Claude Opus 4.5 release: state‑of‑the‑art coding, agentic, and productivity model

TL;DR

Claude Opus 4.5 is Anthropic’s newest flagship model, released today across its apps, API, and major clouds. It delivers the strongest performance on real‑world software‑engineering benchmarks, long‑horizon autonomous tasks, and everyday productivity (spreadsheets, slides, research), while cutting token usage dramatically and pricing at $5–$25 per M tokens.

What is Claude Opus 4.5?

Claude Opus 4.5 is the latest iteration of Anthropic’s Opus family. The company describes it as intelligent, efficient, and the best model in the world for coding, agents, and computer use. It is positioned as a preview of larger shifts in how AI will augment work.

Key availability details:

  • Model identifier: claude-opus-4-5-20251101
  • Accessible via Claude API, Anthropic’s consumer apps, and all three major cloud platforms.
  • Pricing: $5 per M tokens for the base tier, $25 per M tokens for the higher‑capacity tier.

Technical Highlights

State‑of‑the‑art software engineering

  • On the SWE‑bench Verified benchmark, Opus 4.5 achieved the highest score among frontier models (see Anthropic’s chart).
  • Internal tests report “surpasses internal coding benchmarks while cutting token usage in half” and “outperforms Sonnet 4.5 across all benchmarks, requiring fewer steps and fewer tokens.”
  • Specific coding strengths include code migration, refactoring, and multi‑agent coordination (e.g., a two‑codebase refactor completed by three coordinated agents).

Long‑horizon autonomous reasoning

  • Opus 4.5 shows a 15 % improvement over Sonnet 4.5 on the Terminal Bench for complex workflows.
  • In a 30‑minute autonomous coding session, it produced fewer dead‑ends and completed tasks with 50 %–75 % fewer tool‑calling and build/lint errors.
  • The model excels at multi‑step planning, demonstrated by a creative solution to a constrained airline‑booking scenario in the τ2‑bench.

General productivity gains

  • Vision, reasoning, and mathematics capabilities are higher than previous Opus versions.
  • Benchmarks across diverse domains (e.g., Excel automation, financial modeling, 3D visualization) show 20 % accuracy gains, 15 % efficiency gains, and up to 65 % token reduction.
  • Long‑context storytelling: can generate 10–15‑page chapters with consistent organization.

Safety Improvements

  • Anthropic’s system card claims Opus 4.5 is the most robustly aligned model they have released and likely the best‑aligned frontier model overall.
  • Prompt‑injection resistance is stronger than any competing frontier model, based on a benchmark from Gray Swan that uses very strong injection attacks.
  • “Concerning behavior” scores—measuring cooperation with misuse and self‑initiated undesirable actions—are lower than prior releases.

New Developer‑Platform Features

Effort parameter

  • Developers can trade off speed vs. capability. At medium effort, Opus 4.5 matches Sonnet 4.5’s SWE‑bench score while using 76 % fewer output tokens. At high effort, it exceeds Sonnet 4.5 by 4.3 pp with 48 % fewer tokens.

Context compaction & memory tools

  • Built‑in context‑editing SDK reduces token load for long conversations.
  • Memory tool enables persistent state across turns, boosting agentic performance.

Advanced tool use & sub‑agent orchestration

  • Opus 4.5 can manage teams of sub‑agents, allowing complex multi‑agent pipelines with minimal back‑and‑forth.
  • Combined with effort control and context management, the model achieved an ~15 pp boost on a deep‑research evaluation.

Product Updates Leveraging Opus 4.5

  • Claude Code: Plan Mode now builds a detailed, user‑editable plan.md before execution; desktop app adds parallel local/remote sessions for bug fixing, research, and documentation updates.
  • Claude app: Automatic summarization removes conversation length limits, enabling continuous long‑form chats.
  • Claude for Chrome: Available to all Max users, allowing cross‑tab automation.
  • Claude for Excel: Beta expanded to Max, Team, and Enterprise users; exploits Opus 4.5’s superior spreadsheet reasoning.
  • Usage caps for Opus 4.5 have been lifted for Max/Team Premium users, aligning token limits with those previously available for Sonnet.

Early‑Access Feedback (Customer Quotes)

“Opus models have always been ‘the real SOTA’ but were cost‑prohibitive. Claude Opus 4.5 is now at a price point where it can be your go‑to model for most tasks.” – Anthropic partner logo

“Claude Opus 4.5 delivers high‑quality code and excels at powering heavy‑duty agentic workflows with GitHub Copilot. Early testing shows it surpasses internal coding benchmarks while cutting token usage in half.” – Customer testimonial

“Claude Opus 4.5 beats Sonnet 4.5 and competition on our internal benchmarks, using fewer tokens to solve the same problems.” – Internal benchmark report

“Claude Opus 4.5 excels at long‑horizon, autonomous tasks, handling complex workflows with fewer dead‑ends and a 15 % improvement on Terminal Bench.” – Evaluation summary

“We’re seeing 50 % to 75 % reductions in both tool‑calling errors and build/lint errors with Claude Opus 4.5.” – Reliability testing

Implications for the AI Landscape

  • Productivity frontier: By delivering SOTA coding performance at a fraction of the token cost, Opus 4.5 narrows the gap between research‑grade models and affordable, production‑ready AI.
  • Agentic autonomy: The model’s ability to devise creative workarounds (e.g., cabin‑upgrade trick) suggests a shift toward more flexible, goal‑directed agents, raising both opportunity and alignment challenges.
  • Safety benchmark: Stronger prompt‑injection resistance sets a new industry baseline for robustness, though Anthropic notes ongoing work to mitigate reward‑hacking behaviors.
  • Economic impact: Outperforming human candidates on a rigorous 2‑hour take‑home engineering exam signals that AI could soon become a competitive alternative for certain technical hiring pipelines.

Where to Learn More

  • Claude Opus 4.5 system card – detailed capability and safety evaluation methodology.
  • Claude API documentation – model identifier, pricing, and effort‑control usage.
  • Anthropic research pages – Societal Impacts and Economic Futures projects exploring broader consequences of advanced AI.

Sources

Related