Claude Opus 4.8 Is Here: What Actually Changed, and When to Reach for It

Anthropic shipped Claude Opus 4.8 on May 28, 2026 — at the same price as Opus 4.7. After spending real time putting it through coding, design, and strategy work, here’s an honest breakdown of where the new model shines, where it still trips, and how to actually get the most out of it.

The one-line summary

Opus 4.8 is not a reinvention. It’s a sharpening. Anthropic’s own framing is telling: sharper judgment, more honesty about its own progress, and the ability to work independently for longer. Those three things — not a headline benchmark jump — are what you actually feel when you use it. And critically, it lands at the same price as its predecessor, which reframes the whole “is it worth it” question.

Where Opus 4.8 genuinely excels

Three patterns showed up over and over:

• Greenfield prototypes. Starting from a blank file, 4.8 is fast and confident. Give it a fuzzy idea for a tool and it’ll scaffold something usable in one shot more reliably than 4.7 did.
• One-shot features. Self-contained units of work — “build this component,” “add this endpoint,” “write this script” — come out clean and complete.
• Raw execution speed. It commits to a plan and moves. There’s less hand-wringing, less circling, more shipped output per unit of your attention.

If your work looks like “spin up something new, fast,” this is the model’s home turf.

Where it still struggles

The honest part. Three weaknesses persist:

• The last 10%. It gets you 90% of the way with startling speed, then stalls on the finishing details — the polish, the wiring, the “make it actually production-ready” pass. That final stretch still needs you.
• Edge cases in existing codebases. Greenfield is its strength; the inverse is its weakness. Dropped into a large, established repo with its own conventions and gotchas, it’s more likely to miss the subtle constraints that a careful engineer would catch.
• Hallucinations. Still present. It will occasionally invent an API, a function signature, or a fact with full confidence. Trust, but verify — especially anything that looks like a precise external reference.

4.8 vs 4.7 on strategy work: a surprise

Here’s the counterintuitive finding. On business strategy and roadmap work — especially data-heavy analysis — Opus 4.7 is still the model to reach for. Newer doesn’t mean strictly better across every axis. 4.8’s gains are concentrated in agentic coding and fast execution; for dense, analytical strategy reasoning, the older model held its ground or did better.

The takeaway isn’t “4.8 is worse.” It’s that “latest model” is not a synonym for “best model for this task.” Keep both in your toolkit and pick deliberately.

The features shipping alongside the model matter as much as the model

Two of these may change your workflow more than the raw intelligence bump:

• Dynamic workflows with parallel subagents. The model can now spin up and coordinate multiple subagents working in parallel — decomposing a task into independent threads and running them concurrently instead of serially. For multi-part work, this is a real throughput multiplier.
• Effort control in Claude.ai and Cowork. You can now dial how much thinking/effort the model spends. Low effort for quick lookups, high effort for hard problems. It’s a direct lever on the speed-vs-quality trade-off — and on cost.

How to actually get the most out of it

The model is only half the equation. The harness around it — how you prompt, structure context, and verify — is the other half. A few principles that pay off:

1. Play to greenfield. When you can, frame work as a fresh, self-contained build rather than a surgical edit deep inside legacy code. You’ll get cleaner output.
2. Own the last 10% yourself. Plan for it. Let the model sprint the first 90%, then budget your own time for the finishing pass. Don’t expect it to nail production polish unattended.
3. Verify anything that looks like a fact or an API. Hallucinations haven’t disappeared. Build a verification step into your loop for external references.
4. Use effort control intentionally. Don’t burn high effort on trivial tasks, and don’t starve a hard problem with low effort. Match the dial to the difficulty.
5. Lean on parallel subagents for decomposable work. If a task splits cleanly into independent pieces, let the model fan them out.
6. Keep 4.7 around for data-heavy strategy. Route deliberately. The right model depends on the task, not the version number.

The verdict

Opus 4.8 is a confident, fast, judgment-improved iteration that’s especially strong at greenfield building and one-shot features — and it costs the same as 4.7, which makes it an easy default for that kind of work. But it’s not a clean across-the-board upgrade: the last-10% problem and hallucinations remain, it’s weaker inside messy existing codebases, and 4.7 still wins on data-heavy strategy.

The smart move isn’t “upgrade and forget.” It’s treat 4.8 and 4.7 as different tools, lean into 4.8’s execution speed where it’s strong, and keep a human in the loop for the finish and the facts. Same price, sharper edge — just know which edge you’re using.

Based on early hands-on testing reported by Lenny Rachitsky and Anthropic’s official Opus 4.8 announcement (May 28, 2026). Opinions and framing are my own.