's lifecycle table names no successor and Nova 2 Sonic is documented by name only, with no Bedrock model id published in the pages cited here. OpenAI's realtime model is the nearest live speech-to-speech equivalent this catalogue can point at; it is our suggestion, not Amazon's.

Sources

Found something wrong? The data is one YAML file — edit it on GitHub.

This page, for machines

JSON Markdown Badge JSON

Drop the badge in a README so it turns red the day this model is retired:

[![nova-sonic](https://img.shields.io/endpoint?url=https://modeldeprecations.dev/badge/amazon/nova-sonic.json)](https://modeldeprecations.dev/amazon/nova-sonic)

Other Amazon models

All Amazon deprecations →

How to use this site

The full guide, in Markdown. Copy it and hand it to a coding agent verbatim.

# How to use modeldeprecations.dev

[modeldeprecations.dev](https://modeldeprecations.dev) answers one question per model: is it deprecated,
when does it shut down, and what replaces it. Every date on the site is cited to the
provider's own documentation, with the date we last checked it.

Three lifecycle states are used, matching the providers' own vocabulary:

- **active** — still supported, no deprecation announced.
- **deprecated** — the provider has announced it is going away. It still answers requests.
- **retired** — API access is gone. Requests fail.

A model can be active *and* carry a shutdown date: Anthropic publishes a
"not sooner than" date and Google publishes an earliest shutdown date for models
that are current and fully supported. Those are reported as scheduled shutdowns,
not deprecations, because that is what the provider actually committed to.

## Check one model

```bash
curl https://modeldeprecations.dev/api/v1/models/openai/gpt-4-32k.json
```

Or fetch the page as Markdown, which includes the answer, the dates and the sources:

```bash
curl https://modeldeprecations.dev/openai/gpt-4-32k.md
```

## Full catalog

```bash
curl https://modeldeprecations.dev/api/v1/models.json
```

Ids are `provider/model`. Dated snapshots (`gpt-4-32k-0613`) are listed as aliases of
their canonical entry rather than as separate records.

Each entry carries `computed_status`, the lifecycle state recomputed against the build
date — so a model whose shutdown has passed reads `retired` without anyone editing data.

## Badges

A shields.io endpoint per model, for READMEs that should go red when a model dies:

```
https://img.shields.io/endpoint?url=https://modeldeprecations.dev/badge/openai/gpt-4-32k.json
```

## Calendar and changelog

- Subscribe to every shutdown date: https://modeldeprecations.dev/calendar.ics
- Changelog RSS: https://modeldeprecations.dev/changelog.xml

## JSON Schema

```bash
curl https://modeldeprecations.dev/api/v1/schema.json
```

## Contribute

Data lives in YAML under `models/{provider}/{model}.yaml` in the
[GitHub repo](https://github.com/mnfst/modeldeprecations.dev). Every lifecycle date needs a source URL; CI rejects a
date without one. Open a PR.

## For agents

- Machine-readable overview: https://modeldeprecations.dev/llms.txt
- This guide plus every model inline: https://modeldeprecations.dev/llms-full.txt
- Any page also exists as Markdown: append `.md` to a model URL.
- In a browser, this site registers a **WebMCP** tool on `navigator.modelContext`:
  `check_model_deprecation`, `list_shutdowns`, `find_replacement`, and `get_usage_guide`.
- Parameters each model accepts are catalogued on the sibling site, https://modelparams.dev.