Gemini 3.1 Pro doubles reasoning, Sonnet 4.6 ships, Figma gets MCP, Hermes Agent learns from itself
Google dropped Gemini 3.1 Pro with a reasoning score that more than doubled its predecessor. Anthropic shipped Sonnet 4.6. Figma launched their MCP server. And Nous Research released an AI agent that actually remembers and gets better over time.
Gemini 3.1 Pro
linkGoogle dropped Gemini 3.1 Pro on February 19 and it came in hot. 77.1% on ARC-AGI-2 - more than double the reasoning performance of Gemini 3 Pro. Same price too, no increase.
1M token context window. Works across text, audio, images, video, PDFs, and entire code repos. Available through the Gemini API, Vertex AI, the Gemini app, NotebookLM, and their new Gemini CLI.
This was Google's way of saying they're still in the race. While everyone was focused on Claude and OpenAI trading blows in February, Gemini quietly shipped a model that more than doubled its own reasoning capabilities at the same price point.
Claude Sonnet 4.6
linkAnthropic
Sonnet 4.6 dropped February 17 and it's surprisingly close to Opus on most benchmarks. 79.6% SWE-bench (vs Opus's 80.8%). 72.5% OSWorld (vs 72.7%). Users preferred it over Sonnet 4.5 about 70% of the time in Claude Code testing.
The wild part: users even preferred Sonnet 4.6 over the previous Opus 4.5. A "smaller" model beating the old flagship.
1M token context, 64K max output. Same adaptive thinking as Opus. It's now the default model on claude.ai at unchanged pricing ($3/$15 per MTok). For most everyday coding and writing tasks, Sonnet 4.6 is the sweet spot.
Figma MCP + Codex
linkFigma
Figma's official MCP server connects your design files directly to AI coding tools. Select a frame in Figma, and your IDE sees the components, variables, and layout. Not a screenshot - the actual structured design data.
The Codex integration (Feb 26 beta) makes it bidirectional. You're not just pulling from Figma into code, you can push code changes back into the design context. Design and code stay in sync. The thing designers and engineers have been fighting about forever.
Most common workflow: designer finishes a component, engineer opens their IDE, AI pulls in the design context and generates a first pass. Engineer tweaks from there. For design system teams it's even better - tokens and variables map directly.
Works with Claude Code, Cursor, VS Code, Gemini CLI, Replit, and more.
Hermes Agent
linkNous Research
Nous Research dropped Hermes Agent on February 25 and it's the first agent that actually gets better the more you use it. Not in a marketing-speak way - it literally writes reusable skill documents when it solves hard problems so it never forgets how.
The persistent memory is the big deal. It remembers your preferences, your projects, your environment across every session. Most agents start fresh every time. Hermes builds a profile of you and your work over time.
It connects to Telegram, Discord, Slack, WhatsApp, Signal, and CLI from a single process. Runs on a $5 VPS or serverless infra that costs nearly nothing when idle. Single curl command to install, no prerequisites.
The skill generation system uses the agentskills.io open standard, so skills are searchable, shareable, and work across different setups. When the agent figures out how to deploy your specific stack, it saves that as a skill it can reuse.