Claude Sonnet 4.5 and Sora 2: AI That Works for Hours and Creates Cinema-Quality Video

On 18 September 2025, Anthropic released Claude Sonnet 4.5 with a capability that redefined what autonomous AI agents could do: sustained, unsupervised coding sessions exceeding 30 hours. The model could receive a complex software engineering brief, decompose it into tasks, write code across multiple files, run tests, debug failures, and iterate — all without human intervention.

This was not a benchmark demonstration. Anthropic showed Sonnet 4.5 completing real-world engineering projects: building full-stack web applications, refactoring legacy codebases, implementing API integrations, and resolving complex bug reports. The model maintained coherent context, tracked its own progress, and produced production-quality code that passed automated test suites.

For engineering organisations, the implications were transformative. Tasks that previously required a senior developer's full attention for days could now be delegated to an AI agent with a clear brief and appropriate guardrails. The human role shifted from doing the work to defining the work and reviewing the output — a fundamental change in the economics of software development.

OpenAI launched Sora 2 in September 2025, and the creative industry took notice. The updated model generated video with dramatically improved visual fidelity, physical consistency, and — for the first time — synchronised audio generation. Users could describe a scene and receive a video complete with appropriate sound effects, ambient audio, and even spoken dialogue.

The quality gap between AI-generated and professionally produced video narrowed significantly. Sora 2's outputs maintained consistent lighting, physics, and character appearance across extended sequences — weaknesses that had limited the usefulness of earlier video generation models. The addition of audio eliminated the need for post-production sound design on generated content.

For businesses in marketing, content creation, and e-commerce, Sora 2 opened new possibilities for rapid, cost-effective video production. Product demonstrations, explainer videos, social media content, and advertising creative that previously required production teams and budgets could now be generated in minutes at a fraction of the cost.

In September 2025, California passed SB 53, requiring developers of frontier AI models to implement safety protocols before deployment. The bill mandated safety testing, incident reporting, and the ability to rapidly shut down models that demonstrate dangerous capabilities — the first US state-level legislation to impose binding requirements on AI model developers.

SB 53 was significant not because of its specific provisions — which were moderate compared to the EU AI Act — but because of its jurisdiction. California is home to OpenAI, Anthropic, Google DeepMind, and Meta AI. Legislation passed in Sacramento directly affected the companies building the world's most capable AI systems.

For enterprise AI buyers, SB 53 added another layer to the compliance landscape. While the bill primarily targeted model developers, it signalled that AI safety regulation in the US was moving from federal discussion to state action. Businesses deploying AI systems should expect an increasingly patchwork regulatory environment and plan accordingly.

Salesforce reported that Agentforce adoption grew by 300% in the quarter ending September 2025. Agentforce, Salesforce's platform for deploying AI agents within CRM workflows, had moved from pilot to production across thousands of enterprises. Agents handled customer service inquiries, qualified leads, processed orders, and managed support tickets — autonomously, within Salesforce's existing security and governance framework.

The growth validated a key thesis about enterprise AI adoption: businesses adopt AI fastest when it is embedded in platforms they already use. Rather than asking organisations to adopt new tools, new workflows, and new vendors, Agentforce plugged AI capability directly into the CRM system that sales and support teams already lived in. The training curve was minimal, the integration effort was zero, and the ROI was immediately measurable.

For technology leaders evaluating AI deployment strategies, Salesforce's results reinforced the platform-native approach. Before building custom AI agents, evaluate whether your existing platforms — CRM, ERP, ITSM, HRIS — offer embedded AI capabilities that can deliver value with significantly less integration effort.

September 2025 made autonomous AI real — models working for hours without supervision, generating cinema-quality media, and handling enterprise workflows at scale. But autonomy without trust is a liability. The organisations succeeding with autonomous AI shared common governance practices:

• Clear boundaries: Define what the agent can and cannot do. Autonomous coding with test suites and code review is manageable; autonomous customer communication without oversight is risky.
• Progressive delegation: Start with low-stakes, reversible tasks and expand autonomy as confidence in the system grows. Do not jump from copilot to fully autonomous without building an evidence base.
• Audit trails: Maintain comprehensive logs of every agent action, decision, and output. Regulatory requirements aside, audit trails are essential for debugging, improvement, and accountability.
• Human checkpoints: Even the most capable autonomous agents should have defined checkpoints where a human reviews progress, confirms direction, and authorises continuation.

Autonomous AI is not about removing humans from the loop — it is about moving humans to a higher level of the loop, where they define objectives, set boundaries, and review outcomes rather than executing every step. Discuss how to build a trust framework for autonomous AI in your organisation.