On July 24, Anthropic shipped Claude Opus 5. Short version: it cuts the flagship Fable 5's price in half and trails it by a single point.
The weirder part is next — it quietly deleted 80% of Claude Code's system prompt. Anthropic's reasoning: the model is good enough now that it does not need hand-holding.
Pricing: same as its predecessor, half the flagship
Opus 5 costs $5 per million input tokens and $25 per million output. For contrast: Fable 5 is $10/$50, GPT-5.6 Sol is $5/$30.
In human terms: finish a day's work, and Opus 5 costs half what Fable 5 does. On Artificial Analysis's weighted Intelligence Index cost, Opus 5 (max) runs about $2.03 per task, Fable 5 is $2.75, GPT-5.6 Sol is $1.04, Kimi K3 is $0.95.
Opus 5 scored 61 on the Intelligence Index, one point above Fable 5's 60. Half the price for a model that misses by one point.
Capability: it beats its own flagship at coding
On Frontier-Bench (autonomous terminal coding) Opus 5 hit 43.3%, versus 18.7% for Opus 4.8 and 33.7% for Fable 5 — more than double its predecessor and ahead of the flagship.
SWE-bench Pro (repository-level engineering): 79.2%. Caveat: most of these numbers are Anthropic's own; independent trackers are still catching up.
Anthropic openly admits where it loses: Mythos 5 stays stronger on cyber and autonomous biology; on exploit weaponization Opus 5 succeeded only 4 of 14 attempts versus Mythos's 13 of 14.
The part worth remembering: 80% of the prompt is gone
Claude Code engineer Thariq Shihipar said the new model removed 80% of Claude Code's system prompt. Internal coding evals showed no measurable drop after the trim.
The logic in one line: strong models need fewer nannies. The old "check your work" / "proceed step by step" scaffolding now backfires — the model over-verifies, slows down, and burns tokens.
Anthropic gave five principles: let the model judge rather than hardcode rules; invest in tool-interface design; disclose context progressively on demand; keep tool descriptions lean; rely on automatic memory instead of explicit instructions. They also suggest keeping CLAUDE.md under 60 lines (300 at most) and running /doctor to auto-trim.
Reality check: it got chattier by default
Opus 5 turns thinking on by default (a breaking change) and replies run longer than prior versions. Anthropic itself warns: if you do not want long replies, say so explicitly in the prompt, or the bill climbs.
Also, it now refuses less. Good for security teams (fewer false refusals), but a double-edged sword — looser guardrails mean your own usage policy has to keep up.
What this means for you
- Developers: clean up that pile of old prompts. The "please check carefully" / "think step by step" you wrote years ago may now slow things down. Run /doctor.
- Small businesses: for daily knowledge work, long documents, support drafts, Opus 5 is half the price of Fable 5 and good enough — a real cost lever.
- Still deciding: deep refactors and long-horizon reasoning - Opus 5; top-tier cyber or biology - Mythos 5; tight budget, high volume - GPT-5.6 Sol or Kimi K3.
One line: Anthropic is not selling "smarter" this time. It is selling "just as smart for half, and you no longer have to teach it how to work." For teams burning tokens daily, that is the real news.