PLUS: OpenAI's new voice AI, a 10x cheaper agent, and Anthropic's broken Fable model
Happy reading
OpenAI has now fully released its powerful GPT-5.6 family of models, including the science and coding-focused ‘Sol’, after a brief evaluation period requested by the government.
The move makes specialized AI widely available for development in critical fields like biology and cybersecurity. But it also raises a major question: will this release set a new precedent for how top AI labs navigate government oversight in the future?
In today’s Next in AI:
OpenAI’s public release of the GPT-5.6 family
OpenAI's new conversational GPT-Live voice AI
NVIDIA’s agent model with 10x lower costs
Anthropic’s Fable model under fire for safety filters
OpenAI Unlocks GPT-5.6 Sol
Next in AI: OpenAI is now publicly releasing its powerful GPT-5.6 series—Sol, Terra, and Luna—after a brief, government-requested delay for evaluation.
Explained:
The new series includes Sol, its strongest model for coding and science; Terra, a balanced option; and Luna, a lightweight model optimized for speed.
Alongside the models, the company launched GPT-Live, a new voice AI that can listen and speak simultaneously to enable more natural conversations.
This full rollout ends a limited preview requested by the U.S. government, a process OpenAI stated should not become the long-term default for future releases.
Why It Matters:
Making these specialized models widely available accelerates development in key areas like cybersecurity and biology. The move also signals a pushback against turning government evaluation into a mandatory step for AI deployment.
OpenAI Introduces GPT-Live
Next in AI: OpenAI just launched GPT-Live, a new voice model for ChatGPT that can listen and speak at the same time. The update makes conversations feel remarkably natural and fluid, eliminating the awkward pauses of previous voice assistants.
Explained:
The model uses a full-duplex architecture, letting it process your voice and generate responses simultaneously. This ends the rigid, turn-based interactions of older systems, allowing you to interrupt or pause just like in a real conversation.
For complex questions, GPT-Live delegates heavy tasks like web searches to GPT-5.5 in the background, all while keeping the conversation flowing without interruption.
The new experience includes better listening in noisy environments, visual answer cards for topics like weather and sports, and is now rolling out globally to all ChatGPT users.
Why It Matters: This update moves voice assistants beyond rigid, turn-based commands and into the realm of natural, fluid dialogue. This opens the door for more complex, long-running tasks where AI can act as a true collaborator, not just a simple tool.
Nvidia's 10x Agent
Next in AI: NVIDIA's Nemotron 3 Ultra model is achieving top-tier agent performance at a tenth of the cost of leading proprietary models. This efficiency boost comes from a specially tuned harness from the AI framework LangChain, not from retraining the model itself.
Explained:
Instead of expensive retraining, LangChain optimized the system around the model—a process called harness engineering—by adjusting prompts and tools to unlock its full potential.
The results are striking: Nemotron 3 Ultra achieved an 86% accuracy score on LangChain’s agent benchmark for just $4.48 per run, compared to over $43 for the next leading model.
The collaboration produced an open reference blueprint called NVIDIA NemoClaw, giving enterprises a full, customizable stack they can own and run anywhere.
Why It Matters:
This proves that open-source models can compete with the best proprietary systems without needing massive, costly retraining. For developers and businesses, this unlocks the ability to build and deploy powerful, specialized AI agents more quickly and affordably than ever before.
Anthropic's Broken Fable
Next in AI: Anthropic's new "safety-conscious" model, Fable, is drawing sharp criticism from developers and researchers. They argue its overly aggressive safety filters make it functionally useless for legitimate coding and scientific tasks.
Explained:
The model's strict new safeguards are aggressively flagging and rejecting innocuous prompts, particularly those involving biological or complex coding terminology. Researchers report that requests to rewrite open-source bioinformatics software are instantly denied.
This hyper-cautious approach follows a temporary global suspension of Fable in June due to US export controls after a security bypass was discovered. Anthropic responded by deploying a much stricter classifier to prevent future jailbreaks.
The filtering issues go far beyond obvious edge cases, with one researcher detailing how Fable refused to engage with a purely mathematical problem even after it was stripped of all potentially triggering language. The model ultimately only answered a question about the best ice cream flavor.
Why It Matters: This situation highlights the critical tension between advancing AI capability and implementing effective, non-obtrusive safety measures. How Anthropic calibrates Fable's filters will be a key test for deploying powerful, specialized models without stifling research and innovation.
AI Pulse
Noma Labs discovered a critical prompt injection vulnerability in GitHub’s new Agentic Workflows, dubbed "GitLost," that allows attackers to leak data from private repositories by posting a crafted public issue.
OpenAI found that the widely used SWE-Bench Pro coding benchmark has significant flaws, estimating that roughly 30% of its tasks are broken due to issues like overly strict tests and misleading prompts.
Bona Books detailed how the small press nearly published two AI-generated stories in a queer fiction anthology, highlighting the growing challenge of detecting synthetic content in creative industries.
Brown University faced a widespread AI cheating scandal after test scores in an economics class plummeted from a 96% average on a take-home midterm to 48% on the in-person final.