Claude Opus 4.7 Goes GA: 3.75MP Vision, 87.6% SWE-bench Verified

87.6% on SWE-bench Verified + 3.75MP vision

On April 16, 2026, Anthropic shipped Claude Opus 4.7 as generally available. Live across the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. Pricing matches Opus 4.6 — $5 per million input tokens, $25 per million output tokens. The 1M-token context window is included at standard pricing, no long-context premium.

Two numbers drive the release.

87.6% on SWE-bench Verified. Image input max of 2,576px on the long edge (~3.75MP).

Together, they tell you where this model is pointed.

Why this matters — and how we got here

The Opus 4 line, so far

Claude Opus 4 launched in May 2025 as Anthropic's first frontier model explicitly built for agent work — extended autonomous coding sessions, long-horizon tool use. Since then, roughly-quarterly minor updates: Opus 4.1 (August), 4.5 (November), 4.6 (January 2026). Opus 4.7 continues that cadence.

SWE-bench Verified progression:

Version	Release	SWE-bench Verified	Key change
Opus 4	2025-05	72.5%	Agent mode, 1M context
Opus 4.1	2025-08	74.5%	Tool-call reliability
Opus 4.5	2025-11	79.4%	Long-running agent stability
Opus 4.6	2026-01	82.0%	Vision by default
Opus 4.7	2026-04	87.6%	High-res vision + cyber safeguards

SWE-bench Verified measures whether a model can fix real GitHub issues with a working patch. 87.6% means the model produced a passing patch on 87 out of 100 tasks drawn from real issue tickets. For context, GPT-4's initial SWE-bench score was in the low 20s two years ago. This is a steep curve.

Why high-resolution vision matters

Opus 4.7 lifts the maximum input image resolution from 1,568px on the long edge (about 1.15 megapixels) to 2,576px (about 3.75 megapixels). Nearly 3×.

Why resolution matters: at the lower ceiling, dense PDF screenshots routinely lose readable text, UI screenshots drop button labels and small icons, and scientific figures, charts, and maps hit a detail floor the model can't recover. Medical imaging, CAD drawings, compliance documents — all workloads where detail is the signal, not the noise.

Anthropic's announcement explicitly notes this is "the first Claude model with high-resolution image support." Paired with the computer-use agent capabilities, it meaningfully raises the ceiling on how accurately Claude can understand and manipulate what's on screen.

Breaking it down

Coding: 87.6% on SWE-bench Verified

This is top of market. Competitive comparison:

Model	SWE-bench Verified
GPT-5.4	82–84% (reported)
Gemini 3.1 Ultra	~80%
GLM-5.1 (Zhipu)	85.8% on SWE-bench Pro (harder variant)
Claude Opus 4.7	87.6%

GLM-5.1's strong performance is on SWE-bench Pro, a harder variant — apples-and-oranges comparison with standard SWE-bench Verified, so Opus 4.7 effectively sits at the frontier.

Practically: AI coding tools are no longer in "suggest a change" territory. They generate patches, run tests, iterate. The growth of Cursor, Claude Code, and related tools sits on top of model improvements like this one.

Cybersecurity safeguards

Opus 4.7 ships with automatic detection and blocking of requests deemed to have prohibited or high-risk cybersecurity applications. This is a direct response to the "vibe-coding malware" concerns that have been simmering for months — roughly 60–65% of AI-generated code ships with some form of vulnerability, according to prior reporting from The Hacker News.

Building the block at the model level rather than the product wrapper layer is a design choice with real consequences. It makes safety harder to bypass but also means legitimate security research may bump into friction.

Mythos, still

The most interesting fact about this release isn't Opus 4.7 itself. It's what's missing: Mythos. Anthropic explicitly notes in the 4.7 announcement that "Opus 4.7 is not our most powerful model — that's Mythos, currently undergoing testing by a limited number of users."

External estimates position Mythos as a "step change" above Opus 4.6. But Anthropic has held the public release over cybersecurity-safety concerns, and some reports question whether Mythos will ever ship publicly in its current form.

Claude Mythos Preview went to select enterprise customers in early April. It reportedly showed striking capability in offensive cybersecurity — zero-day vulnerability discovery in particular. NPR's April coverage described Project Glasswing, an internal effort using Mythos to scan large open-source codebases and surface flaws that commercial tools had missed.

The wider picture: frontier model field

Opus 4.7 ships into an unusually crowded April.

This month's frontier-class releases:

• Google Gemini 3.1 Ultra — 2M context, early April
• OpenAI GPT-5 Turbo — April 7, native image/audio generation
• Meta Llama 4 — April 5, MoE architecture, Scout variant at 10M context
• Google Gemma 4 — April 2, Apache 2.0 open family
• Anthropic Claude Opus 4.7 — April 16, GA
• Zhipu GLM-5.1 — MIT-licensed 744B MoE

Six frontier-class models in a month. For users, the choice frame has shifted from "pick the smartest one" to "pick the right one for the task."

Strengths by model:

• Claude Opus 4.7 → coding, agents, high-res vision
• GPT-5 Turbo → multimodal generation (image, audio), consumer
• Gemini 3.1 Ultra → ultra-long context, video understanding
• Llama 4 Scout → 10M context + open weights
• Gemma 4 → local execution on a single GPU
• GLM-5.1 → open source + SWE-Bench Pro strength

What this means for you

If you're a developer

If you use Cursor, Claude Code, Aider, Continue, or similar tools, this update is an immediate quality bump — especially on large refactors and hard bug fixes where test-pass rate matters. Pricing didn't change, so there's no cost trade-off.

High-res vision is an upgrade for "design to code" workflows. Drop a high-resolution Figma screenshot into Claude and expect meaningfully better HTML/CSS generation.

If you're in enterprise

Built-in cybersecurity safeguards simplify compliance review. For regulated industries — finance, healthcare — Opus 4.7 is now a strong default "responsible choice" in an AI coding tool procurement decision.

If you work on competing models

Anthropic hitting 87.6% without shipping Mythos suggests a real overhang. When Mythos does land, it will reset competitive targets. Expect OpenAI and Google to recalibrate timing and capability targets on GPT-6 and Gemini 4.

Further reading

• Introducing Claude Opus 4.7 — Anthropic
• Claude Opus 4.7 is generally available — GitHub Changelog
• Anthropic's Claude Opus 4.7 Released: All You Need to Know — FelloAI
• Anthropic's Claude Opus 4.7 Release April 2026 and the Mythos Secret — RoboRhythms
• How AI is getting better at finding security holes — NPR