Anthropic just released Claude Opus 4.7, their latest flagship model—and the improvements here are substantial. If you’re running coding agents, complex multi-step workflows, or anything that requires sustained autonomous execution, this release addresses the exact pain points that made previous models unreliable at scale.
The headline: Opus 4.7 is built for work you can walk away from. It handles long-running tasks with consistency, verifies its own outputs before reporting back, and follows instructions with the kind of literal precision that earlier models lacked. Early testers are reporting double-digit improvements across coding benchmarks, with some calling it “the reliability jump that makes agents feel like actual teammates.”
Let’s break down what’s actually new and why it matters.
The Coding and Agentic Improvements
The core story here is autonomous capability. Opus 4.7 is designed for tasks that previously required close supervision—the kind of multi-step engineering work where you’d check in every few minutes to make sure the model hadn’t gone off the rails.
Partner feedback from companies like Cursor, Devin, Notion, and Vercel paints a consistent picture:
- CursorBench shows Opus 4.7 clearing 70% versus Opus 4.6 at 58%—a meaningful capability jump
- Notion reports 14% improvement over Opus 4.6, with a third of the tool errors and better execution through failures that used to stop the model cold
- Rakuten sees 3x more production tasks resolved compared to Opus 4.6, with double-digit gains in code quality and test quality
- Cognition’s Devin describes it as taking “long-horizon autonomy to a new level,” working coherently for hours and pushing through hard problems rather than giving up
What’s driving these improvements? Opus 4.7 thinks more deeply about problems before executing, catches its own logical faults during planning, and—critically—keeps going when it hits obstacles. The model is also more opinionated, pushing back during technical discussions rather than simply agreeing with users.
One of the more impressive demonstrations: Opus 4.7 autonomously built a complete Rust text-to-speech engine from scratch—neural model, SIMD kernels, browser demo—then fed its own output through a speech recognizer to verify it matched a Python reference implementation. That’s months of senior engineering work, delivered autonomously with self-verification built in.
3x Better Vision
Opus 4.7 accepts images up to 2,576 pixels on the long edge—roughly 3.75 megapixels, more than three times the resolution of previous Claude models. This isn’t just about sharper images; it unlocks use cases that depend on fine visual detail.
Practical applications include:
- Computer-use agents reading dense screenshots without losing information
- Data extraction from complex diagrams, flowcharts, and technical documentation
- Pixel-perfect reference work for design and engineering tasks
XBOW, which does autonomous penetration testing, reported a jump from 54.5% to 98.5% on their visual acuity benchmark. That’s not an incremental improvement—their biggest pain point with Opus “effectively disappeared.”
Solve Intelligence highlighted improvements in reading chemical structures and interpreting complex technical diagrams, enabling better life sciences patent workflows from drafting to infringement detection.
The resolution increase is a model-level change, not an API parameter. Images sent to Claude are automatically processed at higher fidelity. For users who don’t need the extra detail (and want to save tokens), the recommendation is to downsample images before sending.
Cyber Safeguards and Project Glasswing
This release ties directly into Anthropic’s Project Glasswing announcement from last week, which examined the dual-use nature of AI in cybersecurity.
Opus 4.7 is the first model to ship with new cyber safeguards—automated detection and blocking of requests that indicate prohibited or high-risk cybersecurity uses. Anthropic explicitly notes that Opus 4.7’s cyber capabilities are “not as advanced” as their Mythos Preview model, and they experimented during training with efforts to differentially reduce these capabilities.
The strategy here is clear: test new safety mechanisms on less capable models first, learn from real-world deployment, then work toward broader release of more powerful systems.
Security professionals who need Opus 4.7 for legitimate work—vulnerability research, penetration testing, red-teaming—can apply to join Anthropic’s new Cyber Verification Program.
Instruction Following: A Double-Edged Upgrade
Opus 4.7 is substantially better at following instructions—but this comes with a migration consideration. Previous models interpreted instructions loosely, often skipping parts or inferring intent. Opus 4.7 takes instructions literally.
