Builder's Briefing — February 18, 2026
Claude Sonnet 4.6 drops with 1M context, 15-point perf jump, and GitHub Copilot integration
Anthropic shipped Claude Sonnet 4.6 yesterday, and the numbers matter: a 15-point improvement on complex work evaluations over its predecessor, a 1M token context window, and immediate availability in GitHub Copilot at 1x premium pricing (no surcharge). The model is live in the API right now. Alongside the model launch, Anthropic also GA'd code execution, memory, and tool search in the Claude API, plus added dynamic filtering for web search that uses code execution under the hood. This is a coordinated platform push — better model, better tooling, broader distribution.
For builders, the practical upside is concrete. If you're using Copilot, switch to Sonnet 4.6 and test it on your codebase today — the agentic coding improvements plus strong search capabilities make it the most capable mid-tier option in the Copilot model picker. If you're building on the API, the combination of 1M context + GA code execution + web search filtering means you can build agents that process entire codebases and pull live data in a single workflow without stitching together separate services.
The signal for the next 6 months: Anthropic is competing hard on the developer experience layer, not just benchmarks. GA'ing tools alongside the model drop, landing in GitHub Copilot on day one — this is about making Claude the default for builders who ship. Meanwhile, Kimi K2.5 is nipping at their heels on speed and cost (DHH clocked it fixing a bug in 21 seconds vs. Claude's 3 minutes, at $0.003/M tokens). The mid-tier model market is about to get very competitive very fast.
Kimi K2.5 is absurdly fast and absurdly cheap — 3M tokens for under $10
DHH posted a side-by-side showing Kimi K2.5 fixing a bug in 21 seconds vs. Claude's 3 minutes, at roughly $0.003 per million tokens via OpenCode Zen. If you're running high-volume agentic loops where speed and cost matter more than peak reasoning, this is worth benchmarking against your workload immediately.
Claude for Excel gets MCP connectors for S&P, FactSet, PitchBook, and more
The Excel add-in now pipes in financial data from six major providers via MCP. If you're building fintech tools or internal dashboards, this is the signal that MCP is becoming the real integration layer for enterprise data — not just a dev toy.
Cohere releases Tiny Aya: 3.35B multilingual model covering 70+ languages
Open-source, small enough to run on modest hardware, and covers languages that larger models handle poorly. If you're building for non-English markets — especially in Africa, Southeast Asia, or the Middle East — this is the most practical on-ramp for multilingual features without API dependency.
Kling 3.0 lands on Replicate: 4K multi-shot video with audio, 15s clips
4K video generation with style transfer and audio is now API-accessible on Replicate. If you've been building video features with Runway or Pika, test this — the multi-shot coherence and audio integration could simplify your pipeline significantly.
Next.js publishes agent-first design principles and MCP integration guide
The Next.js team laid out their architecture recommendations for building apps that AI agents can use — MCP integration, structured outputs, agent-friendly routing. If you're building a SaaS product, this is worth reading now because your users' AI agents are about to become your API consumers.
obra/superpowers: An agentic skills framework trending hard on GitHub
With 3,600+ stars, this framework codifies a methodology for building with coding agents — think structured prompting patterns and skill decomposition. If you're managing a team that ships with AI agents, this is the most practical "how we work now" playbook currently available.
Linear ships a Cursor plugin for in-IDE project management
You can now pull Linear issues, update status, and manage tasks without leaving Cursor. This is the IDE-as-hub trend continuing — if your team uses Linear + Cursor, turn it on today.
Research: Do AGENTS.md files actually help coding agents?
New arxiv paper evaluates whether project-level instruction files improve agent performance. The HN discussion (58 points, 17 comments) suggests mixed results — worth reading before you spend a day crafting the perfect AGENTS.md for your repo.
OpenAI Codex terminal agent hits 950 stars on GitHub
Lightweight terminal-based coding agent from OpenAI. If you want an agentic coding flow without IDE lock-in, this is a clean starting point that plays well with existing CLI workflows.
Go's 'go fix' gets a deep dive for automated code modernization
Official Go blog post walks through using go fix to mechanically update your codebase to modern idioms. If you maintain a Go project, this is free tech debt reduction — run it on your CI pipeline.
BarraCUDA: Open-source CUDA compiler targeting AMD GPUs
Run CUDA code on AMD hardware without rewriting it. If you're priced out of NVIDIA GPUs or want AMD as a fallback for inference, this is the project to watch — 117 points on HN with active discussion on real-world viability.
Async/Await on the GPU — new programming model for GPU workloads
Vectorware's blog post lays out how async/await patterns can work for GPU compute. If you're writing custom CUDA kernels or building inference infrastructure, this could reshape how you think about GPU task scheduling.
Testing Postgres race conditions with synchronization barriers
Practical technique for reproducing and testing concurrent access bugs in Postgres. If you've ever shipped a "works on my machine" fix for a race condition, this approach gives you deterministic reproduction in your test suite.
Supabase acquires Hydra team, gets pg_duckdb for Postgres analytics
The pg_duckdb maintainers are now at Supabase building an "Open Warehouse Architecture." If you're using Supabase and piping data to a separate analytics warehouse, this acquisition means you'll eventually be able to run analytical queries directly in Postgres — killing a whole category of ETL pipelines.
Supabase enables Row Level Security by default on new tables
New tables created via the dashboard now have RLS enabled automatically. Small change, big impact — this eliminates the most common Supabase security footgun where devs forget to enable RLS and accidentally expose data.
Turso ships as an in-process SQLite-compatible database on GitHub
If you want embedded SQLite with replication and branching, Turso's open-source repo is trending. Worth evaluating if you're building local-first apps or need edge-compatible SQL.
dev/agents exits stealth as Dreamer — full-stack coding agent
New coding agent startup announced by swyx. Another entrant in the crowded agentic coding space, but the full-stack approach (not just code completion) is the direction the market is moving. Worth watching to see if they differentiate on execution quality.
Bria's video background removal model goes live on Replicate
Production-ready video background removal via API. If you're building video editing features or real-time video apps, this saves you from running your own segmentation model.
The agentic coding stack is crystallizing fast: Sonnet 4.6 in Copilot, Linear in Cursor, obra/superpowers as methodology, AGENTS.md under academic scrutiny. If you're building developer tools, design for agent consumption first (see the Next.js MCP guide) — your next power user isn't a human with a browser, it's a coding agent with an MCP connection. And if you're running high-volume agent loops, benchmark Kimi K2.5 against your current model — at $0.003/M tokens and 8x speed improvement over Claude on simple tasks, it could cut your inference bill by 90%+ on the commodity end of your workload.