OpenAI has launched the agent-style coding model GPT-5.3-Codex, combining the code capabilities of the previous Codex generation with the reasoning and knowledge abilities of GPT-5.2, while improving runtime speed by 25%. The company claims new state-of-the-art results on benchmarks such as SWE-Bench Pro and Terminal-Bench, enabling autonomous completion of long-horizon tasks like development, code review, and debugging in terminal and web environments, with support for collaborative guidance. OpenAI classifies it for the first time as a 'high-capability' cybersecurity model, backed by stronger security measures, and launches a $10 million API credit grant program for cybersecurity applications. The model is now available via paid ChatGPT tiers, apps, CLI, and IDE plugins, with API access to follow.

Anthropic has released its flagship model Claude Opus 4.6, significantly enhancing stability in coding, reasoning, and knowledge work, and introducing beta access to a 1 million token context window—the first for an Opus-tier model. The new version achieves leading scores on evaluations including Terminal-Bench 2.0, Humanity’s Last Exam, and GDPval-AA. On the API side, it introduces 'adaptive thinking,' control over reasoning 'effort level,' and context compression features to support long-running tasks and cost management. The model is now live on claude.ai, via API, and across major cloud platforms, priced at $5/$25 per million input/output tokens, with additional premiums for ultra-long contexts. Updates also include expanded research previews for Claude in Excel and PowerPoint.

OpenAI has introduced OpenAI Frontier, an enterprise-grade platform for building, deploying, and managing AI agents capable of performing real-world tasks. The platform enables agents to connect to external data and applications, and offers permission and behavior controls, along with management mechanisms resembling 'onboarding and feedback' to continuously refine agent performance. Pilot customers include HP, Oracle, State Farm, and Uber. Currently available to select users, broader rollout is expected over the coming months. This launch marks a strategic move by OpenAI into enterprise agent infrastructure, positioning it against ecosystems like Salesforce Agentforce, LangChain, and CrewAI. Pricing and detailed technical specifications remain undisclosed.

Kuaishou announced the global full launch of the Kling 3.0 series on February 5, including Kling Video 3.0, Kling Video 3.0 Omni, Kling Image 3.0, and Kling Image 3.0 Omni. The company highlights improvements in narrative and scene control, generation consistency, and fixes for issues like image collapse and text distortion. The Omni versions emphasize enhanced editing and inpainting capabilities, targeting workflows such as subject replacement and scene extension. Compared to prior closed testing, this 'full launch / official release' upgrade improves availability, enabling creators and enterprises to address both content generation and editing needs within a unified model family.

Voice AI company ElevenLabs has completed a new funding round, raising $500 million with a post-money valuation of approximately $11 billion, led by Sequoia Capital. Reports indicate cumulative funding since its 2022 founding has reached $781 million. This significant raise underscores sustained investor confidence in voice generation and voice interaction sectors, fueling expansion in model training, productization, and enterprise-grade voice infrastructure. Given that voice products often serve high-frequency use cases like content creation, customer service, and agent-based interactions, funds are expected to target compute, data, and distribution scaling, though specific allocation and revenue metrics were not disclosed.

According to BBC, Amazon announced plans to invest $200 billion in AI projects and related infrastructure this year, a substantial increase from last year's $125 billion, making it one of the most aggressive capital expenditure plans among tech giants. Following the announcement, Amazon's stock dropped about 10% in after-hours trading, reflecting market caution around high spending, return timelines, and profit pressures. The report notes Meta, Google, and Microsoft are also expanding AI investments, with an estimated combined $650 billion to be spent by these companies in 2026. This trend will continue to drive demand for compute power, data centers, and supply chains, intensifying competition in large models and cloud services.

Databricks' Mosaic AI research team has introduced the MemAlign framework into MLflow, aiming to reduce cost and latency in LLM-based evaluation systems during training and iteration. MemAlign employs a 'dual-memory' design: general evaluation principles are stored in semantic memory, while task-specific feedback in natural language is stored in a scalable vector database (episodic memory), reducing reliance on large annotated datasets and repeated fine-tuning. It supports feedback updates and deletions to prevent 'fixing one issue but creating more.' Testing shows performance comparable to using annotated datasets. Future integration with Agent Bricks' Judge Builder will enhance agent evaluation and governance.

TechCrunch reports that startup Sapiom has raised a $15 million seed round led by Accel, with participation from Okta Ventures, Array, Menlo, Anthropic, and Coinbase Ventures. The company aims to build financial infrastructure enabling AI agents to autonomously and securely purchase and access software, APIs, data, and compute resources—addressing the current bottleneck where agents heavily depend on manual configuration for external service authentication and payments. Focused initially on enterprise B2B use cases, Sapiom emphasizes simplifying 'backend integration' for non-technical developers building agent workflows. This direction is closely tied to agent scalability: as agents gain autonomy in procurement and authorization, enterprise capabilities in permission governance, auditable payments, and risk control become essential.

Media reports reveal that the AI chat app 'Chat & Ask AI' exposed approximately 300 million user conversation records due to a misconfigured backend Google Firebase database, affecting over 25 million users. Leaked data includes full conversation content, timestamps, user-defined bot names, and selected models—some samples contain high-risk topics such as suicide, drug manufacturing, and hacking. The app does not train its own models but acts as an interface to services from OpenAI, Anthropic, and Google, storing and managing conversations independently, making it a security weak point. This incident highlights that even when underlying model providers comply with regulations, third-party 'wrapper' applications may amplify privacy and security risks through poor data governance and default storage practices.