Anthropic released Claude Opus 5 on July 24, 2026, calling it a thoughtful, proactive model that approaches the frontier intelligence of Claude Fable 5 at about half the price. According to the company’s announcement, Opus 5 is designed for everyday use: it is the new default model on Claude Max and the strongest model available on Claude Pro, while API customers can call claude-opus-5 at the same list price as Opus 4.8—$5 per million input tokens and $25 per million output tokens.
The launch matters less as a raw “bigger number” event and more as a product-default shift. Anthropic is telling paying users that the practical daily driver for coding, knowledge work, and agentic workflows is now Opus 5, even though higher-tier frontier systems remain in the lineup for specialized or risk-sensitive tasks.
Where Opus 5 leads—and where it trails
Anthropic says Opus 5 reaches new state-of-the-art results on coding and knowledge-work evaluations such as Frontier-Bench and GDPval-AA, while remaining behind Mythos 5 on cybersecurity tasks. On Frontier-Bench v0.1, the company reports that Opus 5 surpasses other models and more than doubles Opus 4.8’s performance at a lower cost per task. On CursorBench 3.2 at max effort, Anthropic says Opus 5 lands within about 0.5 percent of Fable 5’s peak score at roughly half the cost per task.
Broader bench claims in the announcement include roughly 3× the next-best score on ARC-AGI 3, about 1.5× the next-best pass rate on Zapier AutomationBench at comparable cost, and cost-efficient leadership on OSWorld 2.0 computer-use tasks—reportedly beating Fable 5’s best result at just over one-third the cost. Life-sciences internal evaluations also improve versus Opus 4.8, with notable gains on organic chemistry and protein-related tasks.
Why “default model” is the real product story
Frontier launches often emphasize peak intelligence. Anthropic’s Opus 5 pitch emphasizes usable intelligence per dollar and per minute. An effort setting lets customers trade depth for speed and cost. Fast mode, priced at twice base rates on the Claude platform and via Claude Code usage credits, runs around 2.5× default speed. For teams that live in IDEs, automation tools, and multi-step agents, those economics can matter more than a marginal win on an elite bench.
Early-access anecdotes in the announcement stress verification behavior: rebuilding geometry from raw pixels when a drawing could not be viewed directly, fixing root causes instead of symptoms in open-source tooling, and building test harnesses when live validation feeds were missing. Partner quotes from coding agents, productivity platforms, finance workflows, and legal tools repeatedly cite steadier long-horizon work, clearer judgment, and fewer wasted tool calls.
Alignment, safety, and the Mythos comparison
Anthropic says pre-deployment automated behavioral audits found Opus 5 to be its most aligned model to date, with stronger adherence to Claude’s Constitution than Opus 4.8, Sonnet 5, or Fable 5, plus lower deceptive-behavior rates and reduced susceptibility to misuse tricks. On an overall misaligned-behavior score, Opus 5 is reported at 2.3, the lowest among recent Claude models cited in the post.
On dual-use risk, Anthropic states Opus 5 does not advance the frontier in the most sensitive biology and offensive cyber capabilities and remains behind Mythos 5 there. The company says it intentionally avoided training Opus 5 on cyber tasks, yet general capability gains still improved vulnerability finding—while exploit development remains substantially weaker than Mythos 5 on Anthropic’s OSS-Fuzz evaluation framing. Safeguards allow beneficial cyber and biology uses with classifiers that Anthropic expects to intervene far less often than on Fable 5; flagged traffic on Claude surfaces can fall back to Opus 4.8, and enterprises in the Cyber Verification Program can access a less-restricted Opus 5 variant.
Platform changes shipping with the model
Alongside the model, Anthropic released two beta platform updates: mid-conversation tool changes that do not invalidate prompt cache, and automatic API fallbacks that can route classifier-flagged Opus 5 or Fable 5 requests to another model instead of hard-blocking. Opus 5, like prior Opus models, does not add general-access data-retention requirements. Those details matter for production teams optimizing latency, compliance, and tool-heavy agent loops.
What changes for users this week
If you are on Claude Max, the default brain behind everyday chats and workflows is now Opus 5. Pro users get a stronger top model than before inside that plan’s envelope. Developers should re-benchmark coding agents, computer-use tasks, and long-running analytical jobs at multiple effort levels rather than assuming Opus 4.8 prompts transfer optimally. Security and research teams should read the system card for cyber and biology evaluations before expanding autonomous tooling.
Competitors will be judged against a new bar: not only peak scores, but whether they can match Opus 5’s combination of default-product placement, price parity with the previous Opus, and claimed alignment improvements. For Anthropic, the strategic bet is that the model most people actually use every day will define the brand more than the model that wins only the hardest specialized evals.
FAQ
When did Claude Opus 5 launch?
July 24, 2026, according to Anthropic’s announcement.
How much does Opus 5 cost?
Anthropic lists $5 per million input tokens and $25 per million output tokens—the same as Opus 4.8. Fast mode is offered at twice the base price on the Claude platform.
Is Opus 5 the most capable Claude model?
Anthropic positions it as close to Fable 5 intelligence for many useful workloads at lower cost, and as SOTA on several coding/knowledge benches, while Mythos 5 remains ahead on some cybersecurity and high-risk dual-use evaluations.
Where should developers start?
Use model ID claude-opus-5 on the Claude API and review Anthropic’s prompting guide and system card linked from the launch post.
Primary sources
- Anthropic — Introducing Claude Opus 5
- Anthropic homepage
- Anthropic docs — API model IDs, prompting, and platform feature notes
Related coverage
Readers comparing frontier defaults may also follow employee-led debates on pacing automated AI research and enterprise adoption stories from coding agents and automation platforms cited in Anthropic’s launch notes. Opus 5’s arrival sharpens a market question that will define late 2026: which model is smartest on paper, versus which model becomes the one teams actually leave on all day.
Image: Photo via Unsplash (Unsplash License)
Related reading: Microsoft Routes Copilot Workloads to Its Own MAI Models — What That Means
