- Published on
Claude Code Review Goes GA, Gemini Scam Guard & Haiku 5.5 β AI Updates Oct 21
Every week, AI tools cross a new boundary. This week, three updates stand out: how developers review code is changing, smartphones are now judging scam messages on their own, and the "small model = slow model" formula has been broken once again.
1. Claude Code Review Goes GA β PR Reviews Now Run Automatically
On October 15, 2026, Anthropic moved Claude Code Review to general availability (GA) for Team and Enterprise plans, graduating from research preview.
The setup is straightforward. An admin enables Code Review in Claude Code settings, installs the Claude GitHub App, and selects the target repositories. From that point on, reviews run automatically whenever a new pull request is opened.
Three Review Levels
Admins can choose the depth of review per repository:
| Level | Description |
|---|---|
| Quick | Fast scan, focused on obvious issues |
| Standard | Default. Balanced depth |
| Deep | Most thorough analysis, higher token usage |
3Γ Faster, No Merge Blocking
Reviews are roughly three times faster than in the research preview β most finish in six to ten minutes. Results are posted as inline comments on the PR, and the system never automatically approves or blocks a merge. Billing is based on token usage (billed separately from the plan), and Enterprise organizations can claim at least $10,000 in credits through January 15, 2027.
Tip: Set critical repositories to Standard and fast-iteration experiment branches to Quick to manage token costs. The feature is Team/Enterprise-only, but the /code-review command still runs locally for individual-plan users.
2. Gemini Lands in Android Messages β AI Judges Scam Texts in Real Time
On October 14, 2026, Google introduced Gemini-powered scam detection to Android Messages.
Enhanced Protection Mode
When users opt into Enhanced protection in settings, messages from unknown senders are sent to cloud Gemini models for analysis. The model evaluates the context and intent of each message to detect potential fraud.
For privacy concerns, Google emphasized Private AI Compute β an encrypted, isolated environment where even Google cannot access the data. The mode is off by default; users must enable it manually.
On-Device Gemini Nano β Blocking Links Before You Tap
On devices with Gemini Nano, unfamiliar links are checked on-device before the user taps them. This arrives first on Pixel 10 and later, Galaxy S26 series, and Galaxy Z Flip/Z Fold 8.
Additional security features include automatic hiding of login fields in Chrome during screen sharing, and blocking apps from silently forwarding calls (rolling out in early December).
Tip: Go to Android Messages settings β Spam & scam protection to turn on Enhanced protection yourself. Don''t wait for a notification β enable it now.
3. Claude Haiku 5.5 β The "Small = Slow" Formula Is Over
On October 7, 2026, Anthropic launched Claude Haiku 5.5, describing it as "our fastest model to date."
Benchmarks: 4Γ Leap Over Its Predecessor
The headline figure is on OSWorld 2.1, a computer-control benchmark:
| Model | OSWorld 2.1 Score |
|---|---|
| Claude Haiku 4.5 | 15.7% |
| Claude Haiku 5.5 | 72.4% |
| Claude Sonnet 5.5 | 83.9% |
That''s more than a 4Γ improvement. On real-world computer tasks β browser navigation, screen interaction β this small model now operates at a fundamentally different level than its predecessor.
Pricing and Specifications
- Input tokens: $0.10 per million (for prompts under 100K tokens)
- Output tokens: $0.50 per million
- Context window: 1,000,000 tokens β first in the Haiku line
- Effort setting: Added β first in the Haiku line
- Availability: Claude Platform API, AWS, Google Cloud, Azure
Anthropic''s target use cases: high-volume repetitive tasks (summarization, classification, DB queries), speed-sensitive work (live customer support), subagent execution, and browser automation.
Tip: If you run subagents or automation pipelines in Claude Code, try setting Haiku 5.5 as the default to cut costs. As the OSWorld scores show, "affordable but slow" no longer applies.
Closing Thoughts
This week''s three updates each serve a different audience. Development teams get an automated code quality gate, everyday users get stronger protection against scam messages, and developers and businesses can now do more with small models than ever before. The pace at which AI is moving from a specialty tool to everyday infrastructure keeps accelerating.
Sources:
- Code Review is now generally available β claude.com
- Anthropic Makes Code Review for Claude Code Generally Available β Unite.AI
- Android gets AI defenses against messaging scams β BleepingComputer
- Google adds Gemini-powered scam checks to Messages and Android calls β The Next Web
- Claude Haiku 5.5 debuts as faster small AI model β YourStory
- Anthropic launches Claude Haiku 5.5, its fastest model to date β NewsBytesApp