1. Strongest Agentic Coding Model

• State-of-the-art on Terminal-Bench 2.0 (82.7%), Expert-SWE (73.1%), and SWE-Bench Pro (58.6%), outperforming GPT-5.4 on complex coding tasks.
• Holds context across large systems, reasons through ambiguous failures, and carries changes through surrounding codebases with fewer tokens.

2. Higher Intelligence at GPT-5.4 Latency

• Co-designed, trained, and served on NVIDIA GB200/GB300 NVL72 systems to match GPT-5.4 per-token latency while performing at a significantly higher level.
• Uses fewer tokens to complete the same tasks, making it more efficient as well as more capable.

3. Powerful for Knowledge Work & Computer Use

• Scores 84.9% on GDPval (44 occupations) and 78.7% on OSWorld-Verified for autonomous computer operation.
• Excels at generating documents, spreadsheets, and reports; naturally moves across finding information, using tools, and checking output.

4. Scientific Research Co-Scientist

• Leading performance on GeneBench, BixBench, and FrontierMath; helped discover a new proof about Ramsey numbers verified in Lean.
• Strong enough to meaningfully accelerate progress at the frontiers of biomedical and mathematical research.

1. State-of-the-art software engineering

• A notable upgrade over Opus 4.6 on the hardest coding tasks, with users reporting they can hand off work that previously required close supervision.
• Early partners reported double-digit gains on real-world benchmarks - e.g., Cursor saw CursorBench jump from 58% to 70%, and Rakuten-SWE-Bench resolution tripled versus Opus 4.6.
• Handles complex, long-running tasks with rigor: plans carefully, catches its own logical faults, and verifies its outputs before reporting back.

2. Long-horizon agent reliability

• Full 1M token context window at standard pricing, with state-of-the-art long-context consistency.
• Far fewer tool errors, stronger recovery from tool failures, and better follow-through on multi-step workflows - designed for async work like CI/CD, automations, and managing multiple agents in parallel.
• Stronger file-system-based memory, retaining useful notes across long, multi-session runs.

3. Sharper instruction following and honesty

• Takes instructions literally and precisely - existing prompts may need re-tuning since earlier models were more lenient.
• More honest about its own limits: reports missing data instead of fabricating plausible-but-wrong answers, and resists dissonant-data traps that tripped up Opus 4.6.

4. Substantially improved vision and multimodal reasoning

• Accepts images up to 2,576 px on the long edge (~3.75 MP) - over 3x more than prior Claude models.
• Unlocks dense-screenshot computer use, complex diagram extraction, and pixel-perfect reference tasks.
• Stronger document reasoning for enterprise analysis (e.g., 21% fewer errors than Opus 4.6 on Databricks' OfficeQA Pro).

5. Top-tier professional knowledge work

• State-of-the-art on the Finance Agent evaluation and GDPval-AA, with tighter, more professional finance analyses, models, and presentations.
• Strong on legal work - e.g., 90.9% on BigLaw Bench at high effort, with better-calibrated reasoning on review tables and ambiguous edits.
• Noted by design-focused partners as the best model for building dashboards and data-rich interfaces.

6. Modern effort and budget controls

• Introduces a new xhigh effort level between high and max for finer control over reasoning vs. latency.
• Task budgets (public beta) let developers guide token spend across long runs.
• Recommended to start with high or xhigh effort for coding and agentic use cases.