I built this hub because the job changed. A year ago Claude wrote copy. Now it runs the program. It audits, designs the test, builds it, ships it, and reads the result. This is the whole thing in one place.
I run experimentation programs for $50M to $300M DTC Shopify brands. 1,000+ tests shipped across 150+ brands. For years the work was the work. A strategist pulled GA4. A designer mocked the variant in Figma. A dev pasted it into the theme. An analyst read stat sig two weeks later. Four people, four tools, four handoffs, and a calendar full of "results pending."
Opus 5 collapsed most of that into one orchestration layer. Not because it writes better hooks, though it does. Because it can hold your entire analytics export, your theme code, and a quarter of test notes in a single prompt, then reach across your live stack and actually do the steps.
This page is the thesis for everything else in the hub. Read it once. Then go run a test.
Anthropic's frontier model, released July 24, 2026, built for deep reasoning, agentic (multi-tool, autonomous) work, and long-horizon enterprise tasks. Anthropic calls it a step-change over Opus 4.8 rather than an incremental bump, and prices it at half of Fable 5. The specs are the pitch.
Here is what matters for CRO, and nothing that doesn't.
| Spec | What it is | Why a CRO operator cares |
|---|---|---|
| Context window | 1M tokens, both the default and the maximum. There is no smaller context variant. Max output is 128k tokens | Load a full analytics export, a heatmap session log, an entire theme's Liquid/CSS, or a quarter of A/B-test notes in one prompt. No chunking. No "summarize this first." |
| Pricing | $5 / million input tokens, $25 / million output tokens (standard) | A full store audit costs cents in tokens instead of a strategist's afternoon. |
| Cost controls | Up to 90% savings with prompt caching, 50% with batch processing | Re-running the same context (your brand voice file, your test framework) gets cheap fast. |
| Fast mode | speed: "fast" runs up to ~2.5x output tokens/sec at double price ($10 in / $50 out per M) |
For when you want a variant brief back now instead of in 40 seconds. |
| Tool use + MCP | Function calling plus Model Context Protocol server connectivity, with the Opus 5 docs emphasizing "orchestrating complex multi-tool tasks" and "better tool triggering" | This is the part that wires Claude into Shopify Admin, GA4, Airtable, Figma, and Convert.com. The model doesn't just talk about your store. It reads and writes to it. |
| Thinking and effort | Thinking is on by default. Effort runs low, medium, high (the default), xhigh, and max | A quick rename runs at low effort. A full store audit or a sample-size calculation earns xhigh or max. Opus 5 turns extra effort into better output more reliably than any earlier Opus, so the dial actually matters now. |
A few capability notes worth keeping straight, because I will not print a number I can't trace.
The honest scope note. Anthropic's public consumer Opus page does not publish a standard MMLU/coding leaderboard table. The numbers above are the ones that come from a primary source. Where a spec lives only in launch coverage, I have said so.
What makes a full-stack CRO agent practical instead of theoretical is how Opus handles having a lot of tools available at once.
The Tool Search Tool (advanced tool use, introduced with the 4.5 generation and carried forward) lets the model discover tools on demand instead of loading every tool definition into the prompt upfront. On Opus 4.5, Anthropic reported MCP-eval accuracy climbing from 79.5% to 88.1% with it enabled, and roughly an 85% token reduction versus loading the full toolset.
Note: the 79.5 to 88.1% figures are from the Opus 4.5 announcement, not re-measured for Opus 5.