Anthropic Releases Claude Opus 4.7, Improves Coding and Vision

Anthropic launched Claude Opus 4.7, its newest generally available model, positioned as the most capable Opus-family release for complex coding, agentic workflows, and vision tasks. The model posts measurable benchmark gains, including a 64.3% SWE-Bench Pro result and notable improvements on Terminal-Bench 2.0 and Anthropic's internal coding suites, with Anthropic citing 10-15% lifts on several task collections.

Claude Opus 4.7 (API id claude-opus-4-7) expands multimodal fidelity and developer control. Key platform and model changes include:

• •High-resolution image support up to 2576px / 3.75MP, improving low-level perception, bounding-box localization, and screenshot/document understanding with 1:1 pixel coordinates.
• •A new effort level, xhigh, for stronger reasoning vs cost tradeoffs; the Messages API exposes effort tuning and xhigh is recommended for agentic coding.
• •task budgets (beta) that let the model prioritize token spend across thinking, tool calls, and outputs so agentic loops finish gracefully.
• •Continued support for 128k max output tokens and the same tool integrations available in Opus 4.6.

Anthropic reports Opus 4.7 closes gaps to its Mythos-class systems on some tasks, coming within 1% on the GPQA Diamond exam, while still trailing Mythos on other frontier benchmarks. GitHub is rolling Opus 4.7 into Copilot Pro+, Business, and Enterprise plans, replacing Opus 4.5/4.6 over time and applying a promotional 7.5x premium request multiplier through April 30. Anthropic also tuned the model to be more decisive and opinionated in code review and synthesis tasks, reducing tool errors in multi-step workflows.

Anthropic explicitly frames Opus 4.7 as a safer, broadly usable release compared with the more capable but restricted Claude Mythos. The model includes automatic detectors that block prompts indicative of cyberattack misuse, and the company will use real-world telemetry to develop guardrails for Mythos. To accommodate legitimate security research, Anthropic is launching a Cyber Verification Program and Project Glasswing partnerships to permit vetted cybersecurity teams to run realistic threat simulations under looser constraints.

Opus 4.7 is an incremental but practical step in Anthropic's product roadmap: it improves developer-facing capabilities while preserving a deliberate safety posture. The release maps to three industry trends: higher-fidelity multimodal inputs for software engineering workflows, finer-grained runtime control for agentic systems, and an emphasis on staged exposure of frontier capabilities to manage misuse risk. Compared with competing offerings, Opus 4.7 shows gains against Opus 4.6 and competes broadly with recent releases from other providers, though Anthropic retains Mythos for more advanced internal and partner use cases.

Track how task budgets and xhigh affect latency and token costs in production; monitor telemetry from the Cyber Verification Program to see how detection thresholds evolve before Mythos-class models are widely released. Integration into IDEs and Copilot will determine how quickly these coding gains translate to developer productivity.