Claude Code July 27: Anthropic Launches Claude Opus 5 at $5/$25 per Mtok as the New Default, v2.1.219 Adds 1M Context and Fast Mode, Claude Security Plugin Enters Beta

Anthropic launches Claude Opus 5: near-Fable 5 intelligence at half the price, $5/$25 per Mtok, default model in Claude Code

• On July 24 Anthropic released Claude Opus 5 at $5/$25 per Mtok, the same price as Opus 4.8, with a Fast mode at 2.5x speed, and made it the default on Claude Max the same day.
• TechCrunch reports it outperforms Fable 5 on several benchmarks at half the price, with lighter safety restrictions and new Automatic Fallbacks.
• By July 25, independent benchmarks had Opus 5 winning SWE-bench Pro at 79.2% versus 64.6% for GPT-5.6 Sol, while Sol Ultra took Terminal-Bench at 91.9%.

Editor’s read: Half the Fable 5 price for near-Fable intelligence resets the default-model math: start agents on Opus 5 and escalate to Fable 5 only when a task proves it needs it.

Anthropic TypeScript SDK v0.115.0 adds the claude-opus-5 model, tool_change streaming events and server-side fallback credits

• Released July 24 alongside the Opus 5 launch: the claude-opus-5 model id, new tool_change streaming events, and server-side fallback credits for Automatic Fallbacks.

Anthropic TypeScript SDK v0.114.0 adds the model_context_window_exceeded stop reason

• Released July 23: a new stop reason, model_context_window_exceeded, so your code can distinguish a context overflow from an ordinary end_turn.

Anthropic TypeScript SDK v0.113.0: Managed Agents model effort, initial session events, threads delta streaming

• Released July 22: model effort control for Managed Agents, initial session events, and delta streaming for threads.

Anthropic TypeScript SDK v0.112.5 and v0.112.4: new API refusal category, AWS Bedrock withOptions() fix

• July 20-21 patch releases: v0.112.4 fixes AWS Bedrock options and auth mode being dropped across withOptions(), and v0.112.5 adds support for a new API refusal category.

Claude Code v2.1.219 makes Claude Opus 5 the default Opus model with 1M context, Fast mode at $10/$50 per Mtok, sandbox strictAllowlist

• Released July 24, the launch-day CLI: Claude Opus 5 becomes the default Opus model with 1M context and a Fast mode priced at $10/$50 per Mtok, plus a sandbox strictAllowlist setting and subagents that now nest to depth 3.

Editor’s read: This is the version that flips your default model. Update deliberately and re-check any model pins in scripts and CI.

Claude Code v2.1.220 ships latest bug fixes and reliability improvements

• Released July 25, a day-after-launch stabilization release with bug fixes and reliability improvements.

Claude Code v2.1.218: /code-review runs as a background subagent, MCP connection error details in /mcp, Windows path corruption fix

• Released July 22: /code-review now runs as a background subagent so it no longer blocks your session, /mcp surfaces MCP connection error details, and a Windows path corruption bug is fixed.

Claude Code v2.1.217 caps subagent concurrency at a default of 20 and fixes Windows auto-update

• Released July 21: a subagent concurrency cap (default 20) to stop runaway parallel spawns, emoji shortcode autocomplete, and a Windows auto-update fix.

Claude Code v2.1.216 adds the sandbox.filesystem.disabled setting and fixes long-session slowdown and worktree git escapes

• Released July 20: a sandbox.filesystem.disabled setting, a fix for long-session slowdown, and worktree isolation plus git escape fixes.

Claude Agent SDK for Python v0.2.124 security fix blocks the BatBadBut Windows command-injection class

• Released July 20: blocks Windows .bat/.cmd CLI spawning (the BatBadBut command-injection class) and rejects cmd.exe metacharacters.
• If your agents run on Windows, treat this as a mandatory upgrade.

Editor’s read: A command-injection class fix in an agent SDK is exactly the kind of patch to apply the week it lands, not the quarter.

