Opus 5 is the model you pick when you're tired of picking models

For the last few months my workflow had a stupid shape to it: plan with one model, implement with another, then have a third one review what the first two did. Fable 5 for taste and orchestration, GPT-5.6 Sol for grinding a problem into the ground, and a lot of tab-switching in between.

Anthropic released Opus 5 on July 24. It's smaller than Fable 5, cheaper than Fable 5, and beats it on most published benchmarks. That combination is confusing enough that it's worth unpacking, because the interesting part isn't the benchmark table — it's what it does to model selection.

The pricing is real, the discount isn't

Opus 5 is $5 / $25 per million input/output tokens. Fable 5 is $10 / $50. So: half price, headline confirmed.

Except per-token price is the wrong unit. Different models spend different numbers of tokens reaching the same result, and Opus 5 is less token-efficient than Fable — roughly 37k tokens per task versus Fable's 33k in Artificial Analysis' measurements. Once you measure cost per finished task, the gap collapses from 50% to something closer to 20–25%.

There's a sharper number that explains the whole personality of this model. Across its top three effort levels, Opus 5 averages 103, 91, and 76 turns per task, against 55 for Opus 4.8 at max. Roughly double the agent loop. Everything people complain about downstream — verbosity, over-checking, handing control back to you — is that one number seen from different angles.

Two practical consequences

1. More tokens means slower responses and a context window that fills sooner, which means drifting off-task sooner on long agentic runs.
2. If you're optimising purely for dollars per solved problem at frontier capability, the OpenAI side still wins on raw efficiency.

Where the economics genuinely change is subscriptions. Fable is capped at half your weekly limit on the $100/$200 plans and was pulled from $20 Pro entirely. Opus 5 is the Max default and gives you 100% of your limit while burning it several times slower. If you live inside a Claude subscription rather than the API, that's the whole story.

Why a smaller model beats a bigger one

Fable and Mythos are the same underlying model — the difference is which requests are allowed in and which responses are allowed out, enforced by a classifier sitting in front. Nothing was removed from the weights.

Opus 5 is a distillation of that Mythos-class model, with the restrictions baked into training rather than bolted on. Think of it as filtering the bowl rather than putting a lid on it. Anthropic reports it as their most-aligned release: lowest rate of deceptive behaviour, hardest to trick into misuse, safest against irreversible actions.

That normally costs you real-world capability, and here it mostly didn't — but it did leave a fingerprint. CodeRabbit, who have benchmarked every Opus release since 4.x on their production code-review pipeline, found Opus 5 makes a stronger case as a builder than as a reviewer: it clarifies the goal first, offers several credible approaches, handles open-ended work without lunging at the first solution, and coordinated hundreds of background agents better than 4.8 did. One of their engineers called it less anxious than earlier Opus models. It also stayed slower than Fable 5 and noticeably over-cautious on security tasks — which is exactly what you'd predict from a model trained to be bad at exploitation.

Which is fine for our purposes: it's still good at finding vulnerabilities, which is what you actually want for debugging.

There's also a boring detail with outsized consequences: Opus 5 supports zero data retention. Fable and Mythos requests are logged and audited unconditionally, which quietly disqualified them for a large class of regulated workloads. That constraint is gone.

The benchmarks, and one better than benchmarks

Anthropic's own head-to-head against Fable 5: agentic terminal coding 43.3% vs 33.7%, computer use 70.6% vs 66.1%, business workflows 26.0% vs 17.4%. On CursorBench it lands about half a point behind Fable's peak, at roughly half the cost per task. ARC-AGI-3 came in near 30% — a benchmark that penalises taking more steps than a human would, and where nothing scored above 1% at launch.

The pattern across the rest is consistent: Opus 5 matches or edges Fable at getting work done with tools available, and loses to it on pure factual recall. Smaller model, less in the brain.

The more interesting result is Claire Vo's — she runs a seven-model eval and, importantly, ranks blind. She was scathing about the experience: neurotic, apologetic, and verbose enough that she named the output Claude slop. Then Opus 5 won her leaderboard anyway, taking top scores on front-end design and prototyping. Her summary was that it's her most loathed colleague and it does the best work.

Blind ranking is what makes that useful. Strip the label off and "I hate it" is purely a statement about ergonomics, not capability.

It is not a drop-in replacement

Here's the part that will actually cost you time if you skip it.

Dan Shipper's team at Every spent a week with it pre-release across coding, writing, and their internal agent, and called it a hard model to love. It argued with instructions, stopped before finishing, and behaved badly inside the skills and plugins they'd built for earlier models. His framing: a budget Fable, with the personality tics and not the underlying brilliance. Aaron Levie at Box reports the opposite from their internal evals — double-digit gains on technology and healthcare tasks — so this isn't a universal verdict.

Then Every deleted their existing skills and started over, and it got dramatically better.

That resolves the apparent contradiction with everything above. Prior Anthropic models had to be over-constrained: you'd repeat yourself across the system prompt, CLAUDE.md, and skills, because the model would infer what you really wanted instead of doing what you said. Opus 5 doesn't do that — but your scaffolding still does. Anthropic's own prompting guidance now tells you to remove instructions like "include a final verification step," because the model over-verifies on its own. Layer that on top of a model already running double the turns and you get compounding instructions, wasted loops, and wrong stopping behaviour.

So the migration is not s/opus-4-8/opus-5/:

• Strip verification, narration, scoping, and delegation instructions from skills — Opus 5 does all four proactively.
• Try a lower effort level than you think you need. Multiple reviewers found lower effort produced better results, not worse.
• Note the smaller operational changes: the effort ladder now runs low → medium → high → xhigh → max, fast mode is ~2.5× speed at Fable-equivalent token prices, and the minimum cacheable prompt dropped from 1024 to 512 tokens.
• Migrate one repeated workflow on a copy, with the same inputs, before touching production.

What I'd actually do

• Default to Opus 5 for anything you intend to merge — after rebuilding the prompts, not before.
• Keep Sol for throwaway code, scripts, and assistant-style tasks where completion speed beats diff quality.
• Keep Fable for the longest autonomous jobs, obscure debugging, and anything where you'll read the output line by line. Anthropic says the same thing: Opus for daily work, Fable for the occasional deep job.
• Verify it yourself, blind if you can. Take one real task, run it on two models, then have each review the other's plan without telling either which is which. That comparison taught me more in a day than any benchmark table.

The consensus emerging from launch week is oddly specific: this is the best everyday model available and the least pleasant one to sit next to. If you can live with a colleague who triple-checks everything and apologises constantly, it's a straight upgrade.

Sources: Anthropic's Opus 5 announcement and prompting guidance, Artificial Analysis, CodeRabbit's code-review benchmark, Claire Vo / How I AI, Dan Shipper / Every, and Theo's hands-on review. Pricing verified against published rates as of July 28, 2026.