Claude Opus 5 went live this week, and the headline number is not a benchmark score, it is a price tag: half of what Anthropic charges for Fable 5, its own flagship model. We pulled the actual pricing and benchmark pages rather than trust the hype cycle, and the numbers changed how we are routing agent work across our own builds this quarter.
What actually shipped
Anthropic's own release announcement says Opus 5 comes close to Fable 5's frontier intelligence at half the price, and that framing holds up against the numbers on the Claude pricing page. Opus 5 runs $5 per million input tokens and $25 per million output tokens. Fable 5 runs $10 and $50. That is not a rounding difference, that is exactly half the cost per token, for a model Anthropic is now shipping as the default on Claude Max and the strongest model on Claude Pro.
The benchmark story backs up the price story. On CursorBench 3.2, Anthropic reports Opus 5 performs within 0.5% of Fable 5's peak score at max effort, at roughly half the cost per task. On Frontier-Bench v0.1, Opus 5 reportedly surpasses every other model Anthropic tested, more than doubling its own predecessor Opus 4.8's performance at a lower cost per task. Fable 5 still holds the lead on some of the hardest frontier reasoning work. Anthropic's own Fable product page positions it as the model built for the deepest, longest-running agent work, and its release notes point out Opus 5 trails a rival model on cybersecurity tasks. So this is not "one model wins everything." It is a much cheaper model that closes most of the gap on the tasks a 10-50 person company actually runs day to day: coding, drafting, research, agent chores.
| Model | Input / MTok | Output / MTok | Where it wins |
|---|---|---|---|
| Claude Opus 5 | $5 | $25 | Frontier-Bench v0.1, near-parity with Fable 5 on CursorBench 3.2 at half the cost |
| Claude Fable 5 | $10 | $50 | Still ahead on the hardest frontier reasoning tasks |
Why this matters more than another benchmark chart
Most operators do not need the single smartest model on earth. They need a model that clears the accuracy bar for a specific job, at a cost that does not eat the margin on the automation it powers. When a model that costs half as much closes to within half a percent of the flagship on real coding work, the economics of running agents at volume change. A drafting agent that triages fifty support tickets a day, or a research agent that pulls weekly market data, was never going to need the most expensive reasoning available. Now it clearly does not have to pay for it.
This is the same discipline we apply on every automation build: match the model to the job, not to whichever one is trending. A model swap that cuts your per-task cost in half, with accuracy that barely moves, is the kind of change that shows up directly in a client's margin. It is also exactly the kind of finding our revenue leak heatmap is built to surface: where a business is paying more than it needs to for the same output, whether that is staff hours or API spend.
Route by task, not by brand loyalty
The practical move this quarter is simple. Take every recurring AI task running in your business: a chatbot, a research agent, a coding assistant. Sort each one into two buckets: tasks where the cheaper model clears your bar, and tasks where it does not. Most of what a 10-50 person company runs falls into the first bucket. Reserve the expensive model for the handful of jobs where the extra reasoning changes a real outcome, like a contract review or a pricing model nobody wants to get wrong. Anthropic publishes a full model overview comparing every current model side by side, and that page is worth bookmarking, because this split is going to move again within a few months.
The bigger idea buried in the release notes
Anthropic's announcement has one line that matters more than the price cut: Opus 5 is built to be "much stronger at verifying its work and iterating carefully until it succeeds." That is not a claim about one model. It is a description of the pattern every reliable agent needs, and most businesses are not building it in yet.
Most AI pilots we see stop after one pass. The agent drafts the email, generates the report, writes the code, and the human either accepts it or fixes it by hand. There is no built-in check for "did this actually hit the bar," and no instruction telling the agent to keep working until it does. That is the difference between a demo and a system. A demo produces an output. A system produces an output, checks it against a defined condition, and keeps iterating until that condition is met. Anthropic describes Opus 5 doing exactly this on a task with no direct way to view a technical drawing: it built its own computer vision pipeline to extract the geometry rather than giving up or guessing.
What this looks like in a real build
Every system we build for a client gets a version of this loop. A lead-qualification agent does not just draft a reply, it checks the reply against the qualification criteria before sending. An invoice-processing agent does not just extract the numbers, it flags anything that does not reconcile before it touches the books. The verification step is what turns a clever demo into something you can trust running unattended. It is why our AI Operations Agent and Second Brain builds are scoped around a specific, checkable outcome from day one, not a general-purpose chat window.
Speed still matters inside that loop. Our own build data on lead response keeps landing on the same number: leads contacted within 5 minutes convert far better than leads contacted an hour later. A verification step that adds real delay defeats the purpose. The goal is an agent that checks its own work fast enough that the speed advantage never disappears.
What we are doing about it this quarter
Three concrete changes coming out of this release, and we would make the same three if you asked us to scope your stack this month.
- Re-price every routine agent task. Anything running on a flagship-tier model purely out of habit gets tested against the cheaper option first. If accuracy holds, the cost gets cut in half without anyone noticing a difference in output quality.
- Add an explicit success test to every agent. Not a vague "looks good," a specific, checkable condition: the invoice reconciles, the reply matches the qualification checklist, the report cites the right source documents.
- Keep the expensive model on standby for the hard 10%. The tasks where getting it wrong is expensive, contracts, pricing, anything client-facing, still get routed to the model with the most headroom, whatever it costs per token.
None of this requires a new hire or a six-month build. Our systems go live in days to weeks, not quarters, they run 24/7 once deployed, and every client owns 100% of the code and files we hand over. If you want a second set of eyes on where your current AI spend is going to the wrong model, that is exactly the gap our AI Concierge assessment is built to find. It is priced at 999 euros, credited straight back to the build if you move forward.
The takeaway for this quarter
Model prices and benchmark scores will move again before your next planning cycle. What will not move is the underlying pattern. Cheaper models are closing the gap on routine work faster than most companies are updating their routing, and almost nobody is building the verification step that turns a one-shot output into something you can trust unattended. Fix the routing, add the check, and the model you happen to be using becomes a much smaller decision. Our case studies and the rest of our thinking on this live on the blog if you want to see how the pattern plays out across different builds.