Claude Agent SDK for Python v0.2.127 fixes premature stdin closure that broke background-task tool calls

• Released July 24: fixes a premature stdin closure that broke background-task tool calls.
• If background tasks were dying silently in your Python agents, this is the fix.

Claude Agent SDK for Python v0.2.126 adds terminal_reason on ResultMessage and typed model_usage

• Released July 22: terminal_reason on ResultMessage, typed model_usage for per-model accounting, and the bundled Claude CLI moves to 2.1.218.

Claude Agent SDK for Python v0.2.128 bundles Claude Code CLI v2.1.220

• Released July 25: bundles the v2.1.220 CLI, bringing the Opus 5 default and the week's sandbox hardening to Python agents.

Claude Agent SDK for TypeScript v0.3.219: cancel_queued interrupts, DirectoryAdded hook, fast_mode_disabled_reason

• Released July 24 alongside the Opus 5 launch: cancel_queued interrupts, a DirectoryAdded hook, and a fast_mode_disabled_reason field so you can tell why Fast mode is unavailable.

Claude Agent SDK for TypeScript v0.3.218: background skill dispatch flag, mid-stream 429/529 error status, per-model billing metadata

• Released July 22: a background skill dispatch flag, mid-stream 429/529 error status so rate limits surface during streaming, and per-model billing metadata.

Claude Agent SDK for TypeScript v0.3.216: tool_result_meta sidecar for denied and cancelled tool calls

• Released July 20: a tool_result_meta sidecar for denied/cancelled tool calls, a rewind skippedLinks count, and request-latency fields.

MCP TypeScript SDK 2.0.0-beta.5 aligns with the final 2026-07-28 spec revision: serverInfo moves to _meta, clientInfo now optional

• Released July 21, the core and server packages both hit 2.0.0-beta.5, aligning with the final 2026-07-28 wire revision: serverInfo moves to _meta, clientInfo becomes optional, and the Protocol base-class export is restored.
• The spec revision lands July 28, so this is the week to update.

Editor’s read: If you maintain an MCP server, upgrade before July 28 and re-test the initialize handshake, since serverInfo moving to _meta is a breaking wire change.

Claude Security plugin for Claude Code enters beta: scan changes for vulnerabilities before commit

• Announced July 22: the Claude Security plugin beta scans your changes for vulnerabilities before commit and can run full codebase scans from the terminal.

Claude Cowork adds Record a skill: screen-record a task and Claude turns it into a reusable skill

• Announced July 21 for Pro, Max and Team: record yourself performing a task on screen and Claude converts the recording into a reusable skill.

Nader Dabit: SWE-1.7 Lightning at 1,000 tok/s is 1,200% faster, while Grok 4.5 and Claude Opus 5 sit alone on the Pareto frontier

• On July 25 Nader Dabit reported SWE-1.7 Lightning running at 1,000 tok/s, 1,200% faster with little intelligence tradeoff, and placed Grok 4.5 and Claude Opus 5 alone on the price-intelligence Pareto frontier.

Greg Isenberg: 20 reasons this is the best time to build, from $20/mo superintelligence to LLM search minting the next millionaires

• On July 21 Greg Isenberg listed 20 reasons this is the best time to build: AI agents outnumbering humans, $20/mo superintelligence, and LLM search minting the next wave of millionaires.

levelsio: AI coding equalized everyone, apps that took a year now take hours, and speed is the last differentiator

• On July 20 levelsio argued AI coding has equalized developers: apps that took a year now take hours, speed is the last differentiator, and only AI companies are winning.

Claude Code creator Boris Cherny: Claude Opus 5 is Anthropic's least prompt-injectable model, near-zero attack success with Auto Mode defenses

• On July 24 Boris Cherny said Claude Opus 5 is Anthropic's least prompt-injectable model yet, and that with Auto Mode defenses enabled the measured attack success rate drops to roughly zero.

Editor’s read: Prompt-injection resistance is becoming a model-selection criterion for agents that touch untrusted content, not just a research metric.