A few weeks ago the honest advice was “use Opus 4.8 for everyday work and only reach for the premium Fable 5 model when a task genuinely demands it.” Claude Opus 5 has quietly rewritten that advice. It brings performance close to the frontier Fable 5 model – beating it outright on several benchmarks – at half the price, and at the same cost as the Opus 4.8 it replaces. That’s an unusual thing in AI: a model that’s both better and cheaper than the more expensive option. Here’s what’s actually new, what it means in practice, and the few things worth watching.
What Is Claude Opus 5?
Claude Opus 5 is Anthropic’s new flagship general-access model, released in late July 2026, roughly two months after Opus 4.8. It sits at the top of the mainstream Claude lineup – above Haiku and Sonnet – and just below the restricted Mythos-class frontier models. Anthropic describes it as “a thoughtful and proactive model that comes close to the frontier intelligence of Claude Fable 5 at half the price,” and it’s now the default model on Claude Max and the strongest model available on Claude Pro.
You reach it through the API as claude-opus-5, in the Claude apps, and in Claude Code. The headline specs matter here, so they’re worth stating plainly:
- Price: $5 per million input tokens, $25 per million output tokens – identical to Opus 4.8, and half the cost of the Mythos-class Fable 5.
- Context window: 1 million tokens, both the default and the maximum. That’s five times the effective working memory of previous Opus models, enough to hold entire codebases or large document sets in a single session.
- Output: up to 128,000 tokens in a single response.
- Thinking on by default: Opus 5 decides how much to reason on each turn automatically, controlled by an effort setting from low up to max.
- Data retention: unlike Fable 5, Opus 5 carries no mandatory data-retention requirement – a genuinely important point for businesses handling sensitive information.
This is the natural successor to the model we profiled in our piece on why Opus 4.8 was the sensible default for most businesses. Almost everything that made 4.8 the pragmatic choice still holds – Opus 5 just moves the ceiling up substantially without moving the price.
The Big Change: Frontier Performance at Half the Frontier Price
The story of Opus 5 isn’t a modest incremental bump – on several measures it’s a step change over Opus 4.8, and it closes most of the gap to the far more expensive Fable 5. According to Anthropic’s own reported benchmarks:
- On Frontier-Bench, a software-engineering evaluation, Opus 5 more than doubled Opus 4.8’s score (from around 18.7% to 43.3%) and beat Fable 5’s 33.7% – at a lower cost per task.
- On ARC-AGI-3, which tests solving genuinely novel problems, Opus 5 scored roughly three times the next-best model – a jump from Opus 4.8’s 1.5% to about 30%.
- On CursorBench, a real-world coding benchmark, it landed within half a percent of Fable 5’s peak score at half the cost per task.
- On OSWorld, a computer-use benchmark, it beat Fable 5’s best result at just over a third of the cost.
- On life-sciences tasks, it improved on Opus 4.8 across the board, with notable gains in organic chemistry and protein analysis.
As always, these are Anthropic’s own figures and should be read with the usual healthy scepticism about vendor benchmarks. But the direction is clear and independent reviewers broadly agree: Opus 5 shifts the performance-and-cost curve in the customer’s favour. The one area where it deliberately doesn’t lead is cybersecurity, where it remains behind the restricted Mythos 5 – a safety choice, not a shortcoming.
Persistence: The Capability That Actually Matters
Benchmarks are one thing; behaviour is another. The most striking thing about Opus 5 in practice is its willingness to keep working until a task actually succeeds, rather than stopping at a plausible-looking answer. Three examples from Anthropic’s testing and early-access customers make the point better than any score:
- Asked to rebuild a machine part as a 3D model from a drawing it was deliberately given no way to see, Opus 5 wrote its own computer-vision pipeline to extract the geometry from the raw pixels, then reconstructed the part – repeatedly succeeding where no competing model could after five attempts.
- Given a real bug in a popular open-source package manager, Opus 5 found the underlying root cause and fixed an edge case the community’s own patch had missed. A competing model fixed only the surface symptom and declared the bug resolved.
- An engineer at a trading firm used Opus 5 to build a market-data feed for a new exchange in a single session. With no live feed to test against, the model built its own test harness to verify its code parsed the data correctly – a task previous models couldn’t complete even with detailed plans.
This is the difference between an AI that answers and an AI that finishes the job. For anyone using AI for real, multi-step work rather than one-off questions, that self-verifying persistence is worth more than a benchmark point or two.
Safer, Not Just Smarter
One of the quieter but more reassuring claims is on alignment. Anthropic reports that its pre-deployment behavioural audit found Opus 5 to be its most aligned model to date: it follows its guidelines more reliably than Opus 4.8, Sonnet 5, or Fable 5, shows the lowest rates of deceptive behaviour, is the hardest to trick into misuse, and is the least likely to take reckless actions with hard-to-reverse consequences. For businesses handing AI more autonomy over real tasks, “the smartest model” and “the safest model” being the same model is a genuinely welcome combination – and a contrast with the trust wobbles that surrounded the Fable 5 launch.
What to Watch: The Honest Caveats
No model is all upside, and a few things are worth going in with your eyes open about.
- It’s more verbose, so cost per task can rise even though the price didn’t. Thinking is on by default and Opus 5 produces longer, more thorough output than Opus 4.8. The per-token price is identical, but a given task can use more tokens – so real-world cost per task can go up if you don’t tune the effort setting down for simpler jobs. The flip side is that lower effort settings often match Opus 4.8’s best results while using fewer tokens, so the control is there if you use it.
- It’s not automatically the best at everything. Independent code-review specialists CodeRabbit found Opus 5 produced a more precise stream of review comments than their baseline, but caught fewer of the known bugs and generated roughly four times as many low-value nitpicks. Their conclusion was sensible: Opus 5 is excellent as a precision-focused reviewer within a mix of models, but not necessarily the one you’d want as your sole safety net on high-risk changes. The lesson generalises – test it on your specific task rather than assuming top benchmarks mean top results everywhere.
- Migrating from Opus 4.8 needs a small check. Because thinking is now on by default and counts toward the output limit, teams moving existing API workflows from 4.8 to 5 should revisit their token limits and effort settings rather than swapping the model name and assuming identical behaviour.
So Should You Switch?
For most people and businesses, yes – and for Claude app users it’s largely already happened, since Opus 5 is the new default on the top plans. The value proposition is rare: better performance than the more expensive frontier model on many tasks, at the same price you were already paying for Opus 4.8, with no data-retention catch and the best safety profile of any Anthropic model so far.
The practical advice from our Opus 4.8 guide still stands: start with one well-defined task, treat the output as a strong first draft to verify rather than a finished product, tune the effort setting to the job, and never build a critical workflow around a single provider without a fallback. Within those guardrails, Opus 5 is the most capable, most cost-effective, and most dependable everyday AI model a business can currently put to work – and the gap between “the sensible default” and “the frontier” has just narrowed to the point where, for most jobs, they’re the same thing.
If you want help working out where a model like Opus 5 genuinely fits into your business – which tasks to hand over first, how to control cost, and how to do it without creating new risks – get in touch.