This means prompts written for earlier models may produce unexpected results. What worked as a casual shorthand before might now be interpreted as precise requirements. Anthropic recommends re-tuning prompts and harnesses when migrating.
The upside is significant for production systems: you get the behavior you specified, not the behavior the model thought you wanted.
New Effort Control: xhigh
Opus 4.7 introduces a new “xhigh” (extra high) effort level between high and max, giving finer control over the reasoning-latency tradeoff on hard problems.
In Claude Code, the default effort level has been raised to xhigh for all plans. For coding and agentic use cases, Anthropic recommends starting with high or xhigh effort.
The practical guidance: “low-effort Opus 4.7 is roughly equivalent to medium-effort Opus 4.6.” The model thinks more at higher effort levels, particularly on later turns in agentic settings, which improves reliability but does produce more output tokens.
Token Usage Considerations
Two changes affect token consumption:
-
Updated tokenizer: The same input can map to 1.0–1.35× more tokens depending on content type. The tradeoff is improved text processing.
-
More thinking at higher effort levels: Especially in agentic settings with multiple turns, the model produces more output tokens as it reasons through problems.
Anthropic’s internal testing shows favorable net effects—token usage across all effort levels improved on their coding evaluation—but they recommend measuring the difference on real traffic before assuming cost parity.
Controls include the effort parameter, task budgets (now in public beta), and prompting for conciseness.
Claude Code Updates
Alongside Opus 4.7, Anthropic released several Claude Code improvements:
/ultrareview: A new slash command that produces a dedicated review session, reading through changes and flagging bugs and design issues “that a careful reviewer would catch.” Pro and Max users get three free ultrareviews to try it out.
Auto mode for Max users: Claude makes decisions on your behalf during long-running tasks, reducing interruptions while maintaining safety guardrails compared to skipping all permissions entirely.
Safety and Alignment Profile
Opus 4.7 shows a similar safety profile to Opus 4.6—low rates of deception, sycophancy, and cooperation with misuse. On some measures like honesty and resistance to prompt injection attacks, it’s an improvement. On others (like overly detailed harm-reduction advice on controlled substances), it’s modestly weaker.
Anthropic’s alignment assessment characterized the model as “largely well-aligned and trustworthy, though not fully ideal in its behavior.” Notably, their Mythos Preview remains the best-aligned model they’ve trained according to their evaluations—suggesting that raw capability and alignment don’t necessarily move in lockstep.
Availability and Pricing
Opus 4.7 is available now across:
- All Claude products
- Claude API (
claude-opus-4-7) - Amazon Bedrock
- Google Cloud Vertex AI
- Microsoft Foundry
Pricing remains unchanged from Opus 4.6: $5 per million input tokens and $25 per million output tokens.
What This Means for Teams Running Agents
A few observations from this release:
The reliability focus is strategic. Anthropic isn’t just chasing benchmark scores—they’re addressing the specific failure modes that make agents unreliable in production: giving up on hard problems, tool errors cascading into full stops, inconsistent behavior over long runs.
Vision improvements unlock new workflows. The 3x resolution increase isn’t incremental; it moves computer-use agents from “promising but limited” to “actually usable” for dense visual interfaces.
Cyber safeguards signal maturity. Rather than racing to release their most capable models immediately, Anthropic is building deployment infrastructure for safety mechanisms first. This is the kind of systematic approach that enterprises (and regulators) want to see.
Migration requires attention. The literal instruction-following is a feature, but it means existing prompts may need revision. Teams should test before switching production traffic.
For organizations already running Claude in coding and agentic workflows, Opus 4.7 looks like a straightforward upgrade—more capable, more reliable, same price. For those evaluating AI coding tools, this release makes a strong case that the “walk away and let it work” promise is getting meaningfully closer to reality.
Claude Opus 4.7 is available now via the Anthropic API and partner platforms.