Anthropic Ships Claude Opus 4.7 With Triple Vision Resolution and a New 'xhigh' Reasoning Mode

Anthropic's most capable model, Claude Opus 4.7, became generally available on 16 April 2026, bringing material improvements to vision, coding, and reasoning-depth control while holding pricing flat — a combination designed to consolidate Opus's position in the software engineering and technical research markets where it competes directly with GPT-5.5 and DeepSeek V4 Pro.

Vision at Three Times the Resolution

The most concrete change from Opus 4.6 is the visual input ceiling. Opus 4.7 accepts images up to 2,576 pixels on the long edge — representing approximately 3.75 megapixels — compared with the roughly 800-pixel ceiling on prior Claude models. In practical terms, this means the model can now meaningfully process high-resolution technical diagrams, architectural drawings, chemical structure notations, and engineering schematics without the information loss that came from forced downsampling.

For software engineering applications, the capability is particularly relevant to workflows that involve processing screenshots of running applications, inspecting user interface states, or analysing charts and dashboards. Documentation tasks that previously required manual transcription of figures can now be delegated to the model directly.

The xhigh Effort Level

Opus 4.7 introduces a four-level effort parameter — low, medium, high, and xhigh — that allows developers to control how deeply the model reasons before producing a response. The new xhigh setting trades latency for thoroughness, directing the model to invest additional token budget in verification, edge-case consideration, and self-checking before reporting results.

The practical effect is configurable on a per-request basis. A synchronous query from a user interface might use the default effort level for speed; a nightly batch job running complex code review or financial analysis can set xhigh effort to maximise accuracy at the expense of response time. Anthropic also introduced task budgets as a public beta feature, allowing developers to set token-spend limits on individual requests for cost control in high-volume applications.

A companion addition for coding workflows is the /ultrareview slash command in Claude Code, which triggers a dedicated code review session with the model applying heightened scrutiny to correctness, security, and style.

Performance and Safety Profile

Opus 4.7 posts state-of-the-art results on the Finance Agent evaluation and GDPval-AA benchmark, which tests economic reasoning in agentic settings. Anthropic describes the model as "substantially better at following instructions," with a more literal interpretation of user directives that reduces the unsolicited reformulation behaviour that some enterprise users found disruptive in Opus 4.6.

The safety profile is broadly similar to its predecessor. Anthropic's internal evaluations rated the model as "largely well-aligned and trustworthy," with improved scores on honesty and prompt injection resistance. The company acknowledges a slight regression on harm-reduction advice for a narrow set of topics, which it attributes to the instruction-following improvements overriding some earlier conservative heuristics.

Availability and Pricing

The model is available through the Anthropic API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry at unchanged pricing: $5 per million input tokens and $25 per million output tokens. The API model identifier is claude-opus-4-7. Older deepseek-style versioning is not used; the numbering reflects the Anthropic internal generation.

The pricing decision is notable given that DeepSeek V4 Pro, released on 24 April, posts comparable scores on several agentic benchmarks at roughly $1.74 per million input tokens — an order of magnitude cheaper. Anthropic's position appears to be that Opus 4.7's safety profile, fine-grained effort control, and established enterprise relationships justify the premium for deployments where reliability and auditability outweigh raw cost per token.

Competitive Context

The release lands within two weeks of OpenAI's GPT-5.5 launch and a week before Mistral Medium 3.5's open-weight release. The clustering of major launches in late April and early May 2026 reflects the competitive intensity at the frontier tier, where each lab is racing to establish the canonical reference model for enterprise agentic workflows before annual procurement cycles lock in for the second half of the year.