This week in AI updates: Claude Sonnet 4.6, Gemini 3.1 Pro, and more (February 20, 2026)
Anthropic Release Claude Sonnet 4.6
Claude Sonnet 4.6 includes advanced skills in coding, computing, long-term contextual reasoning, agent programming, information work, and design.
It is now the default model on claude.ai and Claude Cowork, has a context window of 1M (beta), and has the same price as Sonnet 4.5, at $3 per million input tokens and $15 per million output tokens.
“Functionality that would have previously required access to an Opus-class model—including real-world, economic-critical office tasks—is now available with Sonnet 4.6. This model also features significant improvements in computing capabilities compared to previous Sonnet models,” Anthropic wrote in the post.
Gemini 3.1 Pro is now available for preview
Gemini 3.1 Pro is now available for Gemini API developers in Google AI Studio, Gemini CLI, Google Antigravity, and Android Studio. It can also be accessed from Vertex AI, Gemini Enterprise, Gemini operating system, and NotebookLM.
“Building on the Gemini 3 series, the 3.1 Pro represents a step forward in core logic. The 3.1 Pro is an intelligent base, better able to solve complex problems. This is reflected in our progress in rigorous benchmarks. In ARC-AGI-2, the benchmark that tests the ability of the model to solve 3% points confirmed 71% of the new Pro logic. This is more than twice the performance of the 3 logic Pro,” Google wrote in the post.
OpenAI adds Lock Mode, High Risk Labels to ChatGPT
These new features are designed to reduce the risk of rapid injection attacks.
Lockdown Mode limits the way ChatGPT can interact with external systems, reducing the chance of data leakage from a quick injection attack, while new Elevated Risk labels will be displayed on certain products to inform users that interacting with a certain feature may present an increased risk. For example, developers can give Codex network access to do things like look up documents online, but this extra access can also be dangerous. Currently, High Risk labels will be displayed in ChatGPT, ChatGPT Atlas, and Codex.
Microsoft creates a list of pre-built agents for Visual Studio
Prebuilt agents include the Debugger, which uses call stacks, dynamic state, and diagnostic tools to work through errors; Profile, which identifies issues and suggests improvements; Testing, which produces unit tests; and Modernize, which uses framework and dependency development.
“Each agent preset is designed around a specific developer’s workflow and integrates with native Visual Studio tools in ways that the standard assistant cannot,” Microsoft wrote in a blog post.
Agents can be reached through the chat panel by using the agent key or “@”.
GraphRAG enables more context-aware and verifiable responses from LLMs
Graphwise’s GraphRAG offering acts as a semantic layer on top of knowledge graphs that LLMs can use to provide rich and verifiable answers.
According to the company, the standard use of RAG makes the data fragmented, and thus, it can find similar words, but it cannot understand complex relationships, categories, or logical business data. In addition, it is often difficult to see how the LLM arrived at its answer and what sources it used.
Graphwise believes that GraphRAG solves these problems by providing a pipeline where all steps can be evaluated and the answers are supported by documentation and graph structures.
It uses several different search methods, including retrieval from a knowledge graph, vector search from a special vector store, and full search to enable keyword-driven discovery. It uses a knowledge model-driven input processing approach to understand the user’s intent, allow them to enrich concepts using an enterprise taxonomy or ontology, expand queries using related term entities, and construct a graph representing the query.
Checkmarx improves the security of the IDE-native agent application in Kiro
Agentic AI security provider Checkmarx has announced an integration with AWS Kiro IDE to enable developers working on that platform to identify and address security issues as code is written, the company said.
The integration puts Checkmarx Developer Assist directly in Kiro, so developers don’t have to leave the IDE to analyze code for security.
Once developers have used Developer Assist within Kiro and it has been validated, Checkmarx said the tool will analyze the source code and dependencies in the working environment. In addition, it said the tool will automatically display security findings in the IDE, along with contextual data that helps developers fix security issues earlier in the development cycle. That data can be viewed on the Checkmarx One platform, giving stakeholders an overview of project risks.
The Quest Trusted Data Management Platform makes it easy for organizations to create reusable data products
The Quest Trusted Data Management Platform integrates data modeling, data cataloging, data management, data quality, and data marketplace so organizations can deliver AI-ready data across their business.
“Building reliable AI-ready data and reusable data products can take six months, but your business can’t wait, so teams skip metadata, bypass governance flows, and ignore data quality, and every department ends up with their own version of the data product. That results in fragmented, aggregated data that is unreliable at Quest Software.
One of the key capabilities of the platform is the Automated Data Product Factory, which uses generative AI to create data products from natural language recognition, shortening data product design cycles, reducing delivery costs, and enabling business users to create their own data products.
