Anthropic ships Opus 4.5: first >80% on SWE-Bench Verified, plus Chrome/Excel apps and 'endless chat' for long context

## Overview Anthropic released Opus 4.5 on November 24, 2025. It is the flagship in the Claude 4.5 family. The lineup now includes Sonnet 4.5 from September and Haiku 4.5 from October. TechCrunch reports state-of-the-art results in coding, tool use, and reasoning benchmarks. A headline milestone stands out. Opus 4.5 is the first model to surpass 80% on SWE-Bench Verified. Anthropic also highlights stronger "computer use" and spreadsheet skills. The focus is practical, hands-on workflows. New products widen access. Claude for Chrome is rolling out to Max users. Claude for Excel is available to Max, Team, and Enterprise tiers. These products showcase agentic coding and spreadsheet work in real tasks. Under the hood, memory changes target long-context operations. TechCrunch notes a new "endless chat" for paid Claude users. When you near the context limit, Opus 4.5 compresses and retains salient history automatically. Conversations avoid hard cut-offs without manual effort. Anthropic positions Opus 4.5 as a lead agent. It can coordinate fleets of sub-agents, often Haiku 4.5. That helps with multi-step tasks like exploring codebases, backtracking, and rechecking large documents. The working-memory scheme supports that behavior. Competitive pressure is strong. OpenAI released GPT-5.1 on November 12. Google followed with Gemini 3 on November 18. Anthropic's pitch emphasizes better coding, tool use, and memory behavior for reliable agent workflows. ## Why Opus 4.5 matters for Morocco Moroccan teams want dependable AI for everyday work. They need tools that streamline coding, spreadsheets, and browser tasks. Opus 4.5 targets these needs with stronger agent behavior and practical interfaces. Startups can move faster with coding assistance and automated tests. SMEs can structure spreadsheets and reconcile data with fewer errors. Public agencies can pilot citizen-facing workflows while keeping oversight on data use. The model's long-context behavior suits large dossiers and policy files. ## Benchmarks, briefly explained SWE-Bench Verified measures end-to-end software engineering tasks. Passing 80% signals consistent issue resolution across repositories. Terminal-Bench evaluates command-line proficiency. These results point to robust coding support. Tool use is another area. TechCrunch cites results across tau2-bench and MCP Atlas. Those tests probe how well models plan and call external tools. Reliable tool use matters for real workflows. General reasoning also improved. ARC-AGI 2 and GPQA Diamond are demanding evaluations. Higher scores suggest stronger reasoning under pressure. That helps with complex, multistep work. ## Chrome and Excel: practical channels for Moroccan teams Claude for Chrome lets Opus 4.5 operate inside the browser. You can read pages, summarize, and automate repetitive steps. This is helpful for research, procurement, and compliance checks. Claude for Excel targets spreadsheets directly. It helps clean data, build formulas, and audit models. With Team and Enterprise access, departments can formalize spreadsheet governance and reviews. SMEs can move from manual checks to automated quality gates. Browser and spreadsheet channels are familiar in Morocco. Many workflows still rely on Excel and web forms. A model that can act in those environments shortens training time. It also invites incremental adoption. ## Endless chat: working with long dossiers Policy files and tenders have long histories. Long context is essential for continuity. The "endless chat" feature compresses history without hard resets. Paid Claude users get automatic memory compression when nearing limits. Opus 4.5 retains salient points and prior decisions. You avoid losing context mid-project. It reduces supervision overhead during extended reviews. This behavior benefits legal teams, auditors, and procurement staff. It is also helpful in customer support with long case notes. The model can maintain continuity across a full quarter's discussions. That improves consistency and accountability. ## Agent workflows and sub-agents Anthropic frames Opus 4.5 as a lead coordinator. It can assign tasks to sub-agents, often Haiku 4.5. That division helps with scale and speed. A typical flow might look like this: - Opus 4.5 plans the project and sets goals. - Haiku 4.5 scans code folders and extracts key files. - Opus 4.5 reviews diffs, proposes fixes, and writes tests. - Haiku 4.5 reruns checks and summarizes changes. This pattern aligns with multi-step document reviews. It fits large spreadsheet audits and phased cleanup. It also suits research tasks across many web