Anthropic dropped two bombs in 48 hours

Today on FindMeAI:

Two drops in 48 hours. Both change how you work with Claude.

→ Claude Opus 4.7 is live — beats GPT-5.4 and Gemini 3.1 Pro. Self-verifies its own output. Better vision. Same price.
→ Routines launched — cloud automations that run on Anthropic's infrastructure. Your laptop stays closed. Your code gets reviewed at 2 AM.
→ Mythos still looms — 4.7 is powerful, but Anthropic admits it's "less broadly capable" than the unreleased Mythos
→ What to set up this weekend

Claude Opus 4.7 is a notable improvement on Opus 4.6 in advanced software engineering, with particular gains on the most difficult tasks. Anthropic

But the headline feature isn't raw benchmark scores. It's this: Opus 4.7 handles complex, long-running tasks with rigor and consistency, pays precise attention to instructions, and devises ways to verify its own outputs before reporting back. Anthropic

Read that again. It self-verifies. Before handing you the result, it checks whether the result is actually right.

What's new:

→ Better vision — sees images in greater resolution Anthropic → More tasteful and creative on professional tasks — higher-quality interfaces, slides, and docs Anthropic → New "xhigh" effort level between high and max for finer control over reasoning vs. speed Axios → Better file system memory — remembers notes across long, multi-session work Anthropic → New /ultrareview command that runs a dedicated review session flagging what a careful reviewer would catch 9to5Mac

The benchmarks: Opus 4.7 beats Opus 4.6, ChatGPT 5.4, and Gemini 3.1 Pro across multiple categories. Axios But — Anthropic publicly conceded it still trails Mythos Preview Axios, the model they won't release publicly.

Pricing: Same as Opus 4.6 — $5 per million input tokens, $25 per million output. Anthropic No price increase for better performance.

One thing to watch: Opus 4.7 uses an updated tokenizer. The same input can map to roughly 1.0–1.35× more tokens. 9to5Mac Your costs might tick up slightly even at the same per-token price.

This dropped two days before Opus 4.7 and might matter more for your daily workflow.

A routine is a Claude Code automation you configure once — including a prompt, repo, and connectors — and then run on a schedule, from an API call, or in response to an event. Claude

The critical upgrade: Routines run on Claude Code's web infrastructure, so nothing depends on your laptop being open. Claude

Three trigger types:

→ Scheduled — hourly, nightly, weekly. "Every night at 2AM: pull the top bug from Linear, attempt a fix, and open a draft PR." Claude

→ API — every routine gets its own endpoint and auth token. Wire Claude into your alerting, deploy hooks, or internal tools. Claude

→ GitHub — subscribe a routine to kick off in response to GitHub events. Claude creates a new session for every PR matching your filters. Claude

Real use cases already emerging:

→ Backlog management — triage new issues nightly, label, assign, post summary to Slack → Deploy verification — smoke checks after each production deploy, regression scan, go/no-go to the release channel → Alert triage — Datadog fires, Claude pulls the trace, correlates with recent commits, drafts a fix → Docs drift — weekly scan of merged PRs, flags stale documentation, opens update PRs

Think of routines as dynamic cron jobs — they prompt an AI model on a schedule and can take different actions depending on the context they encounter. SiliconANGLE

Limits: Pro users can run 5 routines per day, Max gets 15, Team/Enterprise gets 25.Anthropic Passed OpenAI on Revenue. Both Are Burning Billions.

Here's the part Anthropic is being unusually transparent about.

Opus 4.7 is an improvement over past models but is "less broadly capable" than Claude Mythos Preview. CNBC

Mythos is the model that Anthropic is privately warning government officials makes large-scale cyberattacks much more likely. Axios It's the model they accidentally leaked last month. It's the model they will NOT release publicly.

So what's Anthropic's plan? They're using Opus 4.7 as the first model to test new cybersecurity safeguards — safeguards that automatically detect and block high-risk cyber requests. Anthropic

During training, Anthropic experimented with efforts to "differentially reduce" Opus 4.7's cyber capabilities. Anthropic Translation: they deliberately made it less dangerous than Mythos while keeping everything else strong.

The dual-track strategy is clear: Mythos stays behind closed doors. Opus 4.7 goes to the public with cyber guardrails baked in. What they learn from 4.7's deployment informs whether — and how — Mythos-class models ever get a general release.

Why this matters for builders: The most powerful AI model in existence is being deliberately withheld. What you're getting in Opus 4.7 is the strongest publicly available model — but not the strongest model that exists. That gap is new. And it's widening.3 Signals From This Week That Shape What Comes Next

The Full Claude Code Stack — Where Everything Fits Now

Claude Code just went from "coding assistant" to "operations platform." Here's the current stack, top to bottom:

Layer 1 — Model: Opus 4.7 Self-verifying output. Better vision. xhigh effort level. /ultrareview for code review. The brain.

Layer 2 — Routines (NEW) Cloud automations on Anthropic's infrastructure. Scheduled, API-triggered, or GitHub-event-driven. Runs without your laptop. The autopilot.

Layer 3 — Managed Agents: Build and deploy custom agents via Console, CLI, or Claude Code. Define once, run at scale. The workforce.

Layer 4 — Channels Message Claude Code from Telegram or Discord. Async communication. Work gets done while you're on the train. The remote control.

Layer 5 — Skills + Subagents: Teach Claude repeatable workflows. Spawn specialized workers. Persistent memory across sessions. The institutional knowledge.

Layer 6 — Desktop App (Redesigned) Multiple sessions side by side, integrated terminal, file editor, HTML and PDF preview, drag-and-drop layout. 9to5Mac The command center.

Every layer connects. A Routine triggers at 2 AM → uses an Opus 4.7-powered agent → applies your Skills → posts results to your Telegram via Channels → you review in the redesigned Desktop app at 7 AM.

That's not a coding tool. That's infrastructure.

Two things. 30 minutes total. Do them before tomorrow.

☐ Create your first Routine. Open Claude Code. Type /schedule. Pick your most repetitive daily task — deploy checks, issue triage, PR review. Set it to run nightly. Close your laptop. Wake up to results. Pro plan: 5/day. Max: 15/day. The Register Start with one. 15 min.

☐ Test Opus 4.7 on your hardest problem. Open Claude. Pick the task you normally supervise closely — a complex refactor, a multi-file feature, a gnarly bug. Hand it to Opus 4.7 with xhigh effort. Watch it self-verify before reporting back. Compare to what 4.6 gave you last week. 15 min.

The gap between "uses Claude" and "runs Claude infrastructure" is the new career moat. The builders who set up Routines this weekend will have automated workflows running by Monday morning. Everyone else will still be typing the same prompts by hand.

Close your laptop. Let Claude work. →