Claude Opus 4.7

Claude’s most capable model for reasoning and agentic coding

API

Artificial Intelligence

Development

been running Opus 4.7 in Claude Code for the past couple days and the agentic stuff is noticeably better than 4.6. it actually follows through on multi-file refactors without losing context halfway through which was my biggest complaint before. the effort control slider is nice too — i keep it on high for architecture work and drop it down for quick fixes. only gripe so far is the adaptive thinking sometimes skips reasoning on queries it probably shouldn't, but overall it's a solid step up for daily coding work

I love how it leaves little notes for clarity on what has been done and what is yet to be done. I think it helps to align the model with the users vision and contributes to the amount of time you can spend in one session. With a million tokens it truly feels like the conversation gets easier and easier the more tokens you use. Scary how human it feels after ~ 500k tokens

Multiple people here mention token consumption being brutal. What's the rough token count on a typical multi-file refactor compared to Opus 4? Trying to figure out if the quality jump justifies the cost jump before committing to it for longer sessions.

the agentic coding benchmarks look wild — curious how it handles really long-horizon tasks in practice vs. the SWE-bench numbers. anyone tried it on multi-hour agent workflows yet?

Opus 4.7 sounds impressive for complex reasoning tasks. I'm curious about the agentic coding capabilities - when Claude is running long-running tasks autonomously, I'm wondering how others find it handling decision-making any better when it encounters ambiguous requirements or edge cases. Does it ask for clarification or make intelligent assumptions any better?

I've Been testing 4.7 both coding in Terminal CLI and chatting about my projects on Claude.ai and at least it feels like there is a great leap in understanding of complex dynamics. 4.6 was already Impressive but this seems to be yet on another level. On the Claude.ai chat side I feel like Opus is now pushing me even more towards getting things done and ready for my launch and this is the first time that it took the iniative to really hash out all the angles about what we've been developing, and for the first time I didn't had to ask it to read the project files and am actually impressed in the way it understood some intricacies that I've been explaining again and again to 4.6 even when those small details have been saved in the project memory several times. Now we got straight to the point without me having to explain where we are. Would be great to know how the awareness of the context between a project's chats has changed and how it's managed? Now some chats with old details that have changed in later chats didn't become a problem that I'd have to address. Impressive. On the coding side especially with the complex code base and interactions that I'm working on, for the first time I had the same experience as with the chat side of Opus actually remembering the small details and priorities we've set and actually serving me with choises that are really toward the goals we've set and it pulled from the code base stuff that I've previously had to hash out every time to get to a proper plan. Unfortunately the update nuked my terminal chats from the past 7 months but I got over it fast because when I continued the work with Opus 4.7 and had to start hashing out some stuff that were ready to implement in serveral different chats before the nuke, we actually got those done in one go without any spoon feeding and hand holding. How have people felt about this change and am I imagining this? 😂👍 Ps. I had to change the effort level to max and token limit to 200k in terminal CLI. The json got cleared on the update so the thinking got a downgrade and at first I was disappointed

It’s a powerhouse for design. Been using it since the launch. But it consumes mad tokens

I’ve been using Claude pretty regularly for coding and problem solving, and one thing I’ve really appreciated is how well it handles longer, more complex tasks compared to most tools.

There have been quite a few times where I didn’t have to keep re-explaining context, which made a big difference when working through multi-step problems. Curious how much further 4.7 pushes this, especially around maintaining context and reasoning across longer workflows. Excited to try it out.

@Caveman included?

to quote @leerob: "I really like this model for general agentic work outside coding. It is definitely expensive though."

a product like @Edgee might be a great combo in this context imho

I tried a quick brainstorm on some strategic direction, but didn't really like the response. It was not challenging me, even with explicit instructions to do so. Curious to what others are experiencing. Could it be that this model is even more tailored to e.g., coding than Opus 4.6?

Going to start today 4.7, I have been using Opus 4.6 and have been very happy with its output and performance!

BIG STEP UP, have used it so far!! Watch out though!! Will eat your tokens LOL

The jump from Opus 4 to 4.7 in agentic coding is massive. I've been using Claude Code daily and the difference in how it handles multi-file refactors and complex debugging chains is night and day. The extended thinking really shines when you give it architectural decisions to reason through.

I‘m super excited to test it out! Do you know what the date of knowledge database is? Especially if MacOS / iOS 26 Liquid Glass code is natively supported? With 4.6 I always had to use several mcps to get the right look of my implementations… Thx!

the verification step is interesting. most models just output and hope for the best. how does Opus 4.7 actually verify its own code outputs - static analysis, test generation, or something else?

The session memory improvement is the feature I've been waiting for. Working on a large codebase with Claude Code, the biggest pain was re-explaining architectural decisions every new session. If Opus 4.7 actually retains context across multi-session projects, that alone justifies the upgrade. Curious how the new tokenizer affects costs in practice — 1.35x more tokens on the same input is worth watching.

First impression was very, very positive! As I was preparing for my launch yesterday, it pretty much saved the day! It caught errors that 4.6 was ignoring for long time, helped me design some really valuable scripts and designed some really cool graphics & flows for me.

Maybe I'm just hyped and excited, but I felt like I couldn't do it without this. Came exactly on the right time!

Anthropic released Opus 4.7 today. Same pricing as 4.6 ($5/$25 per million tokens), available across API, Bedrock, Vertex AI, and Microsoft Foundry.

What changed vs 4.6:

• Coding. Biggest gains on long-horizon software engineering tasks. Model now verifies its own outputs before reporting back.

• Vision. Accepts images up to 2,576px (~3.75MP)- over 3x more than any prior Claude. Key unlock for computer-use agents and diagram extraction.

• Instruction following. Now interprets literally. Anthropic warns: prompts tuned for 4.6 may break — re-tuning needed.

• Memory. Better at file system-based memory across long multi-session work.

• Real-world knowledge work. State-of-the-art on Finance Agent eval and GDPval-AA.

New features:

• xhigh effort level between high and max - finer control over reasoning vs. latency. Claude Code default is now xhigh for all plans.

• Task budgets in public beta on the API.

• /ultrareview in Claude Code - dedicated review session flagging bugs and design issues. Three free for Pro and Max users.

• Auto mode extended to Claude Code Max users.

Honest caveats: New tokenizer → same input maps to up to 1.35x more tokens. Opus 4.7 thinks more at higher effort levels, especially on later agentic turns. Safety profile is roughly similar to 4.6 — improvement on honesty and prompt injection resistance, modestly weaker on harm-reduction advice for controlled substances. Still less capable than Claude Mythos Preview, which remains on limited release.

Bottom line: Meaningful upgrade in the three places that matter most- agentic coding reliability, vision for computer-use agents, and knowledge work benchmarks. Solid iteration, obviously shy of Mythos.