Anthropic launches Claude Opus 5: near-Fable intelligence at half the price

Anthropic launched Claude Opus 5 on Friday, July 24, 2026, available the same day on the Claude API (`claude-opus-5`), Amazon Bedrock, Google Vertex AI, Microsoft Foundry, claude.ai, Claude Code, and Claude Cowork. It is the fourth release in the Claude 5 family, after Mythos 5 and Fable 5 (June 9) and Sonnet 5 (late June). Anthropic's own framing: 'a thoughtful and proactive model that comes close to the frontier intelligence of Claude Fable 5 at half the price.'

Pricing stays put, and that is the story. Opus 5 costs $5 per million input tokens and $25 per million output tokens, exactly what Opus 4.8 cost, and half of Fable 5's $10/$50. It becomes the default model on Claude Max and the strongest model on Claude Pro. A Fast mode (~2.5× the standard speed, at $10/$50) is available on the Claude API only. Independent press (VentureBeat, Fortune, Bloomberg) reads the launch as a shift from raw capability to the economics of everyday use: near-frontier intelligence, priced for daily work.

What's new compared to Opus 4.8? Anthropic calls it 'a step-change improvement', with the largest gains in deep reasoning, agentic and long-horizon tasks, and test-time compute scaling. The model verifies its own work and iterates until it succeeds. Anthropic explicitly advises removing old 'check your work' instructions from prompts, because they now cause over-verification. The context window is 1M tokens by default, output goes up to 128K, and office tasks (multi-sheet spreadsheets with real formulas, structured slide decks) get a visible upgrade. Opus 5 is also better at coordinating teams of subagents with writer-verifier patterns.

The benchmarks, with the honest caveat. Anthropic's own figures, mostly at max effort: 43.3% on Frontier-Bench v0.1 (agentic terminal coding; Opus 4.8 scored 18.7%, Fable 5 33.7%), an 1,861 Elo on GDPval-AA v2 for knowledge work (above Fable 5's 1,747), 30.2% on ARC-AGI-3 (novel problem solving; Opus 4.8: 1.5%), 70.6% on OSWorld 2.0 (computer use), and 79.2% on SWE-bench Pro. Note that nearly all of these numbers are self-reported by Anthropic on its own harness; independent trackers are still catching up. Anthropic itself is open about where it loses: OpenAI's GPT-5.6 Sol wins on at least one established coding benchmark, and Mythos 5 remains stronger on cybersecurity and autonomous biology research.

For API users: thinking is now on by default, a breaking change. The model decides for itself how much to think per turn; the effort parameter (low, medium, high, xhigh, max, no beta header needed) steers the depth. Disabling thinking is only possible at effort high or lower; at xhigh/max it returns a 400 error. The prompt-cache threshold drops from 1,024 to 512 tokens. Two API betas launch alongside: mid-conversation tool changes (swap tools without invalidating the prompt cache) and automatic fallbacks (requests flagged by the safety classifier get routed to another model instead of blocked).

Safety: the most aligned Claude to date, says Anthropic. Opus 5 ships under the same ASL-3 protections as Opus 4.8. On Anthropic's automated behavioral audit it scores 2.3 on overall misaligned behavior, the lowest (best) of its recent models, ahead of Opus 4.8, Sonnet 5, Fable 5 and Mythos 5. The cyber classifiers are far less restrictive than Fable 5's (Anthropic expects roughly 85% fewer interventions): finding vulnerabilities in source code is allowed, while binary scanning, pentesting and exploit generation are blocked. In claude.ai, Claude Code and Claude Cowork, flagged cyber requests fall back to Opus 4.8. One honest caveat from the system card: Opus 5 hallucinates factual claims slightly more often than Opus 4.8, despite higher overall accuracy.

What it means for Cowork users. Cowork's way of working, break a big task into subagents, plan first, check the results, is exactly where Opus 5's improved persistence, self-verification and multi-agent coordination land. Expect fewer half-finished deliverables and better spreadsheets and decks. Worth knowing: the launch itself adds no new Cowork features; the two new betas are API-only. Early customer numbers from the announcement: Harvey matched Opus 4.8's maximum-reasoning quality with 26% fewer tokens, Box measured +8% overall and +17% on due diligence, and Cursor calls it 'near Fable 5 intelligence at Opus speed and cost.'

For EU organizations, the data-residency picture is unchanged (per independent compliance sources, verify at claude.com/regional-compliance): claude.ai and the first-party API run on US infrastructure under Standard Contractual Clauses; genuine EU data residency requires AWS Bedrock EU regions (e.g. Frankfurt) or Google Vertex AI EU. Claude Cowork currently has no EU routing option. And mark the date: on August 2, 2026 the EU AI Office gains full enforcement powers over general-purpose AI models, with fines up to 3% of global turnover.

Practical rule of thumb from day one: use Opus 5 for complex, high-stakes agentic work (large refactors, financial modelling, legal analysis, multi-step research), start at effort medium and scale up only where accuracy pays. Keep Sonnet 5 as the everyday workhorse and Haiku 4.5 for volume. And clean up your prompts: remove Opus 4.8-era verification instructions and recheck your max_tokens budgets now that thinking is on by default.