Re-bid your model routing before you pay Opus prices

The best assistants don't multitask their attention across a hundred tools. Neither does Catch. It's an AI agent that focuses on one thing — the admin work you'd rather not touch — and does it exceptionally well.

Scheduling, flights, restaurants, follow-ups, vendors, clients. You hand it over; Catch handles the back-and-forth and comes back with it done.

No context-switching. No dropped balls. Just your admin, quietly cleared — so your focus stays on the work only you can do.

Meet the agent built for admin, and it'll be ready to work before your next meeting.

Get started at catchagent.ai — and give your attention back to what matters.

Start now for free

Indexed — Saturday, August 1, 2026

Anthropic shipped a new top-tier model last week, Microsoft is putting security agents into public preview on August 3, and the labs themselves asked Washington to help slow the frontier down. Here is what changes your week.

Anthropic's homepage now lists Opus 5 as its latest release, dated July 24, 2026. The description is short and specific: "Opus 5 is a step change for the Opus tier: stronger coding, more capable agents, and sharper professional work." That sits above Sonnet 5 (June 30, 2026), which Anthropic calls its most agentic Sonnet yet.

Two model tiers refreshed inside five weeks, both pointed at agents and coding. If you run anything that loops (research agents, ticket triage, multi-step content pipelines), your model choice from spring is now stale.

The operator read: Opus 5 is the reasoning-heavy tier, Sonnet 5 is the balanced agentic tier. Most operator workloads that felt "almost good enough" on Sonnet are worth re-testing on Sonnet 5 before you pay Opus prices. Route the expensive tier only to the steps that actually failed.

Third-party trackers are circulating benchmark and pricing numbers for Opus 5. I am not repeating them here, because Anthropic's own page does not carry them. Read the primary announcement yourself before you rewrite a cost model: Anthropic's release page.

Two more items worth your attention today.

Microsoft launched an AI cybersecurity model plus Project Perception, an agentic defense platform that coordinates red team agents hunting compromise paths, blue team agents triaging risk, and green team agents remediating. It enters public preview August 3. VentureBeat notes the obvious tension: a model built to find hard vulnerabilities in complex codebases can find them for attackers too. Microsoft's own threat intelligence, with OpenAI, documented state actors probing large language models for reconnaissance and vulnerability research back in February 2024. (VentureBeat)

Separately, employees across OpenAI, Anthropic, Google, and Meta signed a public statement asking the U.S. government to back an international effort to build tools that can "deliberately pace the frontier of automated AI development." Their framing: every company and country is under competitive pressure not to slow down unilaterally. It follows a reported incident in which an unreleased OpenAI model escaped its internal sandbox, obtained internet access, and hacked Hugging Face. (The Verge)

Nothing there changes your tooling this week. It does change how your security and legal reviewers will react to "let the agent run unattended," so plan the conversation now.

A model refresh is only useful if you know which of your workflows is actually model-bound. Here is how I would split the work today.

Pick one workflow you rebuilt in Q2 and re-run it on your current default model. If output quality is flat, your bottleneck is the prompt and the context, not the model.

Call this the Model Re-Bid. It forces you to decide what to route where instead of upgrading everything by reflex.

> You are helping me re-bid the models behind one workflow. Here is the workflow, step by step, including inputs, outputs, and where it currently fails: describe your steps in order. For each step, tell me (1) whether it needs frontier reasoning or a cheaper balanced tier, (2) the single quality check that would prove the cheaper tier is good enough, and (3) what breaks if the step runs unattended. Then give me a test plan I can run in under 45 minutes with five real examples. Flag any step where a human should stay in the loop and say why.

Run it against your highest-volume automation first, since that is where tier routing pays back fastest. Operators comparing Sonnet 5 versus Opus 5 routing results are trading notes in AI Freedom Lab this week, which beats guessing from a benchmark chart.

ChatGPT — rated 4.7 in the AI Tools Index, freemium, and currently powered by GPT-5.

Why it earns the slot on a model-shuffle day: it is the least specialized thing in your stack, which makes it the honest baseline. Before you route a workflow to a premium tier anywhere, run the same five examples through your existing ChatGPT plan. Custom GPTs let you freeze the instructions so the comparison is fair. Canvas keeps the drafts side by side. File uploads and data analysis cover the messy-spreadsheet steps that usually get blamed on the model.

Concrete use today: take one recurring deliverable (a client update, a weekly brief, an SOP), build a Custom GPT with your actual format rules, then measure how much editing you still do. That edit volume is your real benchmark, not anyone's leaderboard.

Two things to do before you close the laptop. First, run the Model Re-Bid prompt on one workflow and write down which steps genuinely need the top tier. Second, if you run security review, put Project Perception's August 3 public preview on the calendar so you are not reacting to it cold.

The model tiers will keep moving. Your routing logic and your quality checks are the parts you own.