Anthropic AI Updates: July 25, 2026

1. Anthropic Introduces Claude Opus 5

Anthropic. Anthropic released Claude Opus 5, its new top-tier Opus model aimed at long-running agents, coding, and professional work. The company reported state-of-the-art results on coding benchmarks including Frontier-Bench and CursorBench, higher scores on ARC-AGI 3 reasoning, and gains on OSWorld 2.0 computer-use tasks, along with a stronger ability to check its own work. Pricing stays at $5 per million input tokens and $25 per million output tokens, matching Opus 4.8, with a faster variant offered at roughly double that rate. Source

2. Claude Code v2.1.219 Makes Opus 5 the Default and Deepens Subagent Nesting

Anthropic. Anthropic shipped Claude Code v2.1.219, which sets Claude Opus 5 (claude-opus-5) as the default Opus model with a 1M-token context window and a fast mode priced at $10/$50 per million tokens. The release lets subagents spawn nested subagents up to depth 3 by default, up from 1, and forwards text from those deeper subagents in stream-json when the forwarding flag is set. It also adds a sandbox.network.strictAllowlist setting to deny non-allowlisted hosts without prompting and a DirectoryAdded hook that fires when a new working directory is registered mid-session. Source

3. Anthropic Publishes New Context Engineering Rules for Claude 5 Models

Anthropic. A new Claude blog post describes how Anthropic removed more than 80 percent of Claude Code’s system prompt for advanced models such as Opus 5 and Fable 5 without a drop in performance. The guidance shifts from rigid rules toward trusting model judgment, favoring expressively designed tool interfaces over worked examples and progressive, deferred loading of context instead of front-loading everything. It also recommends lightweight CLAUDE.md files that focus on gotchas rather than details the model can discover on its own. Source

4. Anthropic Offers Guidance on Choosing the Right Claude Model

Anthropic. Anthropic published a guide on selecting a Claude model, recommending teams start with the most capable generally available model because it often completes tasks more efficiently despite a higher per-token cost. The post frames four decision factors, task complexity, latency requirements, access constraints, and unit economics, and maps them across the Mythos, Fable, Opus, Sonnet, and Haiku tiers. It advises testing candidate models with custom evaluations on real production workloads rather than relying on benchmarks alone. Source

5. Anthropic and Andon Labs Test Whether AI Can Control a Drone

Anthropic. Anthropic’s Frontier Red Team, working with Andon Labs, published Project Pilot, an evaluation of whether AI models can autonomously fly a drone through a locate-and-follow surveillance task. The work breaks drone control into five subtasks, reconstructing a 3D environment, localizing the drone, navigating rooms, detecting a target, and following them, and finds that newer models progress further on each, with Claude Fable 5 exceeding baseline on four of five but faltering on 3D reconstruction. The authors warn about the dual-use nature of the technology and the risk that human oversight could be treated as a cost rather than a safeguard once capability and reliability thresholds are crossed. Source

6. Anthropic Shares How Its Designer Uses Claude Design for Early Ideation

Anthropic. In a Claude blog post, Nate Parrott, the product designer who built Claude Design, described using the tool to keep pace with a fast-shipping engineering team by generating many design variations before committing to build anything. He uses it to produce clickable, brand-compliant prototypes, slide decks, landing pages, and animations that can be shared like documents for feedback. Parrott positions Claude Design as a medium for early exploration, collaboration, and buy-in, with production-level implementation remaining Claude Code’s domain. Source