AI Models·July 24, 2026·4 min read

Anthropic launches Claude Opus 5: more performance for the same price — and it now knows how to check its own work

Anthropic launched Claude Opus 5 on July 24, 2026 — its new most-capable model and the direct successor to Claude Opus 4.8. The core message is simple: more performance for the same price. Opus 5 keeps Opus 4.8's rate — $5 input / $25 output per million tokens — but raises the bar on software engineering, knowledge work, and scientific research.

The most interesting change isn't a benchmark number but a behavioral trait: Anthropic says Opus 5 is much stronger at verifying its own work and iterating carefully until it succeeds. In practice, that means a model that stops, checks what it did, and corrects before handing you an answer — exactly what's most needed in coding agents and long tasks where an early mistake ruins everything downstream. It also produces better visual output.

The benchmarks

Anthropic positions Opus 5 as frontier across several key evals, with a constant emphasis on cost:

// Claude Opus 5 — per Anthropic

Frontier-Bench v0.1
Surpasses all other models.
CursorBench 3.2
Within 0.5% of Claude Fable 5's peak — at half the cost.
ARC-AGI 3
Three times the second-best model's score.
OSWorld 2.0 — computer use
Beats rivals at just over a third of the cost.

In science, the head-to-head with Opus 4.8 shows clear gains for the same price:

// Opus 5 improvement over Opus 4.8 (percentage points)

Organic chemistry+10.2 pp
Protein prediction+7.7 pp

Price and availability

Opus 5 costs the same as its predecessor: $5 input / $25 output per million tokens, with a 1-million-token context window. Its Fast mode runs faster at twice the base price ($10 / $50). It's available in Anthropic's API (ID claude-opus-5), on Claude.ai, and in Claude Code, which already uses it by default.

Why it matters

The 2026 trend holds: labs no longer just compete on being smarter, but on delivering more capability at the same price or cheaper. Opus 5 doesn't raise Opus 4.8's rate, yet it improves where it's felt most — code, computer-use agents, and science. And the focus on the model verifying its own work targets the real problem with autonomous agents: not that they're incapable, but that they're sometimes confidently wrong. A model that doubts, checks, and corrects is more useful than a flashier but careless one.

Source: Anthropic

What does this mean for you?

If you already use Claude Opus 4.8 for code or agents, switching to Opus 5 is straightforward: same price ($5/$25), same 1M context, better performance — especially on long tasks where you used to review every step, because the model now verifies and corrects its own work. In the API the ID is claude-opus-5. For simple, high-volume tasks a cheaper model is still better value; Opus 5 shines when reliability matters more than cost per token.