Anthropic Releases Claude Opus 4.7 for Safer Multimodal Workflows

Anthropic released Claude Opus 4.7, labeled its most capable "generally available" model for complex reasoning, agentic coding, and multimodal tasks, while reaffirming that Claude Mythos Preview remains the company's stronger, privately shared model. Anthropic acknowledged that "Opus 4.7 doesn't even advance the company's 'capability frontier,' since Claude Mythos Preview received higher results on every relevant evaluation." Opus 4.7 is launching with built-in cybersecurity safeguards and targeted deployment paths to learn safe operations before any broader Mythos-class rollout.

Opus 4.7 (API id claude-opus-4-7) focuses on long-horizon, tool-heavy agentic workflows and higher-fidelity vision. Key platform and model-level changes include:

• •High-resolution image support at 2576px / 3.75MP, improving low-level perception, localization, and screenshot/document understanding.
• •128k maximum output tokens and compatibility with the existing Opus toolset and platform features.
• •A new effort tuning level, xhigh, to trade compute and token spend for capability; guidance recommends high as a minimum for intelligence-sensitive work.
• •Task budgets (beta) to give the model a running token countdown for multi-step agent loops, improving graceful task completion under constrained budgets.

Opus 4.7 carries automated cyber-risk detection and blocking mechanisms; Anthropic reports it "experimented with efforts to differentially reduce these capabilities" during training. The model is rolling out across Claude products, the API, and integrations like GitHub Copilot (where it will replace Opus 4.5/4.6 in the picker), and cloud platforms including Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. GitHub announced a temporary 7.5x premium request multiplier for Copilot usage while Opus 4.7 availability ramps. Pricing signals from partner channels list $5 per million input tokens and $25 per million output tokens consistent with Opus 4.6.

Anthropic is executing a safety-first product cadence: ship the most capable generally available model they are comfortable exposing broadly, evaluate real-world safeguards, and keep the stronger Mythos family restricted to vetted partners under Project Glasswing. This dual-track approach reflects a broader industry pattern where frontier capability is tested in constrained environments before general release. For practitioners, Opus 4.7 is meaningful because it closes practical gaps in agentic coding, memory across file-system backed sessions, and multimodal document/screenshot tasks that many production workflows need today.

Expect stricter, more literal instruction-following behavior, which will require re-tuning prompts and agent harnesses. High-resolution inputs will improve extraction accuracy but increase token consumption, so downsample when fidelity is unnecessary. The xhigh effort setting and task budgets offer new levers to balance cost, latency, and capability in production agents.

Whether Anthropic's empirical findings about Mythos outperforming Opus 4.7 hold up across independent benchmarks, and how the Cyber Verification Program and Project Glasswing evolve toward any scaled Mythos-class release. Also monitor migration effects in Copilot integrations and cost-performance tradeoffs for xhigh and large image inputs.

Opus 4.7 is a useful, pragmatic advance for teams building agentic, multimodal, and long-horizon workflows, packaged with operational safeguards. It is not a leap in capability compared with Anthropic's private Mythos Preview, but it offers tangible engineering improvements practitioners can adopt now while Anthropic experiments with safer scaling paths.