Anthropic

Introduces a multi-agent “council” system, where several specialized agents collaborate instead of a single model response. Focuses on self-reflection and debate-style reasoning, aiming to improve answer quality through internal discussion. Designed for long-horizon, complex problem solving, especially where multiple perspectives are useful. Emphasizes higher reliability and reduced hallucination through cross-checking between agents. Positioned as a shift toward system-level intelligence, not just a single upgraded model.

Claude Sonnet 5 Release and access: ships as claude-sonnet-5-20260203, available via Anthropic API, Claude Pro, and Google Vertex AI. Coding benchmark jump: posts 82.1% SWE-Bench Verified, positioned above Claude Opus 4.5 at 80.9%. Pricing reset: priced at $3 per 1M input tokens and $15 per 1M output tokens, framed as a major cost drop vs Opus 4.5. Massive context: 1,000,000-token context window for repository-level understanding, whole-codebase prompts, and large refactors without chunking. Agentic engineering stack: built for multi-step workflows with built-in code execution and self-correction, plus a “Dev Team” mode that spawns parallel sub-agents for implementation, testing, and review.

Handles your hardest coding work — A big step up in software engineering. You can hand off complex, long-running tasks with confidence — the model even verifies its own outputs before reporting back. Follows instructions more precisely — Where earlier models interpreted things loosely or skipped parts, Opus 4.7 takes them literally. Sees images 3× better — Vision now supports up to ~3.75 megapixels, more than triple previous models. Great for dense screenshots, complex diagrams, and anything requiring fine visual detail. Better memory across sessions — Smarter about carrying context across long, multi-session work so you don't have to re-explain as much. Updated tokenizer — Improves text processing, but the same input can use ~1.0–1.35× more tokens. Cybersecurity safeguards — Automatically blocks prohibited cyber requests

Claude Sonnet 4.6 Better coding performance, especially on harder bug-finding and long-horizon codebase work. More reliable computer use, with fewer tool errors and better recovery when actions fail. Stronger long-context reasoning, with an optional very large context mode available in beta. Higher efficiency on real workloads, meaning it can achieve similar results with less back and forth and less wasted output. Better “knowledge work” quality, with improved document understanding, structured writing, and cleaner design taste for UI and reports. Claude Opus 4.6 Better financial reasoning (+5.47% on Finance Agent, SOTA on TaxEval) Improved information extraction (BrowseComp, DeepSearchQA) Gains on SWE-bench and agentic benchmarks 128K max output tokens (was 64K) Adaptive thinking replaces budget_tokens (deprecated) New "max" effort level; effort parameter now GA Server-side context compaction (infinite conversations) Data residency controls (inference_geo) Prefilling assistant messages removed Better prompt injection resistance Stronger alignment on sensitive tasks Better first-pass quality for spreadsheets, presentations, financial models Claude in Excel handles longer, more complex tasks