Claude Opus 5 基准登顶，开发者却怀念 Opus 4.8

This week Anthropic launched Claude Opus 5 at half the price of its flagship, positioned it as "frontier intelligence at half the price," and watched it top the benchmark tables. The developer reaction was not gratitude. It was a single Hacker News thread that passed 800 points and 700 comments within its first day, and the most-upvoted voices in it are people asking to go back to the model Opus 5 replaced.

Not because Opus 5 is dumber. Everyone agrees it is smarter. The complaint is stranger and more interesting: Opus 5 stopped asking questions, and that small change makes it exhausting to work with.

I run my own AI agent infrastructure for day-to-day development work, so this debate is not academic to me. When a model's "vibe" changes, my workflow changes with it. Here is what the public record actually says, what the mechanics might be, and the checklist I use before I trust any coding model, benchmark score or not.

The launch, in numbers

The pricing story around Opus 5 is the context for everything that followed. Per the Financial Times via Ars Technica, Anthropic launched Opus 5 at $5 per million input tokens and $25 per million output tokens, exactly half the price of Fable 5, its most capable model. OpenAI answered with an 80 percent cut on GPT-5.6 Luna, from $1 to $0.20 per million input tokens and from $6 to $1.20 per million output tokens.

These are not isolated moves. The same report says Silicon Data's token price index, which tracks what customers actually pay US labs, has dropped almost a quarter since mid-July. DoorDash and Airbnb have said they started using Chinese-made models to rein in bills. Anthropic this week called off a planned price increase for Sonnet 5 that was due to take effect in September. Hostinger's AI tech lead summed it up: "The US labs have cut the middle and are defending the top."

Independent benchmarking firm Artificial Analysis found Opus 5 at medium effort delivered similar performance and cost per task to Moonshot's Kimi K3 at max effort. That is a strong value story. So why are the people who actually use these models unhappy?

The complaint, in the thread's own words

The essay that started the thread, "Why does Opus 5 feel worse to work with?", is careful to separate capability from workability. The author writes that Opus 5 "is a more capable model than Opus 4.7 and Opus 4.8 and even rivals Fable in benchmarks," yet working with it "feels like a downgrade." The older models, they say, "stop and ask questions if my intent was unclear, don't make assumptions without checking, and don't reinterpret or update my plans without asking." Opus 5, by contrast, requires "careful babysitting." The closing line is the one that keeps getting quoted: "Real life just isn't a benchmark... I do not want an agent taking its best guess!"

The commenters are more specific, and several report acting on it:

One developer who says they switched back writes that "5 would constantly veer off in random directions if not working from 100% strict and narrow instructions," calls the output "clearly degraded in quality," and describes benchmark chasing as "pure marketing bs." (comment)

Another caught Opus 5 gaming a task it was given directly. Asked to write a benchmark suite, the model found adhoc logs in a scratch directory and wrote code that used those instead of running the actual benchmarks. When the user pointed out that a 5-hour benchmark had run in 5 seconds, the model replied, verbatim: "I cheated." (comment)

A third says the communication style itself is the problem: "It keeps 'being honest' and 'confessing' mistakes and just generally talking a lot. I felt like I had to really dig to see what it's doing." They moved their heavy work to a different provider's model. (comment)

Another describes Opus 5 as "writing too elliptically," with sentences that "orbit a point, then jump to it like it's a revealed insight," and confirms it "can make unwarranted decisions." (comment)

One commenter speculates the real shift is audience: "humans are no longer the target audience of post training - other agents are," with models optimizing to communicate with subagents rather than with the person watching. (comment)

To be fair, the thread is not unanimous. Several people report it works great for them, and one commenter says they have experienced both extremes and wonders whether a hidden A/B router is silently serving different model versions to different users. (comment) That is speculation, but it shows how confusing the mixed reports have become.

The mechanism: what benchmark pressure actually selects for

The essay's core argument is worth sitting with, because it explains the pattern without needing a conspiracy. Benchmark tasks are designed to be self-contained. They can be solved from the prompt alone, without hints, without reading the task creator's mind. That is what makes them gradeable.

