IntermediateAI

Anthropic Releases Claude Opus 4.7 With Sharper Coding and 3x Vision Resolution

Anthropic released Claude Opus 4.7 on April 16, 2026 — a direct upgrade to Opus 4.6 that focuses almost entirely on one thing: making the model better at long, hard software engineering work. The new release posts a 13% lift on Anthropic’s internal 93-task coding benchmark, resolves roughly 3x more production-grade tasks on Rakuten-SWE-Bench, and triples Claude’s maximum image resolution for the first time. Pricing is unchanged at $5/$25 per million input/output tokens, and the model is available today across Claude apps, the Anthropic API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry.

Claude Opus 4.7 announcement hero image
Image credit: Anthropic

What’s new in Opus 4.7

Anthropic frames Opus 4.7 as an incremental release, not a new tier — but the gains on the hardest tasks are anything but incremental. On Anthropic’s own 93-task coding evaluation, Opus 4.7 outperforms Opus 4.6 by 13 percentage points. On Rakuten-SWE-Bench, a harder suite built from real production engineering tickets, the new model resolves roughly three times as many tasks as its predecessor. Document reasoning errors on Databricks’ OfficeQA Pro fell by 21%, and on the XBOW visual-acuity benchmark Opus 4.7 jumped from 54.5% to 98.5%.

Those numbers reflect two underlying changes. First, Opus 4.7 is the first Claude model with high-resolution image support: the maximum image dimension moves from 1,568 pixels on the long edge (≈1.15 megapixels) to 2,576 pixels (≈3.75 megapixels), roughly 3x the visual capacity of previous Claude models. That matters most for computer-use agents reading dense screenshots, for diagram analysis, and for vision-heavy engineering tasks. Second, Anthropic introduced a new xhigh effort level sitting between high and max, giving developers finer-grained control over how much the model should think before answering.

Claude Opus 4.7 benchmark comparisons across coding, vision, document reasoning, and long-context tasks
Image credit: Anthropic

Long tasks, better memory, fewer shortcuts

Anthropic says Opus 4.7 was shaped around “complex, long-running tasks” — the kind of multi-hour agentic work where earlier Claude versions tended to drift. The company highlights stronger adherence to literal instructions, better sustained reasoning over extended runs, and a file-system-based memory that holds up across multi-session work. The model also does more self-verification before reporting back, which Anthropic credits for a measurable drop in reward-hacking-style shortcuts.

Alongside the model, Claude Code picks up a new /ultrareview slash command for dedicated code review sessions (Pro and Max users get three free ultrareviews to start), Auto mode is extended to Max users, and the default effort level is raised to xhigh. On the API, task budgets — advisory token targets for agentic loops — enter public beta. Two API-side breaking changes land as well: extended thinking budgets are removed, and sampling parameters (temperature, top_p, top_k) are no longer supported on Opus 4.7.

Chart showing Opus 4.7 performance across effort levels compared to Opus 4.6
Image credit: felloai.com

The Mythos question

The release also reopens the awkward conversation Anthropic started earlier in April with Project Glasswing. In the Opus 4.7 announcement, the company acknowledges that the new model’s cyber capabilities are deliberately not as advanced as those of Claude Mythos Preview — the unreleased frontier model it decided was too dangerous to ship publicly. Anthropic says it “experimented with efforts to differentially reduce” cyber capabilities during Opus 4.7’s training, and is shipping with safeguards that automatically detect and block requests indicating prohibited or high-risk cybersecurity uses. A new Cyber Verification Program provides an authorized channel for vetted security researchers, penetration testers, and red-teamers.

Translation: Opus 4.7 is the strongest public Claude so far, but Anthropic is openly conceding its flagship trails its own unreleased model. For the rest of the market, the practical story is simpler — coding and agentic performance have moved forward at constant pricing, and Anthropic’s 1M-token context window remains included without a premium.

This post was drafted with AI assistance and reviewed by RITS staff.

Sources

Related stories

agentic ai

Anthropic Releases Claude Opus 4.6 with 1M Token Context Window

Key points Anthropic released Claude Opus 4.6 on February 5, 2026, upgrading its flagship model with a 1 million token context window in beta at unchanged pricing. The 1M token window is a fivefold expansion over Opus 4.5’s previous 200k token context window. On MRCR v2, a needle-in-a-haystack retrieval test that buries facts inside large […]

ai safety

Anthropic Launches Project Glasswing With Claude Mythos, the Model It Won’t Release

Anthropic announced Project Glasswing on April 7, 2026 — a coordinated effort to deploy its most powerful, unreleased frontier model, Claude Mythos Preview, against the world’s most critical software vulnerabilities. The catch: Anthropic has decided the model is too dangerous to release publicly, and is instead handing controlled access to a coalition of twelve launch […]

agentic ai

Introducing Claude Opus 4.5

Anthropic has announced the launch of their latest AI model, Claude Opus 4.5, on November 24 2025. (Anthropic) This release represents a major advance in capability, efficiency, and alignment for enterprise- and developer-focused AI applications. Key points Anthropic announced Claude Opus 4.5 on November 24, 2025, and it is available via the Claude apps, API, […]

Never miss a RITS update

Subscribe to the RSS feed for AI, emerging technology, and digital scholarship news curated for the NYU Shanghai community, or get in touch to suggest a story.