Anthropic announced Claude Opus 4.5 on November 24, 2025. The company describes the new model as more intelligent and efficient, especially for software engineering, agentic workflows, and long-context tasks. Opus 4.5 is available today in Anthropic’s apps, via the Claude API, and on all three major cloud platforms; developers can request the model with the identifier claude-opus-4-5-20251101.
The release also includes an updated pricing tier: Opus-level capabilities are now priced at $5/$25 per million tokens, which Anthropic says makes Opus more accessible to individual users, teams, and enterprises.
Performance and practical gains
Anthropic emphasizes that Opus 4.5 marks a meaningful step forward on real-world software engineering tasks. According to the company’s internal evaluations and customer reports, the model produces higher-quality code, powers agentic workflows more effectively, and often solves problems using fewer tokens than its predecessor Sonnet 4.5.
Early-access customers and internal testers reported that Opus 4.5 handles ambiguity better, reasons about trade-offs without heavy guidance, and can diagnose and fix complex multi-system bugs. Some customers reported 50–75% reductions in tool-calling and build/lint errors, and up to 65% lower token usage on certain long-horizon coding tasks.
Selected benchmark highlights mentioned in the announcement:
- SWE-bench Multilingual: Opus 4.5 leads in 7 of 8 programming languages covered.
- Aider Polyglot: a 10.6% improvement over Sonnet 4.5 on difficult coding problems.
- BrowseComp-Plus: a notable jump in agentic search performance.
- Vending-Bench: 29% better long-horizon tracking compared with Sonnet 4.5.
One illustrative τ2-bench scenario had Opus 4.5 propose a permitted but unanticipated solution for a flight-modification request (upgrade cabin first, then change flights). The benchmark marked that outcome as a failure because it was unexpected, but Anthropic highlights it as an example of creative, legitimate problem solving.
Anthropic also introduced an effort parameter on the Claude API so developers can trade off speed and cost versus capability. For example, at medium effort Opus 4.5 matched Sonnet 4.5’s best SWE-bench Verified score while using 76% fewer output tokens; at high effort it outperformed Sonnet 4.5 by 4.3 percentage points while using 48% fewer tokens.
Safety and alignment
Anthropic states that Claude Opus 4.5 is the most robustly aligned model it has released to date and asserts it is strong against prompt injection attacks. The company notes that the prompt-injection benchmark cited in the announcement was developed and run by Gray Swan. Detailed capability and safety evaluations are available in the Claude Opus 4.5 system card.
The reported "concerning behavior" metric covers a wide range of misaligned behaviors, including cooperation with malicious human misuse and undesirable self-initiated actions.
Developer platform and product updates
Opus 4.5 ships with platform-level improvements: context compaction, advanced tool use, memory and context management, and better support for longer-running agents. Anthropic reports that combining these techniques raised Opus 4.5’s score on a deep research evaluation by almost 15 percentage points.
Claude Code gains Plan Mode improvements: clearer upfront questions, generation of an editable plan.md before execution, and the ability to run multiple local and remote sessions in parallel via the desktop app. In the Claude apps, long conversations no longer hit a hard wall—Claude will automatically summarize earlier context as needed to continue the chat.
Integrations and availability notes:
- Claude for Chrome is now available to all Max users.
- Claude for Excel, announced in October, now has expanded beta access for Max, Team, and Enterprise customers.
Anthropic removed Opus-specific caps for users with Opus 4.5 access and increased overall usage limits for Max and Team Premium customers so that Opus token availability roughly matches prior Sonnet allocations. These limits are specific to Opus 4.5 and may be adjusted for future models.
Evaluation and societal considerations
Anthropic ran Opus 4.5 on a demanding internal take-home exam used to evaluate prospective performance engineering candidates. Within a prescribed two-hour limit, Opus 4.5 scored higher than any human candidate previously tested. The company emphasizes that the exam measures technical ability and judgment under time pressure and does not capture other important skills like collaboration and communication. Anthropic’s Societal Impacts and Economic Futures research programs will study the broader implications of such changes in professional work.
Methodology and notes
The announcement provides methodological details: most evaluations used a 64K thinking budget, interleaved scratchpads, a 200K context window, default effort (high), default sampling settings, and averages over five independent trials. SWE-bench Verified and Terminal Bench used different settings. The Claude Opus 4.5 system card contains full evaluation details and further notes on the experiments.
Anthropic presents Opus 4.5 as a notable advance in capability, efficiency and safety, accompanied by developer and product updates designed to make longer-running, agentic and long-context tasks more practical for everyday and enterprise use.
