Keep in mind that the people who experience issues will always be the loudest.

I've overall enjoyed 4.6. On many easy things it thinks less than 4.5, leading to snappier feedback. And 4.6 seems much more comfortable calling tools: it's much more proactive about looking at the git history to understand the history of a bug or feature, or about looking at online documentation for APIs and packages.

A recent claude code update explicitly offered me the option to change the reasoning level from high to medium, and for many people that seems to help with the overthinking. But for my tasks and medium-sized code bases (far beyond hobby but far below legacy enterprise) I've been very happy with the default setting. Or maybe it's about the prompting style, hard to say

In my experience with the models (watching Claude play Pokemon), the models are similar in intelligence, but are very different in how they approach problems: Opus 4.5 hyperfocuses on completing its original plan, far more than any older or newer version of Claude. Opus 4.6 gets bored quickly and is constantly changing its approach if it doesn't get results fast. This makes it waste more time on"easy" tasks where the first approach would have worked, but faster by an order of magnitude on "hard" tasks that require trying different approaches. For this reason, it started off slower than 4.5, but ultimately got as far in 9 days as 4.5 got in 59 days.

I think this depends on what reasoning level your Claude Code is set to.

Go to /models, select opus, and the dim text at the bottom will tell you the reasoning level.

High reasoning is a big difference versus 4.5. 4.6 high uses a lot of tokens for even small tasks, and if you have a large codebase it will fill almost all context then compact often.

Glad it's not just me. I got a surprise the other day when I was notified that I had burned up my monthly budget in just a few days on 4.6

In my evals, I was able to rather reliably reproduce an increase in output token amount of roughly 15-45% compared to 4.5, but in large part this was limited to task inference and task evaluation benchmarks. These are made up of prompts that I intentionally designed to be less then optimal, either lacking crucial information (requiring a model to output an inference to accomplish the main request) or including a request for a less than optimal or incorrect approach to resolving a task (testing whether and how a prompt is evaluated by a model against pure task adherence). The clarifying question many agentic harnesses try to provide (with mixed success) are a practical example of both capabilities and something I do rate highly in models, as long as task adherence isn't affected overly negatively because of it.

In either case, there has been an increase between 4.1 and 4.5, as well as now another jump with the release of 4.6. As mentioned, I haven't seen a 5x or 10x increase, a bit below 50% for the same task was the maximum I saw and in general, of more opaque input or when a better approach is possible, I do think using more tokens for a better overall result is the right approach.

In tasks which are well authored and do not contain such deficiencies, I have seen no significant difference in either direction in terms of pure token output numbers. However, with models being what they are and past, hard to reproduce regressions/output quality differences, that additionally only affected a specific subset of users, I cannot make a solid determination.

Regarding Sonnet 4.6, what I noticed is that the reasoning tokens are very different compared to any prior Anthropic models. They start out far more structured, but then consistently turn more verbose akin to a Google model.

Today I asked Sonnet 4.5 a question and I got a banner at the bottom that I am using a legacy model and have to continue the conversation on another model. The model button had changed to be labeled "Legacy model". Yeah, I guess it wasn't legacy a sec ago.

(Currently I can use Sonnet 4.5 under More models, so I guess the above was just a glitch)

I’ve noticed the opaque weekly quota meter goes up more slowly with 4.6, but it more frequently goes off and works for an hour+, with really high reported token counts.

Those suggest opposite things about anthropic’s profit margins.

I’m not convinced 4.6 is much better than 4.5. The big discontinuous breakthroughs seem to be due to how my code and tests are structured, not model bumps.

I definitely noticed this on Opus 4.6. I moved back to 4.5 until I see (or hear about) an improvement.

For me it's the ... unearned confidence that 4.5 absolutely did not have?

I have a protocol called "foreman protocol" where the main agent only dispatches other agents with prompt files and reads report files from the agents rather than relying on the janky subagent communication mechanisms such as task output.

What this has given me also is a history of what was built and why it was built, because I have a list of prompts that were tasked to the subagents. With Opus 4.5 it would often leave the ... figuring out part? to the agents. In 4.6 it absolutely inserts what it thinks should happen/its idea of the bug/what it believes should be done into the prompt, which often screws up the subagent because it is simply wrong and because it's in the prompt the subagent doesn't actually go look. Opus 4.5 would let the agent figure it out, 4.6 assumes it knows and is wrong

Imo I found opus 4.6 to be a pretty big step back. Our usage has skyrocketed since 4.6 has come out and the workload has not really changed.

However I can honestly say anthropic is pretty terrible about support, to even billing. My org has a large enterprise contract with anthropic and we have been hitting endless rate limits across the entire org. They have never once responded to our issues, or we get the same generic AI response.

So odds of them addressing issues or responding to people feels low.

I wonder if it's actually from CC harness updates that make it much more inclined to use subagents, rather than from the model update.

Sonnet 4.5 was not worth using at all for coding for a few months now, so not sure what we're comparing here. If Sonnet 4.6 is anywhere near the performance they claim, it's actually a viable alternative.

In terms of performance, 4.6 seems better. I’m willing to pay the tokens for that. But if it does use tokens at a much faster rate, it makes sense to keep 4.5 around for more frugal users

I just wouldn’t call it a regression for my use case, i’m pretty happy with it.

> Many people have reported Opus 4.6 is a step back from Opus 4.5.

Many people say many things. Just because you read it on the Internet, doesn't mean that it is true. Until you have seen hard evidence, take such proclamations with large grains of salt.

It goes into plan mode and/or heavy multiple agent for any reasons, and hundred thousands of tokens are used within a few minutes.

Opus 4.6 is so much better at building complex systems than 4.5 it's ridiculous.

I fail to understand how two LLMs would be "consuming" a different amount of tokens given the same input? Does it refer to the number of output tokens? Or is it in the context of some "agentic loop" (eg Claude Code)?

I called this many times over the last few weeks on this website (and got downvoted every time), that the next generation of models would become more verbose, especially for agentic tool calling to offset the slot machine called CC's propensity to light the money on fire that's put into it.

At least in vegas they don't pour gasoline on the cash put into their slot machines.

Definitely my experience as well.

No better code, but way longer thinking and way more token usage.

not in my experience

I much prefer 4.6. It often finds missed edge cases more often than 4.5. If I cared about token usage so much, I would use Sonnet or Haiku.

I have often noticed a difference too, and it's usually in lockstep with needing to adjust how I am prompting.

Put in a different way, I have to keep developing my prompting / context / writing skills at all times, ahead of the curve, before they're needed to be adjusted.

Don't take this seriously, but here is what I imagined happened:

Sam/OpenAI, Google, and Claude met at a park, everyone left their phones in the car.

They took a walk and said "We are all losing money, if we secretly degrade performance all at the same time, our customers will all switch, but they will all switch at the same time, balancing things... wink wink wink"