Kursol
AI Breaking News is an AI-generated alert, curated and reviewed by the Kursol team. When major AI developments happen, we break down what it means for your business.
Anthropic released Claude Opus 5 on July 24, and the headline is not the benchmark score — it's the bill. The model approaches Fable 5 across many categories at roughly half the price, whilst token pricing holds flat at $5 per million input and $25 per million output — unchanged from Opus 4.8. It also ships an effort dial: low, medium, or high per request, trading capability against cost task by task. Opus 5 becomes the default on Claude Max and the strongest model available to Pro subscribers. Anthropic reports a new internal best on coding and knowledge-work evaluations including Frontier-Bench and GDPval-AA, whilst sitting behind Mythos 5 on cybersecurity. It is the company's fourth model in under two months.
The Effort Dial Is the Part That Changes Your Budget
Most model releases ask you to pick a tier and live with it. The effort setting moves that decision from procurement to runtime. A classification job that never needed deep reasoning can run at low effort; a contract analysis that does can run at high, from the same model and the same integration.
That matters because the usual cost lever — routing cheap work to a smaller model — carries integration overhead. Every extra model in production means another prompt set to maintain, another quality baseline to track, another failure mode. An effort parameter on one model collapses much of that into a single argument. If you run a two-model setup purely to keep costs down, the second model may no longer be earning its maintenance burden.
Fallback Behaviour Deserves More Attention Than the Benchmark
Anthropic also shipped a beta feature called Automatic Fallbacks, which routes a request to a less powerful model when a prompt trips the safety classifier, instead of returning an error. For anyone running AI inside a customer-facing workflow, that changes the failure profile. A hard error needs handling code, retry logic, and usually a human. A quiet downgrade returns something.
Whether that is an improvement depends on your use case. In a support flow, a slightly weaker answer beats an error message. In a compliance review, a silently downgraded response is worse than a visible failure, because nothing tells you the quality bar moved. Decide which you are before turning the feature on, and log every fallback either way.
TechCrunch also reports that safety classifiers engage about 85% less often on Opus 5 than on Fable 5, and that Opus 5 is not subject to the 30-day data retention policy applied to Fable. Check both against your own data handling commitments before migrating.
What to Do This Week
1. Re-run your cost model with the effort dial included. Take your three highest-volume workloads and estimate what share could run at low effort. If that share is above half, the saving is large enough to justify a migration test now rather than at renewal.
2. Benchmark against your own tasks, not the published evaluations. Frontier-Bench and GDPval-AA are Anthropic's internal measures. Run 500 real requests through Opus 5 at each effort level and compare quality and latency against what you run today.
3. Decide your fallback policy before you enable fallbacks. Write down which workflows may silently downgrade and which must fail loudly. This is a five-minute decision that gets expensive if you make it after an incident.
The Bottom Line
Four frontier models in under two months means the pricing you negotiated in May is already stale. Opus 5 changes what frontier-class work costs more than what it can do, and the effort dial makes that cost a runtime decision rather than a contract one. The useful response is not to migrate immediately — it is to re-price your current workloads against this release and find out whether you are still on the right tier.
If you're not sure whether your current vendor mix matches your workload distribution, take our free AI readiness assessment to see where you stand.
AI Breaking News is Kursol's rapid analysis of major artificial intelligence developments—focused on what actually matters for your business. Subscribe to our RSS feed to stay informed.
FAQ
Per token, yes. In practice, no. The same task can consume fewer tokens at low effort than Opus 4.8 needed at its single setting, so your bill depends on the effort mix across your workloads. Re-run the cost model rather than assuming parity.
Only after testing. Opus 5 approaches Fable 5 in many categories, not all, and sits behind Mythos 5 on cybersecurity. If your workloads cluster where the gap is real, the saving isn't worth it. If they don't, half the price is a strong argument.