Summary: Anthropic released Claude Opus 4.7 on April 16, 2026 as a reliability upgrade for long-running coding and agentic work. Early testers cited longer coherent work, higher task success, and self-verification, while vision support increased to images as large as about 3.75 megapixels. Builders should retest prompts because the tokenizer and literal instruction following can change token counts and behavior.
What’s New in Opus 4.7
Anthropic shipped Claude Opus 4.7 on April 16, 2026, positioning it as a direct successor to Opus 4.6. Better performance on the hardest software engineering tasks is the headline, yet for product builders the substance lies in reliability: this model is built to carry complex, long-running work through with greater consistency, to follow instructions more precisely, and to check its own outputs before it reports back.
The same theme comes up again and again among early testers. Devin reports that Opus 4.7 “works coherently for hours, pushes through hard problems rather than giving up.” Factory measured a 10-15% lift in task success for its Droids, noting it “carries work all the way through instead of stopping halfway.” Notion Agent calls it “the reliability jump that makes Notion Agent feel like a true teammate,” and points out that it is the first model to clear their implicit-need tests.
The point isn’t raw intelligence alone — it’s trust. Vercel observed that Opus 4.7 “even does proofs on systems code before starting work,” a behavior no earlier Claude model displayed. Hex stresses that it “correctly reports when data is missing instead of providing plausible-but-incorrect fallbacks.” When teams run agentic workflows, honesty of that kind shortens debugging and makes automation genuinely usable.
Under the Hood: Technical Details That Matter
A handful of technical changes in Opus 4.7 shape how you’ll work with it:
- Updated tokenizer: Text processing improves with the new tokenizer, but the same input can land at roughly 1.0–1.35× more tokens depending on content type. Your token costs can therefore climb even though the price per token holds steady.
- More thinking at higher effort: Raise the effort level and Opus 4.7 emits more output tokens, particularly on later turns in agentic settings. Hard problems get more reliable answers as a result, at the price of higher token usage.
- New
xhigheffort level: Sitting betweenhighandmax, this new effort setting lets you dial reasoning against latency more finely. Claude Code now defaults toxhighon every plan, and Anthropic suggests starting fromhighorxhighfor coding and agentic use cases. - Task budgets (public beta): Through the API, you can now steer how many tokens Claude spends across longer runs, which helps it decide what to prioritize.
- Higher-resolution vision: Images up to 2,576 pixels on the long edge (~3.75 megapixels) are accepted, more than three times what previous Claude models handled. That opens up work such as reading dense screenshots, complex diagrams, and pixel-perfect references.
- File system-based memory: The model puts file-based memory to better use, keeping notes across long, multi-session work, so less up-front context has to be supplied.
Practical Use Cases and Implementation
Long-Horizon Agentic Workflows
Sustained reasoning across long runs is what testers mention most often. One of them, for instance, reports that 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 the Python reference.” Autonomy at that level is something new.
To take advantage of it, build your agents to keep running longer without a human stepping in. Task budgets keep token spend in check, and for complex tasks it’s worth raising the effort level to xhigh. Bear in mind that higher effort produces more output tokens, so watch your usage.
Vision-Intensive Tasks
XBOW, which runs autonomous penetration testing, recorded a striking gain: 98.5% on its visual-acuity benchmark, against 54.5% for Opus 4.6. Solve Intelligence relies on it to read chemical structures and complex technical diagrams. For any product that involves image analysis, document review, or UI design, the higher resolution support is a straight upgrade.
Coding and Code Review
CodeRabbit reports more than 10% improvement in recall for code review, surfacing bugs that are hard to detect. On CursorBench, Opus 4.7 clears 70%, compared with 58% for Opus 4.6. The model is more thorough on coding tasks and can repair its own code along the way. As one tester puts it, it “cuts out the meaningless wrapper functions and fallback scaffolding that used to pile up.”
Migration Considerations
Coming from Opus 4.6, expect token usage to shift. The updated tokenizer plus more thinking can push token counts higher. Anthropic’s internal testing found improved token efficiency on a coding evaluation, though the company recommends measuring against real traffic. To hold costs down, use the effort parameter, task budgets, or a prompt that asks for conciseness.
Limitations and Trade-offs
Cyber Safeguards
Opus 4.7 is the first model to carry automatic detection and blocking of prohibited or high-risk cybersecurity requests. It falls under Anthropic’s Project Glasswing initiative. During training, the team experimented with differentially reducing cyber capabilities. If your work is legitimate security activity — vulnerability research, penetration testing, or red-teaming — you’ll have to join the new Cyber Verification Program before Opus 4.7 can be used for those purposes.
Safety Profile
The safety profile looks much like Opus 4.6, with concerning behavior showing up at low rates. Honesty and resistance to prompt injection both improve, while the model is modestly weaker at giving overly detailed harm-reduction advice on controlled substances. Anthropic’s alignment assessment calls it “largely well-aligned and trustworthy, though not fully ideal.”
Instruction Following Can Surprise
Because Opus 4.7 takes instructions more literally, prompts composed for earlier models may produce unexpected results. Where previous models might skip parts or interpret loosely, this one takes instructions at face value. Plan on re-tuning your prompts and harnesses to match.
Not a Mythos Replacement
Claude Mythos Preview, Anthropic’s most powerful model, remains more broadly capable than Opus 4.7. The step up from Opus 4.6 is real, but it isn’t a leap to the frontier. For the very best alignment or capabilities, Mythos stays at the top, though its release remains limited.
Concrete Takeaway
Claude Opus 4.7 is a reliability upgrade, not a magic bullet. What counts for product builders is testing it on the workflows that hurt most — above all long-running, multi-step tasks where earlier models quit or hallucinated. Begin at high or xhigh effort, hold spend down with task budgets, and re-tune prompts now that instruction following is literal. Cybersecurity teams should factor in the verification program requirement. The model is available today on the Claude API, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry, priced unchanged at $5 per million input tokens and $25 per million output tokens.
As one tester described it, Opus 4.7 “really feels like a better coworker.” That’s the practical difference: not just smarter, but more dependable. Run your own evals and see whether it holds up.