How Claude Opus 4.5 Broke the 80% SWE-bench Barrier, and What’s Overtaken It Since

When Anthropic released Claude Opus 4.5 in November 2025, it became the first model to clear 80% on SWE-bench Verified and could sustain 20-30 minutes of coherent unsupervised work, up from the five-minute ceiling earlier agentic models hit. That milestone reshaped how builders thought about long-horizon automation at the time. Anthropic has since shipped several more model generations, including Opus 4.6, 4.7 and 4.8, and, as of mid-2026, Sonnet 5 and Opus 5. Here’s what Opus 4.5 actually changed, and how it holds up against what’s replaced it.

Longer Runs, Less Babysitting

Context drift was the practical ceiling for agentic coding sessions on earlier models. Runs would degrade past the five-to-ten-minute mark, constraints would slip, and checkpointing became mandatory overhead. Opus 4.5 pushes that ceiling to 20-30 minutes of continuous unsupervised operation, according to Anthropic, with the model completing tasks without the usual tells of AI-generated output.

That shift changes how you design agents. Instead of breaking work into tightly managed chunks with frequent human review gates, builders can hand off larger tasks, multi-file refactors, enterprise workflow automation, and let the agent run. It won’t eliminate oversight entirely, but the balance between autonomy and control shifts meaningfully in favour of autonomy.

SWE-bench: First Past 80%

Opus 4.5 scored 80.9% on SWE-bench Verified at launch, edging out GPT-5.1 (76.3-77.9% depending on variant) and Gemini 3 Pro (76.2%) at the time. That record didn’t last: Anthropic’s own Opus 4.7 and 4.8 both moved past it within the following six months, and Sonnet 5 and Opus 5 have since become the current flagship models.

In practice, Anthropic’s official announcement reports that testers found the model handles multi-system bugs and ambiguous specs without requiring heavy human guidance. A “hybrid reasoning” mode lets developers dial effort up or down via a parameter, trading speed for depth depending on the task.

Token Efficiency and Pricing

The efficiency numbers are where it gets interesting for builders running long agents: at medium effort, Opus 4.5 matches Sonnet 4.5’s SWE-bench score while using significantly fewer output tokens, according to Anthropic. At highest effort, it exceeds Sonnet 4.5’s performance by 4.3 percentage points while also requiring fewer output tokens than a comparable run.

The 200,000-token context window also reduces how often agents need to refresh context mid-task. On large codebases or extended research workflows, that means fewer interruptions and lower total token spend per completed task. Token costs accumulate fast in long agentic runs, this architecture addresses that directly.

Use Cases Beyond Code

Coding benchmarks dominate the conversation, but Anthropic positions Opus 4.5 for a wider set of long-horizon tasks. Financial institutions are a named target: the model is suited to compliance automation, risk management and financial modelling, where multi-step reasoning over large document sets is the norm. An FIS deployment using Anthropic agents showed what that looks like in practice for AML investigations.

Life sciences applications are also cited, drug discovery research, hypothesis generation, literature synthesis. Internal operations use cases include codebase migrations and workflow automation. The model handles visual inputs including charts, graphs and technical diagrams, which extends its reach to agents working across mixed knowledge bases: code, documents and data outputs in the same session. Anthropic describes it as maintaining consistency across multi-file document tasks like presentations and spreadsheets, though how that performs in production against real enterprise data remains to be seen outside the company’s own testing.

Nine months on, Opus 4.5’s real legacy is less the specific number than the shift it proved was possible: sustained, low-supervision agentic work at production quality. That trajectory has continued past it.