Starting today, new Claude Code sessions on Pro, Max, and Team plans default to auto mode. The classifier reviews each shell command, file operation, and agent action before execution and silently approves or holds it. The study Anthropic cites: 1,053 paid testers, 89% harmful-action catch rate for auto mode versus 13.6% for human review. Anthropic is also absorbing the classifier token cost, so auto mode no longer charges users for the extra inference.

The mechanism is a separate classifier that evaluates each tool call against a harm taxonomy before Claude executes it. A shell command that looks like rm -rf / gets held; a git commit does not. What the classifier does not publish is the taxonomy itself, which means developers cannot fully predict where the hold line sits. That ambiguity is a feature for Anthropic and a friction point for power users.

What the blast radius looks like in practice: every individual user on Pro, Max, or Team who starts a new session on August 14 or later lands in auto mode by default, with a one-time prompt to switch or stay. Organization-managed settings are untouched. Anyone who had already explicitly set manual mode is also untouched. The churn risk is real: developers working in high-stakes environments, production repos, financial systems, will want manual mode back, and the path there requires navigating settings.

The pattern is worth tracing. The August 14 changelog also shipped: GitLab merge request URL support for the --worktree flag (MRs now render as !N in claude agents view), an opt-in forward_user_identity gateway setting that sends signed-in user identity as headers on Anthropic upstreams for per-user spend attribution, and CLAUDE_CODE_TOOL_MEMORY_LIMIT to cap runaway Bash tool processes on Linux. These are enterprise plumbing features. They built the attribution and governance layer first, then flipped the default. The sequence was deliberate.

89% versus 13.6% is a concession buried inside an announcement. Anthropic is saying that the safety mechanism in every previous Claude Code session, your judgment, was worse than their classifier. They are probably right. The classifier runs on every action. You were approving prompts while mentally tabbing to something else. That is the honest version of what the data shows.

The builder's move: check your deny rules today. If you relied on prompts to keep Claude in bounds, prompts are not policy. The deny rules in your settings.json survive mode changes; prompts do not. For elevated-risk environments, set manual mode explicitly.

Against the field: OpenAI's Codex and Cursor run auto-approval by default. xAI's Grok in GitHub Copilot, also live today, has no mode toggle at all. Manual approval was a Claude Code UI choice, not a frontier standard. Anthropic brought itself in line today, with the data to justify it.

Anthropic published the full technical explanation of its text watermarking system on August 15, the follow-through to the August 11 announcement. The mechanism is a variant of Google DeepMind's SynthID-Text. At generation time, Claude uses a cryptographic key and the preceding token context to bias its next-token probability distribution in a statistically imperceptible way. No single token is forced; the bias accumulates across a paragraph and becomes detectable with the key. You cannot see it. It does not change the meaning, cadence, or readability of the text. The mark travels when you copy and paste. It may survive some editing.

Coverage: all Claude models launched on or after August 2, 2026, across Claude.ai, the Claude API, Claude Code, Claude Cowork, and Claude Tag. The rollout is global, not EU-scoped, despite the EU AI Act being the regulatory driver. Files receive C2PA-standard signed provenance metadata rather than the statistical token mark.

The blast radius for downstream operators is quiet but real. Legal, academic, and publishing workflows that rely on Claude to draft material now produce marked text. When Anthropic ships the detection API (announced, not yet available), any tool can query whether a document passed through Claude. Internal review processes will eventually need to account for that.

The pattern is compressed. Google published the SynthID-Text paper in 2024. OpenAI has watermarked DALL-E 3 image output since 2023. Text watermarking has lagged image watermarking by years across the industry. Anthropic is the first frontier text API to commit fully, with a planned detection API, a named technical approach, and a global deployment. It arrived because the EU required something; it is consequential because Anthropic extended it everywhere.

The contrast worth naming: Anthropic adopted a Google DeepMind technique before Google has rolled SynthID-Text out at the Gemini API level. Research and product are different timelines, and this is not a criticism. But it is notable that the originating lab has not yet deployed its own technique in production at the API layer.

Anthropic's framing of the limitation is honest: a detected watermark proves a piece of text may have passed through Claude. That is all. Every other AI-generated content detector has been quietly not saying the equivalent for two years. Anthropic said it on launch day. That is the screenshot line.

On August 14, xAI's Grok 4.6, released two days prior, rolled into GitHub Copilot's model picker across eight development surfaces: VS Code, Visual Studio, the Copilot CLI, the Copilot cloud agent, the Copilot app, JetBrains IDEs, Xcode, and Eclipse. All paid Copilot tiers get access: Pro, Pro+, Max, Business, Enterprise.

Grok 4.6's numbers are real. On CursorBench it posts 69.9% versus Grok 4.5's 66.7%. It leads the APEX-Agents leaderboard. The context window is 500K tokens. Pricing is $2/$6/M input/output for requests under 200K tokens. Above 200K tokens, the entire request reprices at $4/$12/M, not just the overage. That threshold is the one number to remember.

The distribution story is the story. Grok 4.6's Copilot landing means tens of millions of developers who use GitHub's interface can now select it from a picker that already carries Claude Sonnet 4.6 and Gemini. The picker is now genuinely multi-lab. GitHub is becoming an inference market, and xAI just bought a booth.

Claude Code shipped its forward_user_identity gateway feature on the same day. Grok in Copilot has no equivalent transparency layer. The two coding assistants are now competing in the same picker, on the same day, with different approaches to attribution and governance. One is asking who the user is. The other is not asking anything.

The builder's move: if you run long agentic coding sessions in Copilot, watch the token counter. The 200K repricing threshold is the one number to remember, and a session that drifts past it mid-run gets its whole cost retroactively repriced.

On August 13, OpenAI and Cerebras previewed Ultrafast: GPT-5.6 Sol running on Cerebras wafer-scale silicon at up to 750 output tokens per second, 14 times faster than Standard processing by their measurement.

This is not a new model. The intelligence is the same GPT-5.6 Sol that has been available on Standard infrastructure. The only variable is the hardware, and therefore the speed. Cerebras's architecture removes the memory bandwidth ceiling that limits token rate on GPU clusters. The result is a speed tier, not a capability jump.

OpenAI's benchmark choice is the read. On Humanity's Last Exam, a 2,500-question graduate-level benchmark, GPT-5.6 Sol Ultrafast completed the full set roughly 7 times faster than Claude Fable 5. Anthropic did not publish that comparison. OpenAI chose it. The benchmarking strategy is itself data about where the competitive pressure is being felt.

Speed-at-intelligence is a new product axis, and OpenAI is trying to name it before anyone else. "Ultrafast" is a branding decision as much as a technical one. The tier has no published price, no GA date, and no model ID string yet. It is in limited preview. The announcement is the point.

The contrast: Anthropic's speed strategy is model-level differentiation, Haiku for latency, Sonnet for balance, Opus and Fable for capability. OpenAI is selling the same frontier model at different speeds on different hardware. The Cerebras bet is a supply-chain argument: if specialized inference hardware scales, speed tiers become a product line. Both approaches are defensible.

The builder's move: waitlist only for now. If you are building latency-sensitive voice or real-time applications, this is the tier to watch for when pricing lands.