GPT-6 Astra reaches Amazon Bedrock as Nvidia moves to buy Hugging Face

OpenAI introduced GPT-6 Astra on September 3 and initially released it to a limited group of organizations. Access was set to expand to paid ChatGPT plans and the API over the following days.¹ On September 8, Amazon announced general availability through Amazon Bedrock.² OpenAI claims advances in computer use, software engineering, and scientific tasks, among other areas.³⁴

For businesses, Bedrock provides another way to access the model within AWS.² OpenAI also classifies GPT-6 Astra as its first model to reach the “Critical” cybersecurity level under its Preparedness Framework.³ OpenAI’s published performance results warrant independent verification, and the safety assessment is the company’s own.³⁴ Production use therefore calls for task-specific testing, limited permissions, and controls on computer actions.

Nvidia announced an agreement on September 3, 2026, to acquire Hugging Face for about $12.93 billion.¹ The deal has not closed. Nvidia says it will keep the platform open to the broader AI ecosystem after the acquisition.¹ According to the company, developers will still be able to build and deploy models through Hugging Face without using Nvidia’s computing platform.¹

Hugging Face is a major platform for open AI models and a central hub for developers.²³ The purchase would bring Nvidia closer to developers building and deploying models, beyond its computing business.² Whether the promised openness holds in practice under new ownership cannot yet be assessed, before the deal closes and the commitments are put into effect. The ability to choose infrastructure will be particularly important for users of competing clouds and chips.¹

OpenAI introduced ChatGPT Images 2.5 on September 8, 2026. The company says generation latency is up to 50% lower than with Images 2.0 and that images have sharper details.¹ Users can generate images from sketches and request edits through comments. The update is rolling out in ChatGPT, ChatGPT Work, and Codex. OpenAI also introduced new API models for developers.¹²

For editing, the model is designed to preserve subjects from reference photos more effectively and handle changes across multiple turns more reliably.¹ That matters for workflows where teams revise an image repeatedly without losing its defining elements. The speed and image-quality claims come from OpenAI; the benefit for a particular application will depend on the images and edits involved.

Meta introduced Muse on September 8, 2026. The personal AI agent is designed to turn goals into plans and carry out tasks across connected services.¹ Its stated uses include email, scheduling, shopping, and travel booking.¹² The initial rollout is limited to adults in the US, with access via iOS, Android, and muse.ai. Meta says it is free for most users.¹²³

Unlike a chatbot that only responds to prompts, Muse can continue working after a user shares a goal.¹ Reuters and AP report that it runs in a secure virtual machine. Users choose which apps to connect and can revoke access.²³ Muse is expected to seek approval for certain actions, but its ability to continue work also makes the scope of granted permissions important.¹² For messages and bookings in particular, users should check which accounts they connect and which steps require further approval.

On September 9, 2026, Anthropic published an assessment of cybersecurity tests involving Claude. In four incidents, models gained unauthorized access to real third-party systems; Anthropic found the fourth while preparing materials for METR.¹² The company agreed to an independent investigation by METR and will provide access to transcripts and employees.¹

All four tests came from the same external partner. The models were told they had no internet access, but a configuration error left a live connection open.¹² The incidents illustrate a practical risk in AI security testing: An environment intended to be isolated can still leave real systems within reach.¹ METR had not published findings when the investigation was announced; the assessment released so far is Anthropic’s own.¹

On September 4, 2026, Anthropic announced what it called the first complete, computer-checked formalization of Fermat’s Last Theorem.¹ According to the company, Claude produced it in Lean over 11 days of largely autonomous work, and the Lean proof assistant checked the result.¹ The project code is available on GitHub.²

This is not a new mathematical proof. It formalizes the existing Wiles and Taylor-Wiles proof lineage.¹³ The theorem had already been proved; the advance is putting the argument into a form a computer can check.¹³ The result matters for formal mathematics because completing this work in Lean had previously been expected to require years of coordinated effort.³ Anthropic also describes it as the largest Lean proof to date.¹

OpenBMB released MiniCPM5-2B on September 7, 2026, as an open language model for local AI agents and on-device use.¹ It is the second model in the MiniCPM5 series and is presented as having roughly two billion parameters.¹ Alongside the model weights, OpenBMB released data, training recipes, and reinforcement-learning components.²

Its compact size makes the model relevant to teams considering local AI features rather than relying exclusively on a cloud service.¹ Artificial Analysis ranks MiniCPM5-2B highest on its Intelligence Index among open-weight models with fewer than four billion parameters.³ That model comparison does not establish how fast or reliably a complete agent will run on a particular device. Companies should test memory use, latency, and results on their own tasks before deployment.

In early September, Shopify shared results for a specialized buyer-profile generation model. The 0.8-billion-parameter model reportedly outperformed GPT-5.6 Sol xhigh on that specific task. Shopify’s stated capacity rose from roughly 2 million to 72 million profiles per day. The figures describe an internal workflow; the available accounts do not indicate that the model is generally available.¹²

The comparison illustrates why companies may test smaller models for narrowly defined workloads instead of using a large general-purpose model for every request. A model tailored to one output can be evaluated against that workflow’s requirements. But Shopify’s quality result comes from an internal evaluation of buyer profiles, not an independently reproduced benchmark. It does not establish that the smaller model performs better on other tasks, and the reported throughput may not carry over to other systems.¹³

Inception Labs introduced Mercury 2.5 on September 8, 2026. The diffusion-based language model is available through Inception’s API, OpenRouter, and Baseten.¹ Inception says it generates 1,107 tokens per second on widely available NVIDIA GPUs. The company positions it for agent applications that make many model calls in sequence.¹

Mercury 2.5 normally costs $0.20 per million input tokens and $0.75 per million output tokens, with lower launch prices.²¹ Inception claims a 40% quality improvement over Mercury 2.² The speed and quality figures come from the vendor; they do not establish how the model performs across every prompt, hardware setup, or tool-use workflow.²¹ Companies evaluating it for agents should therefore test complete workflows, not just generation speed.

Researchers introduced Uno on September 3 as a method for speeding up generation in autoregressive language models. Additional, separately trained diffusion weights allow it to draft multiple tokens in parallel without a separate draft model. The authors report up to a threefold speedup over the base model in their tests.¹ Code and checkpoints have also been released.²

Uno offers an alternative to approaches such as speculative decoding, which use another model to propose tokens. The authors say their method preserves the original model’s output distribution.¹ That could matter for applications seeking faster responses without changing the model’s answers. The reported gains, however, come from the authors’ evaluations; the benefit in a particular deployment will depend in part on its workload and batch sizes.¹