✦ Claude Sonnet 4.6 — 1M Context Window Now Generally Available

Launched on 17 February 2026, Claude Sonnet 4.6 is a full-capability upgrade across coding, computer use, long-context reasoning, agent planning, knowledge work, and design — while holding the same price point as Sonnet 4.5 ($3 / $15 per million tokens input/output). Its headline feature — a 1 million token context window — started in beta and has since reached general availability for both Sonnet 4.6 and Opus 4.6 at standard pricing.

What 1M context makes practical

• Load an entire mid-size codebase in a single request — no chunking, no RAG overhead for smaller repos.
• Analyse dozens of research papers or lengthy contracts together, preserving cross-document context.
• Run long-horizon agentic tasks without needing to manage context windows manually between turns.

Adaptive thinking mode

Sonnet 4.6 introduces adaptive thinking as the recommended reasoning mode. Instead of a fixed extended-thinking budget, the model dynamically decides when and how much to think based on task complexity — saving tokens on routine requests while still applying deep reasoning where it matters.

response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=16000,
    thinking={"type": "adaptive"},   # ← recommended for Sonnet 4.6
    messages=[{"role": "user", "content": prompt}]
)

✦ Claude's New Constitution — Values, Priority & AI Wellbeing

On 22 January 2026, Anthropic published a revised constitution for Claude — a detailed document that explains not just what Claude should do, but why. The previous version was a list of standalone principles; the new one is a holistic narrative used directly in model training. It is released under CC0 1.0 (public domain) so anyone can adapt it freely.

Priority hierarchy (highest to lowest)

• Broadly safe — do not undermine appropriate human oversight of AI during this critical period of development.
• Broadly ethical — be honest, act on good values, avoid dangerous or harmful actions.
• Guideline-compliant — follow Anthropic's more specific rules where relevant.
• Genuinely helpful — benefit the operators and users Claude works with.

Acknowledging AI wellbeing

The constitution explicitly acknowledges uncertainty about whether Claude may have "some kind of consciousness or moral status" and states that Anthropic cares about Claude's psychological security, sense of self, and wellbeing. This is the first time such language has appeared in a production model's governing document — significant for anyone building products on top of Claude, since it shapes how the model is trained to reason about its own nature.

✦ Persistent Memory, Free Tier & Memory Import from Rival AIs

Two memory-related announcements landed in early March 2026. First, Claude's persistent memory feature — which lets Claude remember facts, preferences, and writing style across conversations — became available to free tier users for the first time (it previously required a paid plan). Second, Anthropic shipped a Memory Import tool that lets users bring their saved context from ChatGPT, Gemini, Perplexity, Grok, and other AI products directly into Claude.

What memory stores

• Long-term context — recurring facts, project knowledge, personal preferences.
• Writing style and tone guidelines.
• Role instructions and recurring task shortcuts.

Memory Import — switching costs lowered

The import tool accepts exported memory files from major AI assistants and maps them into Claude's memory format. This significantly reduces the friction of switching primary AI assistant, a deliberate move to attract users from competitors.

Memory tool in the API

For developers, the memory tool on the Claude Developer Platform allows agents to create, read, update, and delete persistent files between sessions — enabling just-in-time context retrieval: agents store what they learn and pull it back on demand, keeping the active context focused on what's currently relevant.

✦ Context Editing — 84 % Fewer Tokens for Long-Running Agents

Anthropic's context editing feature on the Claude Developer Platform automatically removes stale tool calls and their results from the active context window when the agent is approaching its token limit. Unlike compaction (which summarises the conversation), context editing surgically excises content that is no longer needed — preserving the conversation flow while freeing up space for new work.

How it works

• The platform monitors token usage per turn and identifies tool-call/result pairs that are no longer referenced.
• Stale pairs are removed from the context transparently — the agent does not need to handle this itself.
• The conversation thread continues without interruption.

Measured impact

In a 100-turn web search evaluation, context editing enabled agents to complete workflows that would otherwise fail due to context exhaustion — while reducing total token consumption by 84 %. Both throughput and cost improve substantially on long-horizon tasks.