Claude Opus 5 Review: Brilliant, Frustrating, and Easy to Misconfigure

Direct Answer

Claude Opus 5 is a major technical upgrade and a poor candidate for a blind model swap. Anthropic says it approaches Claude Fable 5 on difficult coding and knowledge work at half the price, while keeping Opus 4.8's $5 input and $25 output price per million tokens. In a week of real use, however, Dan Shipper and the Every team found a model that could be argumentative, stop before finishing, and perform worse inside elaborate skills built for Opus 4.8.

Those two stories are not mutually exclusive. A stronger model can expose assumptions baked into an older harness. Opus 5 verifies, narrates, scopes, and delegates more proactively than Opus 4.8. When a skill already tells the model to do all four, the instructions can compound into wasted work or the wrong stopping behavior.

Field-test credit: Every and co-founder and CEO Dan Shipper. Every tested Opus 5 across production coding and knowledge-work workflows before publishing this early verdict.

Source Note

The video supplies Every's experience and preferences. These are useful field observations, not universal benchmark results. Anthropic's launch post, model documentation, effort guide, prompting guide, and migration guide are the sources for pricing, context limits, model behavior, safeguards, and migration details.

The second embedded video is the four-minute launch recap that matches the supplied transcript. It is useful for seeing Anthropic's benchmark story quickly, but the charts are vendor evaluations or partner evaluations. They do not prove that Opus 5 will improve a specific codebase, skill, or business workflow.

Link Map

What Anthropic Officially Launched

Claude Opus 5 is available as claude-opus-5 through the Claude API and on Claude for Pro, Max, Team, and Enterprise users. It has a one-million-token context window, up to 128,000 output tokens, and thinking enabled by default. It supports the full effort ladder: low, medium, high, xhigh, and max.

Standard API pricing is $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8 and half Fable 5's base token price. Fast mode runs at roughly 2.5 times the standard speed and costs $10 input and $50 output per million tokens. The model also lowers the minimum cacheable prompt from 1,024 to 512 tokens.

Anthropic reports a 0.5 percentage-point gap from Fable 5's peak on CursorBench 3.2 at max effort, at half the cost per task. It also reports state-of-the-art or best-at-cost results on Frontier-Bench, ARC-AGI 3, AutomationBench, OSWorld 2.0, GDPval-AA, and other evaluations. These are impressive launch results, but the relevant question for a team is still whether the new model passes its own acceptance tests.

Launch recap credit: Chase AI. The video walks through Anthropic's benchmark, visual-output, verification, and safety claims. Treat the demonstrations as orientation, then validate the model on your own work.

Every's Field Verdict

Dan's early routing is blunt. He uses GPT-5.6 Sol for roughly 80 percent of daily work, Fable 5 for the largest 20 percent, and has not yet found a stable place for Opus 5. He describes the new model as hard to love: capable of flashes of excellent judgment, but pushy, verbose, and sometimes unwilling to follow a workflow the way Opus 4.8 did.

Why Mature Workflows Can Break

Every's Compound Engineering plugin is not a small prompt. It combines planning, execution, review, debugging, browser testing, specialist agents, and saved learning. That structure helped earlier models finish work reliably. Opus 5 changes several defaults underneath it.

Anthropic says Opus 5 verifies its work without being asked, delegates more readily, narrates progress more often, and can expand task scope using its own judgment. Its official prompting guide recommends removing legacy instructions such as "use a subagent to verify" or "include a final verification step" because they can create over-verification without improving quality.

This explains a common migration failure: the team keeps every old guardrail, raises effort to max, and assumes more intelligence can only help. The result can be more tool calls, more agents, more narration, higher cost, and a model that loses the simple contract of what done means.

The Effort Paradox

Opus 5 defaults to high effort. Anthropic recommends xhigh as a starting point for genuinely difficult coding and agentic work, but it also says low and medium can retain strong quality with a fraction of the tokens and latency. Every's early tests found that stepping down sometimes improved behavior.

Lower effort is not secretly "more intelligent." It can be better matched to a bounded task. The model makes fewer tool calls, spends less time reopening settled decisions, and has less room to widen scope. Max remains useful for frontier problems, but it should be earned by an evaluation rather than selected as a status symbol.

