Anthropic Launches Claude Opus 5 at Half Fable 5 Cost With Bold Benchmark Claims

Anthropic has released Claude Opus 5 at the same API pricing as its predecessor — $5 per million input tokens and $25 per million output tokens — while claiming performance that rivals or exceeds the company's flagship Fable 5 model, which costs twice as much. The new model replaces Opus 4.8 and is positioned as the everyday workhorse for coding, research, document analysis and business automation, leaving Fable 5 for the longest-running autonomous agent workloads ibtimes.sg.

The pricing parity with Opus 4.8 matters because Anthropic says Opus 5 outperforms its predecessor and, in some internal tests, even beats Fable 5. That claim has not been independently verified. The company's own announcement frames Opus 5 as delivering frontier-level performance for everyday professional workloads while balancing capability, latency and cost ibtimes.sg.

Early customer data offers the first external signal. Harvey, a legal artificial intelligence company, ran the same complex research task on Opus 5 and Opus 4.8 at its highest reasoning setting. Opus 5 produced comparable results while using 26 percent fewer tokens. For teams paying by usage, that efficiency gain compounds quickly across large workloads ibtimes.sg.

Zapier provided a second data point, reporting that Opus 5 completed an internal customer-retention workflow that previous models could not finish. Neither case study has been reproduced by third parties, and both come from companies with commercial relationships to Anthropic. The tension between promising early results and the absence of independent benchmark replication defines this launch.

The market context is straightforward. Fable 5 remains the premium tier at $10 per million input tokens and $50 per million output tokens. Opus 5 sits at half that price with a claim to near-frontier quality. If the benchmarks hold under independent scrutiny, the price-performance curve for production artificial intelligence workloads shifts meaningfully. If they do not, Opus 5 becomes another model that looks strong in vendor tests but behaves differently in the wild.

Anthropic's positioning also reflects a broader industry pattern: separating models by intended workload duration rather than raw capability alone. Opus 5 targets the high-volume, latency-sensitive tasks that dominate enterprise usage. Fable 5 targets the smaller slice of workloads that require extended autonomous reasoning. The distinction matters for capacity planning and cost modeling.

What happens next depends on independent evaluation. Researchers and engineers will test Opus 5 against standard benchmarks — SWE-bench, MMLU, HumanEval — and against real production workloads. The 26 percent token reduction Harvey observed is a specific, measurable claim that can be validated or falsified. Until that happens, the prudent stance is to treat Opus 5 as a promising candidate for migration testing, not a proven replacement.

Will independent benchmarks confirm that a half-price model can match the flagship on the tasks that matter most to your pipeline?