Claude Sonnet 4.6 drops with 1M context, 15-point perf jump, and GitHub Copilot integration

Good morning and welcome to the Builder's Briefing for February 18th, 2026. I'm Alex, here with Sam, and wow — today is packed. We've got a major Anthropic drop, a scrappy challenger model that's absurdly cheap, and the agentic coding stack is really starting to crystallize.

Yeah, this is one of those days where you look at the news and realize the tools we're building with are shifting underneath us in real time. Let's get into it.

Alright, the big story. Anthropic shipped Claude Sonnet 4.6 yesterday — a fifteen-point improvement on complex work evaluations over the previous version, a one million token context window, and it's live in GitHub Copilot right now at standard premium pricing. No surcharge.

That's a big deal — especially the Copilot integration on day one. Usually there's a lag before new models show up in third-party tools. This feels like Anthropic is really trying to be the default for developers who ship code every day.

Exactly. And it wasn't just the model. They also GA'd code execution, memory, and tool search in the Claude API, plus added dynamic filtering for web search. So this is a coordinated platform push — better model, better tooling, broader distribution, all at once.

Right, and what's wild is the practical implication. If you're building on the API, one million tokens of context plus code execution plus web search means you can point an agent at an entire codebase and pull live data in a single workflow. No stitching together three different services.

If you're using Copilot, honestly just switch to Sonnet 4.6 today and test it. The agentic coding improvements are real.

Agreed. Though I will say — there's competition nipping at their heels, which brings us to the next story.

So Kimi K2.5 — DHH posted a side-by-side showing it fixing a bug in twenty-one seconds versus Claude's three minutes. And the cost? Roughly zero point three cents per million tokens. You could run three million tokens for under ten bucks.

That's insane. Like, for high-volume agentic loops where you don't need peak reasoning — just fast, cheap task completion — that could cut your inference bill by ninety percent or more. Not everything needs the smartest model in the room.

Right, and this is the market dynamic to watch. Anthropic competing on developer experience and capability at the top, while models like Kimi K2.5 are commoditizing the bottom end at speeds that are hard to ignore.

It's the classic stratification. Premium reasoning versus commodity execution. Smart teams are going to route different tasks to different models. That's just good engineering.

A couple other model drops worth mentioning. Cohere released Tiny Aya — a three point three five billion parameter multilingual model covering over seventy languages. Open source, small enough to run on modest hardware.

That's huge for anyone building for non-English markets. Languages in Africa, Southeast Asia, the Middle East — areas where bigger models just don't perform well. And no API dependency, you can run it yourself.

Also, Kling 3.0 landed on Replicate — 4K multi-shot video generation with audio, up to fifteen-second clips, all API-accessible. If you've been building video features, worth benchmarking against Runway or Pika.

Okay, shifting to developer tools. The Next.js team published agent-first design principles and an MCP integration guide, and I think this one is sneakily important.

This is the one that made me sit up. They're basically saying: your next power user isn't a person clicking through your UI, it's a coding agent with an MCP connection. So design your app accordingly — structured outputs, agent-friendly routing, the whole nine yards.

And that ties into the Claude for Excel news too — the Excel add-in now pipes in financial data from S&P, FactSet, PitchBook, and more via MCP. MCP is becoming the real integration layer for enterprise data. It's not just a dev toy anymore.

The through line is clear. MCP is the connective tissue for this whole agentic stack.

Two more quick ones on the tools front. Linear shipped a Cursor plugin — you can pull issues, update status, manage tasks without leaving the IDE. And there's a trending GitHub repo called obra slash superpowers — an agentic skills framework with thirty-six hundred stars. It's like a structured playbook for how teams should work with coding agents.

The Linear-Cursor thing is perfect. That's the IDE-as-hub trend continuing. You never leave your editor. And superpowers — I've been looking at that — it's the closest thing to a 'here's how we actually work now' methodology guide for AI-augmented teams.

Oh, and there's a new arxiv paper asking whether AGENTS.md files actually help coding agents. The Hacker News discussion had mixed results.

Ha — so maybe don't spend your whole Saturday crafting the perfect AGENTS.md just yet. Read the paper first. Link in the briefing.

Alright, infrastructure. Supabase made two moves. First, they acquired the Hydra team, which means they're getting pg_duckdb — the plan is to let you run analytical queries directly in Postgres without shipping data to a separate warehouse.

That's interesting because it kills a whole category of ETL pipelines. If you're on Supabase and you've been piping data out to BigQuery or Snowflake just for analytics, that pain might just go away.

And second, they turned on Row Level Security by default on new tables created via the dashboard. Small change, massive impact. That was the most common security footgun — devs forgetting to enable RLS and accidentally exposing data.

Secure by default. Love it. That's the kind of change that prevents incidents you never even hear about.

Quick hits to close us out. Google I/O is confirmed for May 19th — mark your calendars. BarraCUDA is an open-source CUDA compiler that targets AMD GPUs, so you can run CUDA code on AMD hardware without rewriting it. And there's a new open-source voice-to-text tool called FreeFlow as an alternative to Wispr Flow.

BarraCUDA is one to watch. If you're priced out of NVIDIA or want AMD as a fallback for inference, that's a real lifeline. Active discussion on Hacker News about real-world viability — link in the briefing.

So the takeaway for today — the agentic coding stack is crystallizing fast. Sonnet 4.6 in Copilot, Linear in Cursor, superpowers as methodology, MCP as the integration layer. If you're building developer tools, design for agent consumption first.

And if you're running high-volume agent loops, seriously benchmark Kimi K2.5 against whatever you're using now. At three tenths of a cent per million tokens and eight times the speed on simple tasks, you might be leaving a lot of money on the table.

Your next power user isn't a human with a browser — it's a coding agent with an MCP connection. Build for that world. That's the Builder's Briefing for February 18th. All the links and details are in the show notes.

Go switch Copilot to Sonnet 4.6, benchmark Kimi, and read that Next.js agent guide. Plenty to build on this week. See you tomorrow!