Views expressed here are my own and do not represent those of my employer or any organization I'm affiliated with.
Claude Opus 4.8 on high effort is a complete yapper.
I have never used another model that can turn a straightforward task into such an elaborate internal debate. It starts with a reasonable plan, questions the plan, proposes another plan, rejects that one, returns to the first, and then discovers a third interpretation of a requirement that was clear from the beginning.
“Actually, no.”
“But wait.”
“On second thought.”
Dude. Just do the task.
I am not asking a model to avoid thinking. I use high reasoning effort because I want it to inspect the problem, catch edge cases, and make good decisions. What I do not want is an AI agent performing a live dissertation before it edits the file I asked it to edit.
That distinction matters. Reasoning should improve the work. It should not become the work.
Overthinking is not the same as rigor
Long reasoning can look impressive because it makes the model's uncertainty visible. You can watch it consider alternatives, revisit assumptions, and narrate every possible failure mode. There are tasks where that depth is useful.
Most of my daily engineering work does not need it.
If I ask an agent to rename a component, fix a test, add a route, or update some content, I want it to read the repository, identify the relevant constraints, and start working. It should stop when it has enough information to act safely. If the implementation reveals a problem, it can adjust.
Opus 4.8 often seems unwilling to cross that line. It keeps searching for one more concern to resolve before committing to an action. The result is not always a better decision. Sometimes it is merely a delayed decision wrapped in more words.
Good reasoning reduces uncertainty until action becomes responsible. Overthinking keeps producing uncertainty after action is already reasonable.
You pay for the yapping twice
The behavior would be easier to tolerate from a cheap model. Opus is positioned and priced as a premium option, so I expect its extra intelligence to buy me leverage.
Instead, I sometimes pay once for the model to think through an unnecessarily long chain of possibilities and again with my own attention while I wait for it to get to the point. Even when the final result is good, the path can feel wildly inefficient.
That is especially frustrating in an agentic workflow. I am not sitting down for a philosophical conversation. I am trying to move a real task through a real codebase. The model has tools. It can inspect the files, run the tests, learn from the result, and correct its approach. It does not need to solve every hypothetical version of the problem before touching the actual one.
The premium experience should feel decisive, not indulgent.
Anthropic knows this can happen
Apparently, this is not just me being impatient.
Anthropic's own prompting best practices include a section called “Overthinking and excessive thoroughness.” The discussion specifically names Opus 4.6, which Anthropic says can do more upfront exploration at higher effort levels, gather extensive context, and pursue multiple lines of research without being asked. It also acknowledges that excessive thinking can inflate token usage and slow responses.
Anthropic's suggested intervention is almost comically close to what I want to tell the model:
“When you're deciding how to approach a problem, choose an approach and commit to it.”
The rest of its guidance says not to revisit the decision unless new information contradicts it. If the approach fails, course-correct then.
The model-specific guidance gets even closer to the behavior I see:
“Claude Opus 4.8 has a tendency to favor reasoning over tool calls.”
Anthropic also says Opus 4.8 decides how long a response should be based on how complex it judges the task to be. It notes that adaptive thinking can activate more often with large or complex system prompts, and that the model consumes more tokens in interactive coding sessions because it reasons more after each user turn.
Anthropic says high effort should balance intelligence and token usage, while max can be prone to overthinking. I use Opus 4.8 on high, and it still regularly crosses that line for me.
Anthropic frames this as a tunable tradeoff, and says the extra reasoning often improves results. Fair enough. But the documentation confirms that the failure mode is real: Opus can spend too many tokens thinking, delay tool use, and reason more than the task needs.
In other words, Anthropic knows Claude can be a yapper.
Opus 4.6 set a much higher bar
My frustration with Opus 4.8 is sharper because Opus 4.6 was amazing.
Opus 4.6 felt like a genuinely premium model. It could sit with a difficult problem, understand the important constraints, and produce thoughtful work without turning every decision into a ceremony. Its reasoning felt controlled. It usually spent its extra intelligence on the parts of the task that deserved it.
Opus 4.8 does not feel like an improvement in my everyday use. It may perform better on particular evaluations, and other people may have a completely different experience, but I do not think it is a very good model for the way I work. The extra deliberation rarely buys enough quality to justify the added time, tokens, and attention.
That is a personal judgment based on using these models as coding agents, not a scientific comparison. It is still the comparison that matters when I am deciding which model to reach for every day.
I have not tried Opus 5 yet, so I do not have an opinion on whether it fixes any of this. My criticism is specifically about Opus 4.8 and the regression I feel from Opus 4.6.
Sonnet 4.6 just gets to work
Claude Sonnet 4.6 on high effort has become my daily driver for exactly this reason.
Sonnet reads the task, does enough analysis to understand the shape of it, and starts making progress. It still reasons. It still notices constraints. It still changes direction when the evidence demands it. The difference is that its reasoning usually feels attached to the task instead of fascinated with itself.
That makes Sonnet feel faster even when raw generation speed is not the main difference. Progress begins earlier. I can see the agent inspecting the right files, making the requested change, and verifying the result. Its intelligence shows up in the work rather than in a monologue about the work.
For a daily driver, that rhythm matters more to me than winning the hardest benchmark. Most tasks are not once-in-a-year architecture problems. They are bounded pieces of engineering work where context, judgment, execution, and verification matter more than exploring every branch of the decision tree.
Sonnet 4.6 is good at keeping those things in proportion.
The smartest model should know when to stop thinking
There is an odd assumption in AI products that more reasoning must be better. If five minutes of thought produces a strong answer, perhaps fifteen minutes will produce an exceptional one.
That is not how good engineering works.
Experienced engineers do not consider every possible abstraction before writing code. They recognize which decisions are expensive to reverse, spend time on those, and make the rest with the information available. They use implementation and testing to create new evidence. They know that thinking has diminishing returns.
An effective coding agent needs the same judgment. It should know the difference between a dangerous ambiguity and a harmless detail. It should know when to investigate, when to ask, when to make a reasonable assumption, and when to stop narrating and execute.
That last skill is underrated. Restraint is part of intelligence.
I still understand the appeal of Opus
There are problems where I want the biggest model I can get. A messy architectural decision, an unfamiliar system, a subtle production failure, or a high-stakes review may justify more exploration. Opus can be valuable when the cost of overlooking something is much higher than the cost of waiting.
But that does not make it the best default.
The best model for everyday work is not the one that can produce the longest analysis. It is the one that reliably finds the shortest path to a correct, well-verified result. Sometimes that path requires deep thought. Often it requires reading the task carefully and getting on with it.
Right now, Sonnet 4.6 on high effort gives me that balance far more consistently. Opus 4.8 may be the model I call when I want every angle interrogated, but Sonnet is the model I trust to sit down beside me and work.
Read the task. Think enough. Do the work.
Please, for the love of tokens, stop yapping.