Gemini 3.1 Pro preview rolls out as Claude API adds prompt caching

Google introduced Gemini 3.1 Pro on February 19, 2026, and began a preview rollout for developers, businesses, and consumers.¹² The rollout covers the Gemini API, Vertex AI, and Gemini Enterprise, as well as the Gemini app and NotebookLM.¹² Google describes the model as an upgrade for complex tasks and says it also underpins the recent Gemini 3 Deep Think update.¹

Gemini 3.1 Pro scored 77.1% on the ARC-AGI-2 reasoning benchmark, more than twice the score of Gemini 3 Pro, according to Google.¹ That benchmark result does not establish an equivalent improvement for every business task. Google says it will validate the changes during the preview before making the model generally available.¹ Businesses can test it now, but should account for its preview status when planning production use.¹²

Anthropic introduced automatic prompt caching for the Claude API on February 19, 2026.¹ Developers can enable it with a cache_control field; the API then places the cache breakpoint instead of requiring developers to choose its position manually.¹² The feature is described for the Messages API and aims to reduce the cost and processing time of repeated inputs.¹²

Previously, developers had to place cache breakpoints themselves.¹² The new option simplifies applications that repeatedly send long, unchanged prompt prefixes.³² The price benefit applies to cache hits, not every request: writes have an additional cost, and changes to a prompt prefix can prevent a hit.³ No general latency figure is available for the automatic caching feature.³

OpenAI introduced “OpenAI for India” on February 18. It reached an agreement with Tata Group and Tata Consultancy Services (TCS) to develop AI-ready data center capacity in India. The planned initial capacity is 100 megawatts. OpenAI also announced university partnerships for ChatGPT Edu and additional locations alongside its existing New Delhi office. The new offices are expected to open later in 2026.¹²

By combining local infrastructure, education partnerships, and a larger on-the-ground presence, OpenAI is deepening its position in India. The planned capacity could eventually grow to as much as 1 gigawatt, but that expansion is not available capacity today.¹ Companies should distinguish the agreement from an operational hosting option: it describes infrastructure to be built, not computing capacity they can use immediately. Reuters confirmed the Tata agreement and its planned 100-megawatt starting point.²

WordPress.com has introduced the WordPress AI Assistant inside its editor and media workflow.¹ Users can ask it to change page layouts and structure, generate or refine text, and create or edit images.¹ The feature is available on eligible paid WordPress.com plans.¹

The assistant extends WordPress.com's AI tools beyond building a new site to editing one over time.¹ It also works in block notes in the newer editor, where users can invoke it with “@ai.”² The announcement applies to WordPress.com; it is not a general update for self-hosted WordPress installations.¹ That distinction matters for companies assessing whether the feature fits their existing websites and publishing workflows.

On February 18, 2026, Cursor detailed sandboxing for local coding agents on macOS, Linux, and Windows. The company said it had been rolling out the feature over the preceding three months. Agents can execute tasks within defined boundaries without repeated approvals. They must request permission to cross those boundaries, most often to access the internet.¹

For businesses, the key question is which actions an agent can take independently and when it must ask. Cursor also offers access controls that enterprises can enforce.² The company reports 40% fewer interruptions than with unsandboxed agents; that does not establish a comparable productivity gain.¹ The local sandbox applies to agents on a user's machine. It is separate from Cursor's cloud agents, which run in their own virtual machines.¹³

Google announced new features for NotebookLM-generated presentations on February 17, 2026. Users can revise individual slides with text instructions and download decks as PPTX files for PowerPoint.¹² The rollout began in phases. Accounts differ on availability: one names Google AI Pro and Ultra subscribers as the first recipients, while another says access expanded to all users by February 20.³²

Previously, generated decks could only be downloaded as PDFs, making changes and reuse in presentation software less convenient.³⁴ The new revision feature lets users request changes to a specific slide’s text, images, or layout.³² PPTX export provides a way to continue working on a deck outside NotebookLM. Direct export to Google Slides is not yet available, though Google says it is coming.¹²

Tavus introduced Phoenix-4 on February 18, 2026. The model generates head-and-shoulders video for AI avatars in live conversations and is designed to adapt facial expressions and behavior to the exchange.¹² Tavus says it combines emotional-state generation, active listening, and continuous facial motion. It can listen and respond at the same time.¹ Phoenix-4 is available through the Tavus platform and APIs; users can create custom avatars or use prebuilt replicas.¹²

For companies deploying video assistants, the notable feature is continuous visual response: an avatar is meant to appear attentive even when it is not speaking.¹² Tavus positions Phoenix-4 as a successor to its earlier avatar models.¹² The published information does not establish how reliably it performs under real-world network conditions or how it compares with competing systems. Tavus also does not provide pricing in its announcements.¹²