TL;DR:
- Claude Opus 4.8, released at Anthropic’s Code with Claude 2026 event, scores 69.2% on SWE-Bench Pro — a substantial jump from previous versions and the highest score among available models at time of release.
- Dynamic Workflows lets Claude Code spin up parallel subagents for independent sub-tasks, then aggregate results — significantly speeding up multi-file refactors and large codebase changes.
- The new security plugin covers 25 vulnerability classes including injection, SSRF, authentication flaws, and insecure cryptography — and integrates directly into the coding flow rather than as a separate scan step.
Anthropic ran its Code with Claude event in early 2026, and the updates were substantial. Claude Code had been improving incrementally since launch, but this release represents a more significant step change — both in raw model capability and in the workflow tooling around it.
Claude Opus 4.8
The headlining model update. Claude Opus 4.8 scores 69.2% on SWE-Bench Pro, the benchmark used to measure how well AI models resolve real GitHub issues. For context, GPT-4o at time of writing sits below 50% on the same benchmark; earlier Claude versions topped out around 55–60%.
What SWE-Bench Pro measures: given a real GitHub issue description and the surrounding codebase, does the model produce a patch that resolves the issue and passes the test suite? It’s a closer proxy to real software engineering work than many synthetic benchmarks, because it requires understanding existing code, reasoning about intended behaviour, and producing changes that integrate cleanly.
In practice, the improvement shows most clearly in:
- Multi-file changes: Opus 4.8 is better at tracking interdependencies across files and making consistent changes across a codebase
- Ambiguous specifications: given an underspecified problem, it makes reasonable inferences rather than asking clarifying questions or producing something obviously wrong
- Test generation: the quality of tests it generates alongside new code has improved noticeably
One caveat: SWE-Bench Pro scores measure average performance across a large benchmark dataset. Individual tasks vary significantly — Opus 4.8 can still produce wrong or incomplete solutions, particularly for highly domain-specific or unusual codebases.
Dynamic Workflows
The most practically significant addition for complex coding tasks. Dynamic Workflows lets Claude Code decompose a high-level task into parallel sub-tasks, execute them simultaneously using multiple subagents, and then aggregate the results.
The workflow structure:
- You give Claude Code a high-level task (“refactor the authentication module to use the new token interface”)
- Claude Code analyses the codebase and identifies independent sub-tasks (update the login handler, update the token refresh logic, update tests, update documentation)
- Parallel subagents work on each sub-task simultaneously in separate contexts
- Claude Code aggregates the results, resolves any conflicts, and produces a coherent set of changes
For large codebases or tasks that touch many files, this reduces end-to-end time substantially — a refactor that might take 15 minutes sequentially can complete in 4–5 minutes with parallel subagents.
The practical limitation: sub-tasks need to be genuinely independent. If task B depends on output from task A, it can’t run in parallel. Claude Code handles this with a dependency graph — truly independent tasks run in parallel; dependent tasks run sequentially. The orchestration is mostly automatic, though you can inspect and modify the task graph before execution.
The Security Plugin
New in this release: a security scanning plugin that runs automatically on any code Claude Code produces or modifies. It covers 25 vulnerability classes including:
- SQL, command, and LDAP injection
- Server-side request forgery (SSRF)
- Insecure direct object references
- Authentication and session management flaws
- Insecure cryptographic implementations
- Hard-coded credentials
- Path traversal
- XML and JSON injection
The integration is non-blocking by default — Claude Code will flag a security issue inline, explain what the risk is, and suggest a fix, but it doesn’t halt execution waiting for you to resolve it. You can configure it to block in CI mode.
What makes this useful compared to running a separate SAST tool is that Claude Code can explain the vulnerability in context. Rather than “CWE-89: SQL injection at line 42,” you get an explanation of how the specific injection path could be exploited in your code, and a suggested fix that fits the surrounding codebase. The educational value alongside the detection is significant.
Redesigned Interface
The GUI has been overhauled. The most visible changes:
- Task timeline: a visual representation of what Claude Code is doing, including subagent activity in Dynamic Workflows — you can see the parallel execution happening in real time
- Diff preview: proposed changes are shown in a split-view diff before being applied, rather than being applied and then shown in git diff
- Context inspector: a panel showing what files and context are currently loaded, with the ability to add or remove context items before or during a task
The context inspector in particular addresses a long-standing friction point: it wasn’t always clear what Claude Code had “seen” during a task, which made it hard to diagnose why it was or wasn’t aware of certain files. The inspector makes this explicit.
Getting Started
Claude Code is available as a CLI (npm install -g @anthropic-ai/claude-code), a VS Code extension, and a JetBrains plugin. The Dynamic Workflows and security plugin features are available on all platforms as of the Code with Claude 2026 release.
Dynamic Workflows requires an Anthropic API key with a plan that supports concurrent requests — check Anthropic’s pricing page for current limits. The security plugin is included at all tiers.
If you’re using an older version of the CLI, running claude-code --version and then npm update -g @anthropic-ai/claude-code is all you need to get the latest.