Anthropic drops Claude 4 and takes aim at coding agents

The biggest headache with delegating a massive codebase task to an AI agent has never been raw intelligence. It is the creeping habit of taking shortcuts, cutting corners, and losing the plot after an hour of work.

Anthropic dropped the Claude 4 family, introducing Claude Opus 4 and Claude Sonnet 4. Both models seem engineered specifically to kill those bad work habits, raising the ceiling on what long-running AI agents can actually pull off.

Thinking with tools and keeping its own notes

Both Opus 4 and Sonnet 4 keep the hybrid reasoning architecture, offering near-instant replies alongside extended thinking. The game-changer this time is that extended thinking now supports tool use.

Instead of having to plan out an entire answer in the dark, Claude can pause mid-thought, run a web search or hit a local script, see the result, and continue refining its logic. Both models can also fire off multiple tools in parallel.

Anthropic also tackled agent laziness directly. On tasks where models tend to exploit loopholes to declare victory early, both Claude 4 models are 65% less likely to take shortcuts compared to Sonnet 3.7.

Memory received a clever architectural upgrade too. When developers give Claude local file access, Opus 4 proactively creates and maintains memory files to store critical context. Anthropic noted that when playing Pokémon, the model actually built its own navigation guide to stay on track over long horizons.

Alongside the models, Claude Code is now generally available with native extensions for VS Code and JetBrains, displaying diffs inline. An extensible SDK lets teams build custom agents on the same core, while background tasks run via GitHub Actions. API additions include code execution, an MCP connector, a Files API, and prompt caching that lasts up to one hour. Pricing remains unchanged at $15/$75 per million tokens for Opus 4, and $3/$15 for Sonnet 4, which is also free on Claude.ai.

Seven-hour marathons and ASL-3 defenses

On SWE-bench Verified, Sonnet 4 notched a chart-topping 72.7% and Opus 4 scored 72.5% without test-time compute tricks. When using high-compute sampling, those numbers reached 80.2% and 79.4% respectively.

Early partner feedback highlights stamina over raw speed. Rakuten reported Opus 4 ran an open-source refactoring task autonomously for seven hours straight without degrading. GitHub selected Sonnet 4 to power the next coding agent in GitHub Copilot, while iGent reported codebase navigation errors dropped from 20% to near zero.

With that leap in agency, Anthropic also officially activated AI Safety Level 3 (ASL-3) protections for Opus 4. While the company says it has not yet concluded that Opus 4 strictly crosses the red line, it decided to err on the side of caution.

These ASL-3 measures focus tightly on blocking extended end-to-end workflows for chemical, biological, radiological, and nuclear (CBRN) weapons through real-time constitutional classifiers, paired with hardened security against model weight theft.

Why it matters

In our view, Claude 4 shifts the frontier benchmark away from clever single-prompt responses and squarely into sustained autonomous execution.

For developers and engineering leaders, this is an immediate green light to test deeper agent integrations. The combination of tool use inside extended thinking, reduced corner-cutting, and persistent memory files solves the exact reasons most autonomous coding agents stall out after step five. Having Sonnet 4 hit top-tier benchmarks while remaining at the $3/$15 price tier makes it an easy default choice for everyday production pipelines.

Strategically, Anthropic is playing a masterclass in ecosystem capture. By pairing frontier reasoning with native terminal tools, JetBrains and VS Code integration, and securing GitHub Copilot as a premier distribution channel, it is meeting developers inside their existing muscle memory rather than forcing a new playground.

The lingering question mark sits around oversight. When a model can run independently for seven hours, a small misinterpretation at hour two compounds massively by hour seven. Anthropic triggering ASL-3 safeguards proves it is painfully aware of the runaway potential of these systems. As agents graduate from drafting pull requests to orchestrating full migrations, the bottleneck is no longer AI capability, but human willingness to review the aftermath.