Amazon Reportedly Scales Back Nova AI Models, Shifts Resources to New Frontier Research Team
Amazon is reportedly scaling back active development of its Nova Premier, Nova Omni, Reel, and Canvas AI models, according to Business Insider. Resources are shifting to a new Frontier Model Research group led by Pieter Abbeel, with a new foundation model reportedly planned for re:Invent this fall.
Amazon Reportedly Pulls Back on Nova Models
Amazon is scaling back active development of several of its in-house Nova AI models, according to Business Insider, which cited unnamed insiders. The affected models reportedly include the flagship Nova Premier and Nova Omni language models, the Reel video generator, and the Canvas image generator.
According to the report, these models will remain online in "keep the lights on" mode for existing customers but will no longer receive active development. Amazon first introduced the Nova model family in December 2024.
AGI Division Cuts
The reported pullback follows layoffs within Amazon's AGI division and the closure of the AGI Lab, which Amazon founded in 2024 following its acquisition of Adept, an AI startup focused on agentic systems. Business Insider's sourcing suggests these organizational changes preceded the shift away from active Nova development.
Resources Move to Frontier Model Research
According to the report, Amazon is redirecting resources to a group called Frontier Model Research, led by Pieter Abbeel. Abbeel joined Amazon through its acquisition of Covariant, a robotics startup. The group is reportedly preparing a new foundation model, with a release targeted for Amazon's re:Invent conference this fall.
An Amazon spokesperson told Business Insider that AI models remain "one of Amazon's most important projects," though the company did not confirm specifics of the Nova cutbacks or the new model's capabilities.
Context: Amazon's Dual AI Bet
Amazon has pursued a two-track AI strategy: building its own foundation models under the Nova brand while simultaneously investing heavily in external AI labs. The company is among the largest backers of both Anthropic and OpenAI, with combined investment commitments reportedly approaching $100 billion. That dynamic has drawn scrutiny — Amazon competes with the same companies it funds, a tension that appears to be resolving in favor of external partnerships over in-house model development for at least some Nova product lines.
Neither Amazon nor Business Insider disclosed specific performance benchmarks, financial figures tied to the Nova program's downsizing, or a timeline for when "keep the lights on" support might end for affected models.
What This Means
If accurate, this reporting suggests Amazon has concluded that competing head-to-head with OpenAI and Anthropic on general-purpose foundation models isn't where its AI resources are best spent — at least for the Nova Premier, Omni, Reel, and Canvas lines. Redirecting talent toward Pieter Abbeel's Frontier Model Research group, with roots in robotics via Covariant, hints Amazon may be positioning its next model push around embodied AI and agentic systems rather than chasing frontier LLM benchmarks directly.
This wouldn't mean Amazon exits foundation models entirely — a new model is reportedly coming at re:Invent — but it does suggest a narrower, more differentiated bet rather than a broad Nova product family competing on all fronts. Combined with continued heavy investment in Anthropic and OpenAI, Amazon looks to be hedging: fund external frontier labs for general-purpose capability while building proprietary models where it sees unique advantage, likely in robotics, logistics, and AWS infrastructure integration. None of this is confirmed by Amazon directly, and key details — including what happens to existing Nova customers long-term — remain unclear.
Related Articles
Amazon Bedrock adds schema-enforced structured outputs for all Anthropic Claude models on both endpoints
Amazon Bedrock now supports structured outputs for all Anthropic Claude models on both the bedrock-runtime and bedrock-mantle endpoints. The feature uses constrained decoding to enforce a JSON schema during token generation, and it includes strict tool use for agentic workflows.
OpenAI Python SDK v3.39.0 adds gpt-6.1-terra model identifier
OpenAI's Python SDK has released v3.39.0, dated 2026-10-26. The release lists one feature: a new gpt-6.1-terra model identifier added to the API.
Audible's controllable TTS attributes dialogue with 97% accuracy in English and improves character differentiation by 3.
Audible's technology team describes a controllable text-to-speech system that identifies which character speaks each line with 97% accuracy in English and 94% in Spanish. Rendering distinct character voices improved dialogue differentiation by 3.7 points on a MUSHRA test, with statistically significant listening-experience gains for eight of ten voices.
Amazon Polly Generative engine upgrade adds 14 voices, first Hindi and Finnish support, at unchanged pricing
AWS has upgraded the Amazon Polly Generative engine in place, adding 14 voices across 9 locales and launching Generative voices in Hindi and Finnish for the first time. The catalog now totals 57 voices across 11 languages and 25 locales. Voice names, API calls, and pricing are unchanged.
Comments
Loading...