The AI Catchup - October 22, 2026

Anthropic launched an agent that comes to you, and Claude Code Review left research preview. The items that matter:

Claude Duet: a proactive agent app on Claude Opus 5.5 that sends a morning brief, checks your connected apps every few hours, and asks before it sends, pays, signs, or deletes
Claude Code Review is GA: automatic pull request reviews for Claude Team and Enterprise, with most reviews done in 6-10 minutes instead of about 20
Cursor canvases and Environments API: Cloud Agents hand back interactive canvases you share by link, and large environments load 5x faster
OpenAI 12-hour prompt caching: keep a stable prompt prefix cached for 12 hours instead of the 30-minute default, on GPT-5.6 and later
OpenAI's legacy-language results: GPT-6.1 Sol scores 86.6% on OpenAI's COBOL migration test at low reasoning effort, up from 5.4% for GPT-5.6 Sol
The catch: Duet is a US-only waitlist for Pro and Max and spends your normal usage limits; a 12-hour cache write costs 2.5 times the normal input rate, and Code Review is off-limits to organizations with Zero Data Retention or HIPAA configuration

Every claim below is sourced in the linked write-ups.

Claude Duet: An Agent That Comes to You

Before: Claude answered when you opened a chat. Now: Claude Duet, a research preview since October 21, briefs you each morning, messages you when something needs attention, and keeps working on its own cloud computer after you close the app.

Model: Claude Opus 5.5 only; Duet decides how much effort a job needs
What it does: finds missed Gmail and Slack messages, checks QuickBooks or Stripe for unpaid invoices, and hands back larger jobs as files such as a research brief, a financial model, or a slide deck
Approval gates: it waits for you before it sends a message, makes a payment, signs a document, or deletes anything
Access: US Pro and Max subscribers by waitlist, on iOS, Android, and claude.ai; Team and Enterprise are not included yet
Watch out: no extra charge, but it draws on the same five-hour and weekly limits as chat, and its background checks count too; it cannot see files on your computer, so it is no substitute for Claude Code

What Duet does, who gets it, and how it compares to ChatGPT dots.

Claude Code Review Is Generally Available

Before: Code Review was a research preview, and a review took about 20 minutes. Now: it is generally available to Claude Team and Enterprise organizations, and Anthropic says most reviews finish in 6-10 minutes.

Setup: an organization Owner enables it, installs Anthropic's GitHub App, and picks repositories; developers configure nothing
When it runs: when a pull request opens, after each push, or on request; anyone with write access can comment @claude review
Depth: Quick, Standard, or Deep per repository; Standard is the default
Watch out: it never approves or blocks a pull request, and it does not support customer-managed encryption keys yet

Setup, review levels, and the limits on who can use it.

Cursor Cloud Agents Build Canvases

Before: a Cloud Agent's result came back as a text response. Now: Cloud Agents can build interactive canvases, and your team opens a finished one from a link.

Environments API: in beta; create and manage the cloud environments your agents run in from code
Speed: large environments load 5x faster, per Cursor's changelog
Watch out: check what a canvas contains before you share the link, as you would any team dashboard

What canvases and the Environments API add.

OpenAI Prompt Caching Lasts 12 Hours

Before: a cached prompt prefix had a 30-minute minimum lifetime by default. Now: set prompt_cache_options.ttl to 12h in the Responses API and it stays eligible for cached-input pricing for at least 12 hours.

Who gets it: Responses API requests on GPT-5.6 and later; older models use the separate prompt_cache_retention setting
Cost: a 12-hour cache write costs 2.5 times the uncached input rate; reads keep the model's cache-read rate, 0.1 times input for most models and 0.05 times for GPT-6.1 Sol
Minimum: the prefix needs at least 1,024 input tokens
Watch out: a longer TTL does not guarantee a hit; caches are not shared across organizations or regions

When the longer lifetime pays off, and how to split a prompt across breakpoints.

OpenAI's Models on COBOL, Fortran, and Visual Basic

Before: GPT-5.6 Sol passed 5.4% of OpenAI's COBOL migration test at low reasoning effort. Now: GPT-6.1 Sol scores 86.6% at that setting and 97.9% at higher effort, by OpenAI's own evaluation.

MultiLangBench: GPT-6 Astra averaged 87.4% building a SQLite-compatible database engine across 20 languages, against 26.8% for GPT-5.6 Sol
Cheaper option: GPT-6 Luna averaged 40.3% at an estimated $0.77 per sample
Watch out: these are OpenAI-designed tests, not a new model and not a promise for your codebase

What each test measures, and how to run your own.

Quick Hits

ChatGPT Slides: rolling out to eligible Pro, Business, and Enterprise users with Space access; build a deck from Work or Space, edit it with teammates, and export to PowerPoint, PDF, or Google Slides on the web
GPT-Live-1.1 in ChatGPT Voice: GPT-Live-1.1 and the smaller GPT-Live-1.1 Quip are rolling out, with more natural accents and new voices in six languages plus Bubbie in English
Claude Platform tracing: OpenTelemetry traces for the Messages API, Managed Agents, and Claude SDKs in public beta, sent to any OTLP endpoint; metadata-only by default
OpenAI PHP Version Checker: two Codex skills, phpcheck-url and phpcheck-server, flag PHP versions that may need security updates; a version check, not proof a site is patched
New on AI Catchup: Cursor, Claude Code, and Codex trackers list current plans, prices, limits, and models, each row quoting the vendor's own page; sign up on a tracker for an email when that vendor changes something

Ship It This Week

Turn on Claude Code Review for one busy repository at Standard and compare its findings with your human reviews.
Set prompt_cache_options.ttl to 12h on one long, stable system prompt, then check cached reads against writes in your usage data.
Run a small COBOL or Fortran module through GPT-6.1 Sol against your own acceptance tests before trusting the benchmark.
Ask a Cursor Cloud Agent for a canvas instead of a text summary on your next status report.

Account and Billing Notes

Claude Code Review credits: eligible paid Claude Enterprise organizations can claim $10,000 to $50,000 in Code Review credits, once, by January 15, 2027; the 30-day clock starts at the claim, so set up repositories first. Details.
Cursor Router correction (Teams and Enterprise): we corrected our page on October 22; Cost mode bills as a Cursor-model request from the weekly pool, while Balance and Intelligence bill at the list price of the model a request is routed to. Details.
Gemini 4 Argon in Cursor: listed as a third-party model at $2 input and $10 output per million tokens; on individual X Subscription plans it bills from prepaid credit on Premium and above, and Plus and Free do not include it. Details.

Claude Code Review is the release to try this week: automatic pull request reviews that now finish in minutes. If a teammate still waits on a reviewer for every small pull request, forward this.

Until next week, stay caught up.