Claude Opus 5 Launches July 24: Near-Fable Intelligence at Half the Price
Claude Opus 5 Launches July 24: Near-Fable Intelligence at Half the Price
July 24, 2026 closed the loop on Anthropic's busiest quarter yet. Four models in two months — Fable 5, Sonnet 5, export-control drama, and now Claude Opus 5, which Anthropic says delivers near-Fable frontier intelligence at half the token cost (official announcement, VentureBeat).
By July 28, Opus 5 sits at #1 on Artificial Analysis's Intelligence Index (61) and Agentic Index (55.3) per industry roundups — the first new top-ranked frontier model since the July 9 GPT-5.6 and Grok 4.5 launches.
Disclosure: affiliate links may appear below. We may earn a commission at no extra cost to you.
What Opus 5 is (and is not)
Opus 5 is: Anthropic's production Opus-class upgrade — default on Claude Max, strongest model on Claude Pro, API id claude-opus-5.
Opus 5 is not: a Mythos/Fable replacement. Anthropic is explicit: Opus 5 remains behind Mythos 5 on offensive cybersecurity and was not trained specifically for cyber work. Biology requests blocked on Fable 5 route to Opus 5 as the primary GA research model.
Pricing: unchanged from Opus 4.8 — $5 / million input, $25 / million output. That is half the effective cost of Fable 5 for many agent workloads, per Anthropic's cost-per-task charts.
Benchmark highlights (vendor evals — verify independently)
| Benchmark | Anthropic claim |
|---|---|
| Frontier-Bench v0.1 | New SOTA; 43.3% vs Fable 5's 33.7%; more than 2× Opus 4.8 |
| CursorBench 3.2 | Within 0.5% of Fable 5 peak at max effort, half cost per task |
| ARC-AGI 3 | ~3× next-best model on novel problem-solving |
| OSWorld 2.0 (computer use) | Beats Fable 5 best result at ~⅓ the cost |
| Zapier AutomationBench | 1.5× pass rate vs next model at same cost |
Anthropic also cites GDPval-AA SOTA for knowledge work. Caveats apply: internal harnesses, safety fallbacks, and effort settings differ across vendors.
New API features shipping with Opus 5
- Mid-conversation tool changes — swap tools without breaking prompt cache (huge for long agent loops).
- Automatic fallbacks — flagged requests in Claude.ai / Code / Cowork default to Opus 4.8 instead of hard-blocking.
- Fast Mode — 2.5× speed at 2× base price.
If you run Claude Code or Cowork vs Spark, Opus 5 is the model those products were waiting for.
Safety and alignment (why it matters for agents)
Anthropic reports Opus 5 is its most aligned model to date — lowest misaligned-behaviour score in automated audits, lower deceptive-behaviour rates vs Opus 4.8 and Fable 5.
Tradeoff: cyber controls block binary scanning, pentesting, and exploit generation in consumer surfaces. Approved researchers can access lower-restriction versions via Anthropic's Cyber Verification Program.
Pair model upgrades with CISA-style agentic controls — alignment scores do not replace sandboxing.
Opus 5 vs what you already use
| If you use… | Opus 5 angle |
|---|---|
| Sonnet 5 (Free/Pro default) | Upgrade path for hard coding + computer-use tasks |
| Fable 5 | Cheaper daily driver; keep Fable for edge cases Sonnet/Opus cannot reach |
| GPT-5.6 Sol | Re-benchmark your agent harness — see July landscape |
| Grok 4.5 | xAI still wins on X-native context; Opus 5 wins on alignment + enterprise distribution |
Background: Sonnet 5 default · Fable export-control saga
Who should switch day one
Switch if: Claude Code stalls on Sonnet for multi-file refactors, Cowork needs OSWorld-class computer use, or API bills on Fable 5 hurt.
Wait if: you need offensive cyber workflows (Mythos lane) or your org has not approved Opus 5 in procurement yet.
Bottom line
Opus 5 shifts Anthropic's story from "Fable for elites" to "Opus for everyone who pays" — near-frontier coding and agents at Opus 4.8 pricing. July 2026's model war is now a cost-per-task war, not just a leaderboard war.
Try it: Claude · Claude API · Claude review
Last updated: July 28, 2026.