anthropic shipped claude opus 5 — top of frontier-bench at half fable 5's price

Anthropic released Claude Opus 5 on July 24, 2026 — its fourth model in under two months. The pitch: near-Fable-5 intelligence at half the price. Pricing is $5 per million input tokens and $25 per million output tokens, identical to Opus 4.8 and half of Fable 5's $10/$50 rate card.

the headline coding number is on Frontier-Bench v0.1, a specialized agentic evaluation that measures whether models can build working software from engineering specs in a real terminal environment. Opus 5 posts 43.3% — roughly double Opus 4.8's 21.1% on the same benchmark. on ARC-AGI-3, which measures abstract reasoning, it posts 30.2%. Anthropic also reports DeepSWE v1.1 at 68.8% and FrontierCode Main at 53.4%.

fourth model in two months

the pace is worth noting. Anthropic shipped Sonnet 5 in May, then Haiku 4.5, then Opus 4.8, and now Opus 5. each release has pushed coding capability meaningfully forward. Opus 5 is the first in the family to challenge Fable 5 — Anthropic's own frontier model — on benchmark scores while coming in at substantially lower cost per token.

for vibe coding specifically: Frontier-Bench is the most relevant benchmark here because it doesn't just measure whether the model can answer a question about code — it measures whether the model can complete an agentic loop: read specs, write code, run tests, fix failures, and ship a working artifact. a 30.2% score means Opus 5 completes roughly one in three of those full engineering loops successfully on a hard test set. that's a meaningful bar for anyone using Claude Code or building multi-step agent workflows.

when to use opus 5 vs fable 5

if you're running agentic coding tasks at scale — CI pipelines, overnight refactors, PR review agents — Opus 5's cost profile makes it a better default than Fable 5 for the same benchmark results. if you need Fable 5's specific edge cases (the benchmark difference is small; the real-world difference for your codebase may be larger or smaller), you pay the premium. for interactive vibe coding in Claude Code, the practical answer is: try Opus 5 first.

Anthropic is also shipping a Fast Mode for Opus 5, the same mechanism it uses on Fable 5 and Opus 4.8 — trades some reasoning depth for lower latency. for quick exploratory prompts and conversational use, Fast Mode is the default you'll likely want. for agentic tasks where you're paying per token anyway, full mode is probably worth it.

the pattern Anthropic is establishing

each release in this cycle has held the same pricing while raising the capability ceiling. the effect is that the cost-per-task for complex agentic coding keeps dropping even if the per-token rate stays flat — better models complete tasks in fewer turns. if that pattern holds into the next release cycle, the economics of running AI coding agents continuously gets materially more attractive.