News

Anthropic launches Claude Opus 5 — new #1 on AA Intelligence Index at unchanged $5/$25

On July 24, 2026, Anthropic released Claude Opus 5 (claude-opus-5-20250724) — the new top of the Claude lineup. The headline: it matches Opus 4.8’s $5/$25 pricing while closing most of the gap to Fable 5 on benchmarks and surpassing it on several agentic measures.

Key numbers

BenchmarkOpus 5Fable 5Opus 4.8
AA Intelligence Index61 (#1)61.4 (prev #1)
SWE-bench Pro79.2%80.3%69.2%
Frontier-Bench43.3% (SOTA)
ARC-AGI-330.2% (4× next)
OSWorld 2.070.6%~65%
Zapier AutomationBench26% (~1.5× next)

What changed

  • Thinking on by default. Opus 5 ships with extended thinking enabled — no need to opt in. This is the first Claude model to do so.
  • Near-Fable at half the price. SWE-bench Pro 79.2% is only 1.1 points behind Fable 5’s 80.3%, at half the cost ($5/$25 vs $10/$50).
  • Agentic benchmarks pull ahead. OSWorld 2.0 (70.6%) and Zapier AutomationBench (26%) both surpass Fable 5 — Opus 5 is now the best computer-use and automation model in the Claude line.
  • Pricing unchanged. $5 input / $25 output per 1M tokens, with prompt caching and batch discounts matching the Opus 4.8 rate card.

The trade-off

Anthropic’s own system card notes a slightly higher hallucination rate compared to Opus 4.8 — the model gained massive agentic capability at the cost of the extreme calibration discipline that defined its predecessor. For high-stakes work where silent failures are expensive, this is the variable to monitor as production reports accumulate.

Why it matters

  1. The frontier price floor just dropped. Near-Fable capability at $5/$25 means most users no longer need to choose between quality and budget — the gap is now ~1 point on the hardest benchmark.
  2. Opus 4.8 enters legacy. Anthropic still serves it, but with Opus 5 at the same price, the default for new projects shifts immediately.
  3. The Claude lineup is now four-tier. Haiku → Sonnet 5 → Opus 5 → Fable 5, each roughly doubling in price with diminishing returns — a clean escalation ladder for agent fleets.

Full specs, benchmarks and pricing: Claude Opus 5 model page. For landscape context, see the 2026 LLM guide.

Sources

← All news