AI Dev Weekly #22: Claude Code Auto Mode Goes Live, Stripe Buys OpenRouter for $7.5B, GLM-5.3 Ships
AI Dev Weekly is a Thursday series where I cover the week’s most important AI developer news, with my take as someone who actually uses these tools daily.
Four stories this week. Claude Code quietly stopped asking permission before most tool calls. Stripe closed one of the largest AI infrastructure acquisitions of the year. Zhipu shipped a model that matches Kimi K3 on price-adjusted performance. And SpaceX’s Cursor acquisition became official the same week Grok 4.6 landed with a new top-tier reasoning mode.
1. Claude Code auto mode is now the default — here’s what actually changed
Starting August 14, new Claude Code sessions on Pro, Max, and Team plans open in auto mode by default, per Anthropic’s own announcement. Instead of approving every file write or shell command yourself, each tool call now routes through a separate classifier model that decides whether the action is safe to run without asking you.
What’s confirmed:
- Applies to Pro, Max, and Team plans. Enterprise, the Claude API, and cloud-partner deployments (Bedrock, Vertex, Foundry) stay opt-in for now.
- If you already set a different permission default yourself, that choice is preserved — you’ll get a one-time prompt asking if you want to switch, but it won’t be overridden silently.
- In Anthropic’s own controlled study, human reviewers manually approving each action missed 143 of 1,053 planted dangerous commands. The classifier caught 937 of them, per reporting on Anthropic’s study (via implicator.ai).
My take: The classifier catching more dangerous commands than human reviewers is a genuinely useful data point, but it’s Anthropic’s own study measuring Anthropic’s own classifier — worth independent validation before treating “937 out of 1,053” as the last word on how safe auto mode is in practice. The bigger practical question is whether it changes your workflow at all: if you already picked --auto-edit or a custom permission mode, nothing changes for you. If you’ve been running default Claude Code on Pro/Max/Team, this is a genuine behavior shift you should know is happening, not something to discover mid-session. Check your permission mode with /permissions if you’re not sure which one you’re on.
2. Stripe closes its OpenRouter acquisition at $7.5B+
Stripe confirmed on August 19 that it has closed its acquisition of OpenRouter, the AI model routing layer used by developers to access 400+ models through one API. Reported deal size varies by outlet — Bloomberg and CNBC put it above $7 billion, Forbes cites over $8 billion, and the Wall Street Journal had earlier reported talks around $10 billion. What’s consistent across all reporting: this is roughly 6x OpenRouter’s own $1.3 billion valuation from a funding round less than three months earlier, in May 2026.
Why Stripe: OpenRouter sits at the layer between applications and every major AI model — Stripe’s stated interest is in tokens becoming a spend category it can process and route the same way it processes card payments. Multiple outlets frame this as Stripe extending its infrastructure bet from money movement to AI-token movement.
What it doesn’t change today: OpenRouter’s API, pricing, and model catalog continue operating as-is. No changes to the developer-facing product have been announced alongside the acquisition close.
My take: If you route model calls through OpenRouter (and a lot of the setups we cover here do), the acquisition itself isn’t an action item yet — but it’s worth watching for two things over the next few months: whether OpenRouter pricing or model access changes under Stripe ownership, and whether Stripe bundles OpenRouter access into its existing payments/billing products in a way that changes how you’d integrate it. Nothing to do today beyond noting who now owns a piece of infrastructure a lot of us depend on. See our OpenRouter complete guide if you haven’t set it up yet.
3. GLM-5.3: same base model, matches Kimi K3 through post-training alone
Zhipu AI (Z.ai) released GLM-5.3 on August 14, and the interesting part isn’t the benchmark number, it’s how they got there. Per multiple outlets covering the release (felloai.com, qubrid.com), GLM-5.3 runs on the same 744-billion-parameter base as GLM-5.2 — no new pretrain, no architecture change. Every reported gain comes from extended post-training.
The numbers:
- Artificial Analysis Intelligence Index: 60, matching Kimi K3 and one point behind GPT-5.6 Sol
- Cost per task: $0.68 (per Artificial Analysis, via the-decoder.com), 1.5x GLM-5.2’s $0.44 but 19% cheaper than Kimi K3’s $0.84
- API pricing: $1.40 input / $4.40 output per million tokens — unchanged from GLM-5.2
- Access: live via API and the GLM Coding Plan (starting $18/month); public weights delayed to roughly August 28
My take: A model getting meaningfully better through post-training alone on an unchanged base is a genuinely useful data point about where the easy gains are right now in the open-model race — it suggests there’s still headroom in post-training that doesn’t require a new pretrain cycle, which is cheaper and faster to iterate on. The catch is the same one we’ve flagged before with fast-moving Chinese model releases: API access and a coding plan are live today, but the open weights that let you self-host or independently verify the claims aren’t out until the end of the month. If self-hosting is the point for you, this is a “wait two weeks” release, not a “switch today” release.
4. Cursor’s SpaceX deal closes officially, same week as Grok 4.6
SpaceX officially closed its $60 billion acquisition of Cursor this week, an acquisition that had been reported and discussed for months but became final now — and almost simultaneously, xAI (now SpaceXAI) released Grok 4.6, explicitly positioned around long-running agentic and coding work.
We’ve already covered Grok 4.6 in detail: same $2/$6 pricing and 500K context as Grok 4.5, plus a new xhigh reasoning tier and broad benchmark gains on xAI’s own reporting (see the full Grok 4.6 guide for what’s verified vs. not). The timing with the Cursor deal closing isn’t a coincidence — Grok 4.5 was already the first model co-trained directly with Cursor’s engineering team, and 4.6 continues that relationship now that the acquisition is formally done rather than pending.
My take: Nothing changes for existing Grok/Cursor users this week specifically — pricing and access are the same as they were under Grok 4.5. What’s worth watching is whether SpaceXAI now leans harder into Cursor-exclusive features now that the deal is no longer “pending,” the way Meta’s Muse Code leans on data-sharing pricing tiers. If you’re a Cursor user who picked it partly for the Grok integration, that bet looks more durable now that the acquisition is actually closed rather than reported-but-not-final.
That’s it for this week. Next Thursday: whether GLM-5.3’s open weights actually ship on schedule, early reports on how auto mode performs at scale in Claude Code, and whatever else breaks.
Want this in your inbox? Subscribe to AI Dev Weekly.
Previous issue: AI Dev Weekly #21
FAQ
Does Claude Code auto mode affect me if I’m on the free tier or API?
No. The August 14 default change applies to Pro, Max, and Team plan sessions only. Claude API access, Enterprise plans, and cloud-partner deployments (Bedrock, Vertex, Foundry) remain opt-in for auto mode.
Is GLM-5.3 open source?
The API and GLM Coding Plan are live now. Public weights are expected around August 28, 2026 — not available at the time of this issue.
What happens to OpenRouter now that Stripe owns it?
No announced changes to OpenRouter’s API, pricing, or model catalog as of the acquisition close on August 19. Worth monitoring, not something requiring action today.
Is Grok 4.6 different now that the Cursor deal is closed?
Not in pricing or specs — those are unchanged from when Grok 4.6 launched. The deal closing formalizes a relationship that was already shaping Grok’s development (Cursor co-training started with Grok 4.5).