Anthropic launches Claude Opus 5 with a lower cost per task
Anthropic just made the expensive frontier feel optional for a lot of daily work. Claude Opus 5 is live, pitched as near–Fable 5 intelligence on coding and knowledge-work evals at roughly half the per-task cost, while keeping the same $5 / $25 per million token rates as Opus 4.8. It is the new default on Claude Max and the strongest model on Claude Pro.
That is the product move: squeeze Fable-class usefulness into the Opus price lane, then ship the APIs that make agents stickier.
Pricing parity with Opus 4.8 and the half-cost claim
List price did not jump. Anthropic held Opus 5 at $5 input / $25 output per million tokens — identical to Opus 4.8 — and argues the real win is performance per dollar. On CursorBench 3.2 at max effort, Anthropic says Opus 5 lands within 0.5% of Fable 5’s peak while costing about half per task. On Frontier-Bench v0.1, it more than doubles Opus 4.8’s mean reward at a lower cost per task and leads the field in Anthropic’s internal harness.
Effort settings let buyers trade intelligence for tokens. Opus 5 also leads Anthropic’s charts on ARC-AGI 3, Zapier AutomationBench, OSWorld 2.0, and GDPval-AA at comparable spend. Treat the board as vendor evals, not a court ruling. The commercial claim is still blunt: same Opus sticker, denser results.
Fast mode and mid-conversation tool changes
Opus 5 ships with Fast mode at roughly 2.5× default speed for 2× base price on the Claude Platform and via Claude Code usage credits.
Two betas land beside the model. Developers can change which tools Claude may use mid-conversation without busting the prompt cache. Automatic safety fallbacks let flagged Opus 5 or Fable 5 requests route to another model instead of hard-blocking. Anthropic also says Opus 5 has no general-access data retention requirements, matching prior Opus policy.
Where Opus 5 ships across API and clouds
Availability is day-one broad. Model ID on the Claude API is claude-opus-5. Anthropic says the release is on all platforms, including Amazon Bedrock, Google Cloud, and Microsoft Foundry, plus Claude.ai, Claude Code, and Claude Cowork. Max users get it as the default; Pro users get it as their top model.
If you were waiting for “Opus-smart, Fable-adjacent, already in my cloud bill,” that is the pitch. The migration story is less about a new price tier and more about whether your agent stack can absorb Fast mode and mid-run tool swaps without rewriting orchestration.
System-card alignment notes vs Mythos cyber
Anthropic’s pre-deployment story is split. On an automated behavioral audit, Opus 5 scores 2.3 on overall misaligned behavior — the lowest of recent Claude models — with stronger Constitution adherence and less deception than Opus 4.8, Sonnet 5, or Fable 5. Safeguards largely track Opus 4.8, with tighter guardrails on a narrow cyber band: source-code vuln finding stays more open; binary scanning, pen-testing, and exploit generation stay blocked. Flagged product traffic can fall back to Opus 4.8; Cyber Verification Program members get a less-restricted cut.
Capability caveat: Opus 5 does not advance the dual-use frontier. It remains behind Mythos 5 on biology research and offensive cybersecurity. On OSS-Fuzz-style testing, Anthropic says Opus 5 can approach Mythos at finding vulnerabilities yet lags hard on turning them into working exploits — and the company says it still avoided training Opus 5 on cyber tasks.
Analysis: Opus 5 is Anthropic selling the everyday frontier. Fable stays the prestige tip. Mythos stays the scary tip. If the half-cost coding claim holds outside Anthropic’s harnesses, a lot of “always pin Fable” roadmaps get rewritten this week.
