Anthropic debuts Sonnet 4.6: long context, stronger computer use, and creative-friendly planning
Updated 13:00 EST / February 17, 2026
Anthropic rolled out Sonnet 4.6, a mid-tier Claude model that now brings a 1 million token context window (beta) to more users. It's the new default for Free and Pro plans, and pricing stays the same at $3/$15 per million input/output tokens. Performance is edging close to prior Opus levels, with big gains in computer use, long-context reasoning, and agent planning.
For creatives, this means larger projects, richer references, and fewer workarounds. Think: entire brand bibles, raw transcripts, multi-source research, and complex briefs-held in working memory at once.
Why this matters for your workflow
- 1M token context (beta): Load full style guides, campaign calendars, interview transcripts, and research decks without constant trimming.
- Computer and browser control: Automate multi-step tasks in Chrome, fill complex web forms, and work through spreadsheets and documents.
- Agent planning: Break big creative tasks into steps, then execute across files, folders, and the browser.
- Near-Opus quality at mid-tier cost: Better outputs without premium pricing.
Sonnet 4.6 sits inside the Claude family. If you want model guides and practical workflows, explore Claude.
Computer and browser use: real progress, clear guardrails
Anthropic has been training models to operate real apps-Chrome, LibreOffice, VS Code, and more. Users report human-like steps on tasks such as working through complex spreadsheets or completing multi-page web flows and then pulling data together across tabs.
It's not beating a skilled human yet, but it's catching up. Box CTO Ben Kus said Sonnet 4.6 outperformed Sonnet 4.5 by 15 percentage points on heavy reasoning Q&A across enterprise documents.
Anthropic also released Claude Cowork, a Mac app (Windows coming) that lets the model read files and control your mouse, keyboard, and browser. That enables multi-step work like organizing assets, editing documents, and structured browsing-useful for content cleanups, link sourcing, and research sprints. If you're exploring agent workflows, see AI Agents & Automation.
Safety and reliability
Full computer control raises obvious risks-prompt injection, hijack attempts, and hallucinations. Anthropic says Sonnet 4.6 improves resistance to manipulation and hallucination, scoring similarly to Opus 4.6 in internal safety evaluations. Still, human oversight is wise for any step that touches credentials, payments, or brand-critical assets.
Coding and long-context thinking (great for content systems)
Early testers of Claude Code report Sonnet 4.6 reads more context before it edits, consolidates logic instead of duplicating it, and avoids overengineering. With 1M tokens, it can reason across entire codebases.
For creatives, treat code like content systems: websites, design tokens, CMS templates, and automation scripts. It can also hold long contracts, research papers, or a season's worth of scripts and keep the threads straight as you iterate.
API and power features
- Adaptive and extended thinking (beta): Dial in cost-to-performance for bigger problems.
- Context compaction (beta): Auto-summarizes when the window fills, so you don't lose earlier decisions.
- Code execution, web fetch, memory, and tool calling: Now generally available, improving production use.
- Web search + fetch: Automatically writes and runs code to filter search results.
- Model Context Protocol in Excel: Available for Pro and above-handy for campaign sheets and reporting.
Pricing and access
- Default model: Sonnet 4.6 is now the default for Free and Pro.
- Pricing: $3 per million input tokens, $15 per million output tokens.
- Context: 1M token window available in beta.
Quick wins for creatives this week
- Brand brain: Drop your brand guidelines, voice rules, and recent campaigns into one session. Ask Sonnet to produce a multi-format content kit-emails, captions, landing page copy-with consistent voice and references.
- Research pack: Load a dozen sources (articles, notes, transcripts). Have it map insights, pull quotes, and draft outlines with citations you can verify.
- Content QA assistant: Use Claude Cowork to rename and sort assets, check links, and prep a clean folder structure before handoff.
- Website/system pass: Point it at your theme files or CMS templates to suggest cleanup and remove duplicated logic without breaking styles.
- Spreadsheet-heavy tasks: Let it fill forms, reconcile lists, and produce summary sheets for campaign performance or content calendars.
What to watch
Sonnet 4.6 still trails skilled humans in tricky, multi-app flows. Keep humans in the loop for critical reviews, sensitive data, and UI steps that can fail silently. The upside: pace of progress is strong, and many day-to-day tasks are already within reach.
Bottom line: More context, steadier reasoning, and growing computer use make Sonnet 4.6 a practical pick for creatives who juggle large references and repetitive steps. If you skipped earlier versions, this one is worth testing on real work.
Your membership also unlocks:
