Claude Opus 5 Sets New SOTA on Frontier-Bench — and Beats Fable 5 at Half the Price
Anthropic shipped Claude Opus 5 on Friday. The headline number is a 43.3% score on Frontier-Bench v0.1 — an agentic terminal coding benchmark that measures whether a model can complete real software engineering tasks end to end, including writing its own tooling when the task requires it. That is the highest score any model has posted on the benchmark. Fable 5 sits at 33.7%. Opus 4.8 was at 18.7%.
The model is priced identically to Opus 4.8: $5 per million input tokens, $25 per million output tokens. It becomes the default on Claude Max and the strongest model available on Claude Pro.
The benchmark that matters here
Frontier-Bench v0.1 is not a knowledge recall test. It drops a model into a terminal with a software task and grades whether the task is actually completed. Scores are inherently low — the benchmark was designed so that a 50% score would be extraordinary. The gap between Opus 5 (43.3%) and Fable 5 (33.7%) is therefore not a rounding error. Fable 5 was the previous frontier model and it is being outrun on this specific dimension by a cheaper model.
The Anthropic system card gives a concrete example of what 43.3% looks like in practice. On one task, Opus 5 was given a drawing of a machine part and asked to reconstruct it as a 3D FreeCAD model, but with no direct way to view the drawing. The model wrote its own computer vision pipeline to extract the geometry from raw pixels, then rebuilt the part. It succeeded repeatedly. No competing model could complete the task after five attempts.
SWE-bench and DeepSWE
SWE-bench Verified sits at 96.0%, above Fable 5’s 95.0% and well above Opus 4.8’s 88.6%. On SWE-bench Pro, Fable 5 still edges ahead — 80.0% versus 79.2% — which tracks with Anthropic’s framing that Fable 5 retains an advantage on some long-horizon coding tasks.
On DeepSWE v1.1, the ranking is: GPT-5.6 Sol 72.7%, Fable 5 69.7%, Opus 5 68.8%. Opus 5 is third on that benchmark, not first. That matters for anyone building agentic coding pipelines where the specific benchmark maps closely to their workload.
ARC-AGI-3
The result that will attract the most attention outside of software engineering is ARC-AGI-3: 30.2%. Opus 4.8 scored 1.5%. GPT-5.6 Sol scored 7.8%. The benchmark tests novel problem-solving — tasks the model cannot have memorised — and a 4x gap over the next-best model is not a benchmark rounding error.
Anthropic is cautious about what this means. The system card notes that Opus 5 remains behind Mythos 5 on offensive cybersecurity and long-running autonomous biological research. ARC-AGI-3 tests a specific class of fluid reasoning. Whether that translates directly to production tasks depends on the task.
GDPval-AA and knowledge work
On GDPval-AA v2, which scores models on knowledge-intensive professional tasks, Opus 5 posts an ELO of 1,861. Fable 5 sits at 1,747. GPT-5.6 Sol at 1,736. Opus 5 is first on this evaluation.
The Zapier AutomationBench result is notable for enterprise deployments: Opus 5’s pass rate is roughly 1.5x the next-best model at the same cost per task. Even at its lowest effort setting it passes more tasks than any other model on the leaderboard.
What this means for the index
Opus 5 enters the Stack Futures index as CL-OP5. The fundamentals score will update as Terminal-Bench, Live-Bench, and Arena ELO data accumulates over the coming days. The initial entry is anchored on SWE-bench Verified (96.0%) and the legacyAgentic score, with tape inherited from Fable 5 until Arena coverage is established.
At identical pricing to Opus 4.8, the cost-per-task improvement is material: the same $5/$25 rate delivers a model that more than doubles Opus 4.8 on Frontier-Bench. For agentic coding workloads specifically, the upgrade is straightforward.