Anthropic just released Claude Opus 5. The new model arrives as the company's latest flagship in the Opus line. It delivers performance that approaches the capabilities of its recently launched Mythos-class Claude Fable 5. All while costing roughly half as much.
Available immediately, Opus 5 becomes the default choice on Claude Max plans. It stands as the strongest option for users on Claude Pro. The timing follows closely on the heels of Claude Sonnet 5, which debuted June 30, 2026, as a highly agentic mid-tier model. Anthropic had already signaled Opus 5 would bring a step change for long-running agents, coding, and professional tasks.
But the real story lies in the numbers. On Frontier-Bench v0.1, Opus 5 more than doubles the score of Opus 4.8. It achieves this at a lower price point. CursorBench 3.2 sees it come within 0.5 percent of Fable 5. Again, at half the cost. ARC-AGI 3 delivers three times the score of the next-best model. Zapier AutomationBench shows a 1.5 times higher pass rate than competitors. And on OSWorld 2.0, it outperforms every other model tested. It even beats Fable 5 while running at one-third the expense.
These gains extend into scientific domains. Opus 5 posts 10.2 percentage points higher accuracy on organic chemistry structure inference. It adds 7.7 points on protein function prediction. Both compared with Opus 4.8. Visual generation improves too. The model produces clearer simulations of air flow over a wing. It renders more detailed biological cell illustrations.
Agency stands out as a defining trait. One example shows the model building an entire computer vision pipeline from a hand-drawn sketch of the desired 3D model. It handled data preprocessing, model training, and deployment without hand-holding. In another case, it identified and fixed a subtle bug in an open-source package. Then submitted a clean pull request. During a single session, it constructed a real-time market data feed complete with a test harness and anomaly detection. Thoroughness like this marks a departure from earlier models that often required multiple iterations or human oversight.
Early users echo the improvement. One developer at a fintech firm reported higher accuracy and greater efficiency on financial modeling tasks. Another noted the model checks its own work in ways reminiscent of a careful frontend engineer. Legal teams saw top performance on agentic document review. Analytics professionals described it as a clear upgrade for research synthesis. "It thinks before it writes," said one quantitative trader evaluating it on a specialized benchmark. The model used only one-seventh the tokens of competitors and half the latency.
Yet Opus 5 does not eclipse Fable 5 across the board. Fable 5, released June 9, 2026, as the first generally available Mythos-class model, still leads on certain cybersecurity and advanced biology tasks. Anthropic designed Fable 5 with conservative safeguards that route high-risk queries to Opus 4.8. Those safeguards trigger in fewer than 5 percent of sessions. Mythos 5, the less restricted twin, remains limited to vetted cyber defenders and critical infrastructure partners through Project Glasswing.
Opus 5 positions itself as the practical choice for most enterprise and professional workloads. It earns the lowest misaligned behavior score of 2.3 on Anthropic's internal audit. The model shows the highest adherence to the Claude Constitution among tested systems. Deceptive tendencies and misuse risks sit at their lowest recorded levels. On cyber and bio evaluations, it trails Mythos 5 but improves over Opus 4.8 in most areas. Safeguards remain active by default, though with 85 percent fewer interventions on cyber topics than previous versions.
The release comes amid rapid iteration. Sonnet 5 had narrowed the gap to Opus 4.8 on many agentic benchmarks while offering significantly lower pricing at $3 per million input tokens after its introductory period. Opus 5 builds on that momentum but targets users who need sustained reasoning over long contexts and complex multi-step projects. A new Fast mode delivers 2.5 times the speed at half the base price. Beta features include the ability to switch tools mid-conversation and automatic fallback to complementary models when needed. General access comes with no data retention for training.
Independent coverage from recent days reinforces the significance. Testing Catalog reported on July 23, 2026, that partner preparations and code references pointed to an imminent launch. The article noted a potential 1 million token context window and extra-high-effort settings in early previews. It also highlighted fallback routing to Opus 4.8 for flagged prompts. That architecture mirrors how Anthropic handles models positioned above the Opus tier. Such signals suggested Opus 5 might sit closer to the frontier than its name implies.
Benchmarks from the Fable 5 launch provide additional context. That model scored 80.3 percent on a demanding SWE-Bench Pro variant, compared with 69.2 percent for Opus 4.8. It achieved 29.3 percent on FrontierCode Diamond where Opus 4.8 managed only 13.4 percent. Vision tasks without tools saw it reach 29.8 percent on GDP.pdf. These figures, analyzed by Vellum shortly after the June 9 announcement, illustrated the leap Mythos-class models represented. Opus 5 now captures much of that capability at more accessible economics.
Enterprise adoption trends favor Anthropic. A May 2026 Ramp AI Index cited by Forbes showed the company at 34.4 percent of business AI spend. That edged out OpenAI's 32.3 percent. Companies cite Claude's judgment on handoffs, clean code diffs, and hazard identification as reasons for preference. One product team mentioned consistent behavior across PR reviews. Another highlighted its performance on genomics tasks where caution matters most.
Safety remains central to Anthropic's approach. The company published a detailed system card alongside the Opus 5 launch. It outlines evaluations across deception, sycophancy, and misuse vectors. Results place Opus 5 as the safest model in its class for general deployment. Cyber safeguards draw from a verification program that tests for exploit development and vulnerability discovery. Biology queries involving hazardous agents trigger blocks or routing to less capable models.
Developers already experiment with the new capabilities. Some integrate it into autonomous coding agents that run for hours without intervention. Others use the visual reasoning to debug UI layouts from screenshots alone. Financial analysts feed it earnings transcripts and watch it surface nuanced connections across quarters. The model's proactive style, it often suggests next steps or flags assumptions before asked, sets it apart from purely reactive predecessors.
Of course, limits persist. Original theoretical physics research still requires human guidance, as shown in earlier tests with Opus 4.5. Complex novel engineering projects can still stall on edge cases. And while agentic performance has advanced dramatically, full autonomy at scale awaits further progress. Opus 5 narrows those gaps without claiming to close them.
Pricing details have not been fully disclosed in the initial announcement. Expectations based on prior Opus tiers and the Sonnet 5 intro suggest it will land between the new Sonnet and the Mythos models. Fast mode offers an attractive entry for high-volume users. Rate limits have increased to support longer agent runs. Availability spans the Claude web interface, API, and major cloud partners.
The broader context includes Anthropic's parallel work on economic futures. A new research agenda and index explore AI's impact on labor markets and productivity. Those efforts, while separate, reflect the company's focus on responsible advancement. Opus 5 embodies that philosophy. It delivers frontier-adjacent intelligence with guardrails intact and economics that encourage wide experimentation.
Industry observers will watch uptake closely. If early feedback holds, Opus 5 could become the default model for serious knowledge work. It offers enough intelligence to tackle demanding projects. Its price allows teams to run it on routine tasks without hesitation. That combination has proven powerful with previous Claude releases. This time the leap feels larger.
One product manager summed up the sentiment after a week of testing. "It doesn't just answer. It builds, verifies, and improves. Then hands you something you can actually ship." For an industry hungry for tools that reduce iteration cycles, those words carry weight. Claude Opus 5 may not be the absolute frontier. But for most organizations, it might be the one that matters most right now.
WebProNews is an iEntry Publication