Opus 4.8 to Opus 5 Migration

1. Freeze one real evaluation. Choose a repeated task with known inputs, expected outputs, tests, and a human reviewer.
2. Run the Opus 4.8 baseline. Record completion, defects, repair prompts, tool calls, tokens, latency, and review minutes.
3. Change only the model first. Use claude-opus-5 and keep the same task. This exposes the actual behavior change.
4. Remove duplicated verification. Delete instructions that force extra checks or a verifier subagent when tests already define acceptance.
5. Cap delegation. Permit subagents only for large, independent tracks. Set a deterministic maximum if the harness supports it.
6. Constrain scope. Tell the model to finish the requested task, make routine judgment calls, and raise only ambiguities that would materially change the result.
7. Sweep effort. Test low, medium, high, and xhigh. Reserve max for a task that demonstrably benefits.
8. Check the API changes. Thinking is on by default. Disabling it is allowed only at high effort or below. Revisit max_tokens because thinking and visible output share the output limit.
9. Promote by accepted result. Change the team default only when Opus 5 improves quality, cost, speed, or review burden across several runs.

Anthropic also provides a bundled migration command for Claude Code: /claude-api migrate this project to claude-opus-5. It can update model IDs and flag relevant parameter changes, but it cannot decide whether a long-lived skill still matches the new model's behavior. That requires an evaluation.

Copy-Ready Opus 5 Test Brief

Complete the requested task end to end.

Scope
- Deliver exactly what was requested.
- Make routine judgment calls yourself.
- Ask only when two reasonable interpretations would materially change the result.
- Do not add adjacent features or refactor unrelated code.

Delegation
- Use a subagent only for a large, genuinely independent workstream.
- Do not use subagents to repeat or verify work you can test directly.
- Use no more than [NUMBER] subagents.

Definition of done
- [Required output or behavior]
- [Tests or acceptance checks]
- [Files or surfaces that must be updated]
- [Actions that require approval]

Verification
- Run the listed acceptance checks once after implementation.
- Fix failures you caused.
- Report the result, remaining risks, and anything you could not verify.

Communication
- Keep progress updates brief.
- Lead the final response with what changed and whether the acceptance checks passed.

Run this brief at more than one effort level. Keep the inputs and acceptance criteria fixed. If medium passes and max adds cost without improving the accepted result, medium is the better configuration for that workflow.

Opus 5 vs Sol vs Fable

Routing does not need to be permanent. A useful first policy is Opus 5 medium for bounded professional work, Sol for fast interactive execution, Fable for the rare assignment where direction is the hardest part, and Opus 4.8 as the fallback for workflows that have not yet migrated cleanly.

Price Is Not Cost per Result

"Half the price of Fable" is true at the base token-rate level. Anthropic also reports half the cost per task on specific evaluations. Neither number predicts your production bill by itself.

Count the full accepted-task cost: uncached input, cache writes and reads, thinking and output tokens, tool calls, subagents, retries, elapsed time, human review, and repair work. A medium-effort run that ships cleanly can be cheaper than a low-effort run that needs three repairs. A Fable plan plus an Opus implementation can beat either model alone.

Safety behavior also affects routing. Anthropic expects Opus 5's cyber classifiers to intervene about 85 percent less often than Fable 5's, while keeping stronger restrictions on penetration testing, binary vulnerability scanning, and exploit generation. In Claude products, flagged requests can fall back to Opus 4.8; API users can configure server-side fallbacks. Log the model that actually completed the request before comparing results.

Video Chapters

Bottom Line

Opus 5 appears to be a genuine generation change. The official benchmarks, one-million-token context, lower-effort efficiency, visual gains, stronger code review, and Opus pricing make it one of the most important models to evaluate.

Every's frustration is the more useful launch lesson. A model can be stronger and still be worse inside a system designed around its predecessor. Do not compensate by writing an even longer prompt or turning effort to max. Simplify the behavioral scaffolding, keep the acceptance criteria strict, and measure the lowest effort level that reliably finishes your actual work.

Sources

• Every: We Tested Claude Opus 5. It's Frustrating with Flashes of Brilliance
• Every on X and Dan Shipper on X
• Chase AI: Opus 5 Is Here and It's Better Than Fable?!
• Anthropic: Introducing Claude Opus 5
• Anthropic: Claude Opus product page
• Anthropic: What's new in Claude Opus 5
• Anthropic: Prompting Claude Opus 5
• Anthropic: Effort parameter
• Anthropic: Claude migration guide
• EveryInc Compound Engineering plugin on GitHub