Anthropic just dropped Claude Opus 4.7, and the headline is simple: this model is designed to do real work without you babysitting it. If you’re running any kind of AI agent in production (or thinking about it), this matters. A lot.
Opus 4.7 is now generally available across the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. Same price as the last version. Better at basically everything that counts for production agent work.
What happened
- New model, same price. Claude Opus 4.7 replaces Opus 4.6 as Anthropic’s top generally available model. It costs $5 per million input tokens and $25 per million output tokens.
- Big coding upgrade. Opus 4.7 shows major improvements in advanced software engineering, especially on the hardest tasks that previously needed close human supervision.
- Better vision. The model can now see images in higher resolution, which matters for document processing, diagram reading, and interface work.
- Self-verification. Opus 4.7 devises ways to verify its own outputs before reporting back. It catches its own logical faults during planning.
- New cyber safeguards. Anthropic built in automatic detection and blocking of prohibited cybersecurity uses. Security pros can join a new Cyber Verification Program for legitimate pen testing and red-teaming.
- Not Mythos. Opus 4.7 is less powerful than Claude Mythos Preview (Anthropic’s most capable model, which remains in limited release). But it beats Opus 4.6 across a range of benchmarks.
The numbers
- 13% higher resolution rate on a 93-task coding benchmark vs. Opus 4.6, including four tasks neither Opus 4.6 nor Sonnet 4.6 could solve.
- 70% on CursorBench vs. 58% for Opus 4.6. That’s a meaningful jump in real-world coding capability.
- 3x more production tasks resolved on Rakuten-SWE-Bench compared to Opus 4.6, with double-digit gains in code quality and test quality.
- 98.5% visual acuity on XBOW’s benchmark vs. 54.5% for Opus 4.6. Not a typo. That’s nearly double.
- 14% improvement over Opus 4.6 on complex multi-step workflows, with fewer tokens used and a third of the tool errors.
- 10% to 15% lift in task success for Factory Droids, with fewer tool errors and more reliable follow-through.
- 90.9% accuracy on BigLaw Bench (Harvey’s legal benchmark) at high effort, with better reasoning on ambiguous document editing tasks.
- 0.715 overall score across six research-agent modules, tying for top score with the most consistent long-context performance of any model tested.
5 things that matter most about Claude Opus 4.7 for service businesses
- It finishes what it starts. Multiple testers report that Opus 4.7 carries work all the way through instead of stopping halfway. If you’ve ever had an AI agent bail on step 7 of a 10-step workflow, you know why this is a big deal.
- It pushes back. Several early testers noted the model brings a more opinionated perspective instead of just agreeing with the user. Replit’s president said it “pushes back during technical discussions to help me make better decisions.” That’s the difference between a tool and a teammate.
- It’s honest about what it doesn’t know. Hex’s CTO said it correctly reports when data is missing instead of making up plausible-sounding answers. For any business where accuracy matters (so, all of them), this is huge.
- Vision went from weak to strong. The jump from 54.5% to 98.5% on visual acuity means document processing, slide creation, diagram interpretation, and interface building all got dramatically better overnight.
- Lower effort gets you the same quality. Hex reported that “low-effort Opus 4.7 is roughly equivalent to medium-effort Opus 4.6.” Translation: your existing workflows get cheaper and faster without any changes.
The hot take
The real story here isn’t that a new model is slightly better on benchmarks. It’s that we’ve crossed a reliability threshold. When a model can run for hours, push through errors, verify its own work, and tell you when it doesn’t have enough data, you’re not looking at a chatbot anymore. You’re looking at a junior employee who actually follows the SOPs. Most service businesses are still using AI as a fancy autocomplete. The companies that retool their agent workflows around this level of reliability will open up a gap that’s going to be very hard to close.
The Agency OS play
If you’re already running AI agents on Opus 4.6, switching to Opus 4.7 is the easiest win you’ll get this quarter. Same price, better results, fewer failures. Update the model ID in your API calls this week. Don’t overthink it. The model identifier is claude-opus-4-7.
If you’re running a law firm, look at that 90.9% accuracy on BigLaw Bench and the ability to distinguish tricky contract provisions. That’s contract review and document editing work you can start automating with more confidence. If you’re in finance, the improved General Finance scores (0.813 vs. 0.767) and better data discipline mean your research agents will hallucinate less. If you build dashboards or client-facing reports in any vertical, the vision and design improvements are worth testing immediately.
The bigger move: audit your existing AI workflows for the tasks you pulled back to humans because the model wasn’t reliable enough. The “stops halfway” problem, the “makes stuff up when data is missing” problem, the “can’t read this PDF” problem. Opus 4.7 specifically targets all three. Make a list of those failed automations, re-run them on the new model, and measure the difference. You’ll probably find two or three workflows that are now worth deploying to production.