Claude Unleashes 1 Million Token Context, macOS Control for Developers

Key Takeaways

• Anthropic made its 1-million-token context window generally available for Claude Opus 4.6 and Sonnet 4.6 on March 13, 2026, at standard pricing — no more surcharges for large context requests.
• Developers can now feed entire codebases or document archives into a single prompt, reducing reliance on chunked RAG pipelines and multi-stage prompting for complex agentic workflows.
• A new “Computer Use” research preview, rolled out March 23, 2026, lets Claude directly control macOS interfaces — opening up autonomous, multi-step agent deployments that work with existing software rather than just APIs. Anthropic has quietly shipped one of the more practically useful Claude updates in months: a 1-million-token context window now available at standard pricing, plus a research preview that lets Claude actually drive a Mac like a human would. For anyone building agentic systems, both changes matter — and they’re available now.

1 Million Tokens: What It Actually Changes for Builders

The general availability of the 1-million-token context window for Claude Opus 4.6 and Sonnet 4.6 — live since March 13, 2026 — removes the premium surcharge that previously made large context requests expensive at scale. Roughly 750,000 words of input per prompt. The economics now work for production use.

The real win here is architectural simplicity. If you’ve been running chunked RAG pipelines to handle large codebases or document archives, you know the tradeoffs: added latency, retrieval misses, fragmented context. A 1M token window doesn’t replace RAG in every scenario, but for many workflows it removes the need entirely. What this unlocks in practice:

• Code Comprehension and Generation: Agents built on LangChain, LlamaIndex, or AutoGen can now ingest an entire repository in a single pass — understanding interdependencies and architecture without repeated retrieval calls. Coding assistants and autonomous dev agents get meaningfully better as a result.
• Legal and Financial Analysis: Lengthy contracts, regulatory filings, or financial statements become single-shot operations. Faster extraction, cleaner summarisation, fewer errors from missing context.
• Long-Horizon Agentic Tasks: Agents can carry full project context across extended tasks without explicit memory management. Less scaffolding code, better task persistence.

Anthropic also raised the media limit from 100 to 600 images or PDF pages per request when using the 1M token context. If you’re building multimodal pipelines that process mixed document types, that’s a significant lift in what a single request can handle.

Computer Use: Autonomous Agents That Work With Actual Software

“Computer Use” has been in the conversation for a while, but the macOS research preview — available to Pro and Max subscribers since March 23, 2026 — is the most deployable version yet. Claude can now open files, run dev tools, and navigate GUIs through clicks and keystrokes, the same way a human operator would.

For agentic builders, the key use case is bridging the integration gap. Not every system has a clean API. Legacy tools, niche enterprise software, internal dashboards — Computer Use lets you build agents that interact with these environments without writing custom connectors for each one. Concretely, this enables:

• Cross-Application Workflow Automation: Multi-step tasks across different apps — pulling data from a webpage, dropping it into a spreadsheet, generating a report — without an explicit API call at each step. Think n8n or Make.com-style orchestration, but the AI is driving the UI directly.
• Deeper Developer Tooling: Claude Code can now use Computer Use for more nuanced debugging, testing, and CI/CD interactions. The integration with Microsoft Office apps — including Opus 4.6 handling native pivot table editing in Excel — shows where Anthropic is pushing for embedded enterprise workflows.
• Legacy System Access: For systems without robust APIs, Computer Use is a practical workaround. Any software accessible via a graphical interface becomes fair game for agent automation.

It’s still a research preview, and Anthropic is clear that error rates are higher than text-based tasks. For repetitive, bounded workflows the utility is already real — but don’t deploy this unsupervised on anything consequential yet. The safety design is sensible: agents check in more frequently on high-stakes actions, and an auto-mode safety layer reviews each action before execution. That’s the right architecture for where this capability is today. If you’re thinking about how to structure that kind of oversight, the agent orchestration patterns are worth revisiting.

Model and API Updates Worth Knowing

Opus 4.6 and Sonnet 4.6

Opus 4.6 (released February 5, 2026) defaults to “adaptive thinking” — the model decides when deep reasoning is needed rather than running extended compute on every request. That’s a meaningful cost optimisation for production agentic systems. It also supports 128,000 maximum output tokens, double the previous limit, and a beta Compaction API that handles server-side context summarisation for effectively unlimited conversation length.

Sonnet 4.6 (February 17, 2026) brings improved agentic search with lower token consumption and extended thinking on par with Opus. For everyday coding assistance, content pipelines, and data analysis, Sonnet 4.6 hits a strong price-to-performance point — flagship capability without flagship costs.

API and Platform Updates

A few March 2026 API changes are worth integrating into your tooling. New capability fields — max_input_tokens, max_tokens, and a capabilities object — were added to the Models API on March 18, letting you programmatically check what each model supports. A thinking.display: "omitted" field, launched March 16, strips thinking content from streamed responses — useful for real-time applications where only the final output matters.

Automatic caching for the Messages API launched February 19. Add a single cache_control field; the system handles the rest, moving the cache point forward as conversations evolve. Less manual cache management, lower costs on repeated context.

Anthropic also announced a $100 million Claude Partner Network on March 12, 2026 — early API access, go-to-market support, and potential direct capital for startups building on Claude. Worth watching if you’re scaling a Claude-native product.

Agentic Features

Claude Code Security (February 20, 2026) scans codebases for vulnerabilities and flags targeted patches for human review — a sensible addition for AI-driven dev workflows where generated code needs a security pass before it ships.

Persistent memory, rolling out through March 2026, lets Claude carry project context, debugging patterns, and preferences across sessions with no setup required. That’s real workflow continuity, not just a feature checkbox.

Claude Dispatch, launched March 17, adds a persistent agent thread in Cowork — Anthropic’s desktop agent tool — so you can assign tasks from mobile and get notified on completion. Parallel coding workflows from Claude.ai on Anthropic-managed instances are also now available, giving teams more flexible deployment options. Taken together, these updates make Claude a more complete platform for production agentic development — not just a model you call, but infrastructure you build on. For more on AI agents and automation tools, visit our AI Agents section.

Originally published at https://autonainews.com/claude-unleashes-1-million-token-context-macos-control-for-developers/