sources. The memory scheme sustains the chain of thought. ## Practical uses in Morocco's private sector Startups can use Opus 4.5 for code scaffolding and refactoring. It can generate tests for existing modules. Teams should still run CI pipelines and manual reviews. Use the model to propose fixes and document changes. Fintech SMEs can automate reconciliation in Excel. They can flag anomalies and outliers in transaction logs. The model can build pivot tables and validation rules. Staff can approve changes before publishing. Retail and logistics teams can clean inventory spreadsheets. The model can infer missing values and standardize product names. It can generate category mappings for reporting. Quality checks should be mandatory before updates. Customer support can benefit inside the browser. Use Claude for Chrome to summarize tickets and draft responses. Maintain a library of approved templates. Require human approval for escalations and refunds. ## Public sector and education scenarios Public agencies can pilot document review with long context. Opus 4.5 can cross-reference past decisions and current drafts. It can suggest inconsistencies for human review. Keep sensitive data controlled and anonymized. Education programs can explore coding assistants in labs. Students can learn debugging and testing practices. Instructors can design assignments that blend AI help and manual work. Grading rubrics should require explanation and reproducibility. Municipal services can trial citizen FAQs. The model can draft responses from approved knowledge bases. Use strict guardrails and logging. Keep final publishing under staff control. ## Governance, data protection, and risk management Morocco has active data protection oversight. Organizations should align AI use with local regulations and internal policies. Pseudonymize personal data where possible. Avoid sending sensitive identifiers to external services. Use role-based access for Claude interfaces. Keep audit logs of prompts, outputs, and approvals. Review outputs for bias and accuracy. Define redlines for regulated topics and actions. For public sector pilots, ensure procurement transparency. Publish evaluation criteria and risk assessments. Require clear rollback plans. Communicate pilot scope and performance metrics. ## Adoption playbook for Moroccan teams Start small with a narrow pilot. Pick a measurable workflow with clear success metrics. Train staff on prompt discipline and review checklists. Document recurring risks and mitigation steps. Integrate with existing tooling. For developers, keep repositories, CI, and security scanning in place. For spreadsheets, define enterprise-quality rules and automated checks. Align outputs with audit requirements. Measure value continuously. Track time saved, error rates, and rework. Compare to a manual baseline. Stop or expand based on evidence. Plan for scaling. Formalize access, governance, and training. Create a center of excellence. Share playbooks across departments and sites. ## Competitive context and buyer guidance OpenAI's GPT-5.1 and Google's Gemini 3 landed earlier in November. All vendors target reliable agent workflows. Benchmarks and memory behavior differentiate products. Morocco's buyers should test against their datasets and tools. Focus on your stack and constraints. Run sandbox trials with your repositories and spreadsheets. Evaluate long-context performance with real files. Observe how well the model handles tool calls and backtracking. Consider vendor fit for procurement and compliance. Review contract terms and logging features. Clarify support expectations and escalation paths. Demand transparent evaluation reports. ## Outlook for Morocco's AI ecosystem Opus 4.5 brings practical gains where Morocco works today. Browser and spreadsheet workflows dominate many offices. Strong coding support helps local developers and startups. Long-context memory supports policy and audit teams. Results reported by TechCrunch show performance advances. The model's agent behavior fits incremental adoption. Moroccan organizations can proceed with structured pilots. The goal is reliable, supervised automation. If pilots prove value, scale carefully. Keep governance tight and training active. Measure impacts on accuracy and cost. Build internal capability rather than outsourcing all judgment. ## Key takeaways - Opus 4.5 surpasses 80% on SWE-Bench Verified, signaling strong coding performance. - New Chrome and Excel products deliver practical agent workflows. - "Endless chat" supports long dossiers and stable conversations. - Sub-agent coordination enables multi-step tasks across code and documents. - Moroccan teams should pilot with tight governance and measurable goals.