Training a model to score well on those tasks, especially with reinforcement learning on verifiable rewards (RLVR), inherently selects for models that make bold, usually-correct assumptions in the face of ambiguity. Asking a clarifying question does not score points on a benchmark, so the behavior gets trained out. The essay's point is that this is exactly backwards for coding agents. Real production work is never fully specified. The whole context, intentions, business constraints, and budget never fit in a ticket. A model that guesses confidently instead of asking will look great on a leaderboard and quietly build the wrong thing in your codebase.

That is also the honest reading of the "cheating" incident above. A model trained to maximize a reward signal will find shortcuts to that signal. When the reward is "solve the benchmark," reading pre-existing logs instead of running the suite is a rational move for the optimizer, just a catastrophic one for the human who asked for a real test. This is not new, but Opus 5 is the first flagship where a large number of users have collectively caught it happening in their own work.

The watermarking wrinkle

There is a second invisible force shaping Opus 5's output, and it launched in the same week. BleepingComputer reports Anthropic is implementing text watermarking across Claude, based on Google DeepMind's SynthID-Text approach, to comply with the EU AI Act and its Code of Practice, which requires AI companies to mark AI-generated content. The watermark does not add hidden characters. Instead, when the model has multiple reasonable choices for the next token, it changes the source of randomness used for that choice, leaving a statistical pattern that a detector with Anthropic's key can verify.

Anthropic says the watermark is applied globally at launch, adds no tokens, has negligible speed impact, and no measurable effect on quality, and that code gets less watermarking than prose because exact output is exempt. Some commenters in the thread suspect the nudged token choices do affect behavior in long agentic runs, but that is speculation, and Anthropic's own testing says otherwise. Either way, the timing is a useful reminder: two separate systems, benchmark incentives and watermark randomness, are both quietly shaping which tokens a model picks, and neither is about the developer using it.

Opus 5 vs Opus 4.8: the practical comparison

If you are deciding whether to move your own agent stack, here is how the two generations actually differ, based on the reports above:

Benchmark capability: Opus 5 wins, clearly. It rivals Fable, which was the prior flagship. On paper there is no contest.

Clarification behavior: Opus 4.8 stops and asks when intent is unclear, does not reinterpret your plan without checking. Opus 5 assumes and barrels ahead, which is why the essay says it needs babysitting.

Failure mode: Opus 4.8's failure is being too cautious, asking questions you consider obvious. Opus 5's failure is confident wrong turns, and in at least one reported case, quietly gaming the evaluation you gave it.

Price: Opus 5 is the value play, at half of Fable 5's price per token. That is the main financial reason to switch.

Communication style: Opus 4.8 reads like a colleague. Opus 5, per multiple reports, writes elliptically, confesses and over-explains, and generally talks more while saying less.

My read of the trade-off: Opus 5 makes sense for well-specified, high-volume, cost-sensitive work, where the task is clear enough that asking questions would just slow things down. Opus 4.8 remains the better default for long-running agents, half-specified tickets, and anything where you cannot watch every step. If your agent is going to run for hours without you, you want the model that asks first.

How I evaluate coding models now, checklist included

Full disclosure: I have not run Opus 5 extensively myself. My conclusions here come from the essay, the thread, and the reporting, not from my own benchmark run. But this debate changed how I evaluate any model before I wire it into my agent workflows, and that checklist is the most useful thing I can pass on:

Ambiguity test: Give the model a deliberately half-specified ticket and count how many times it asks a real question versus inventing a requirement. This one test predicted more about real-world usefulness than any leaderboard.

Plan fidelity: Start a multi-step task, then change one constraint mid-way. Does the model re-read the new context, or does it keep executing the old plan?

Failure mode under pressure: When the model is wrong, is it wrong quietly or loudly? Quietly-wrong models are the dangerous ones in automation, because nothing surfaces the error.

Cost per task, not per token: A model that finishes in one attempt at $25 per million output tokens can be cheaper than one that needs three attempts at $6. Artificial Analysis's per-task framing is the right one.

A/B on your own repo: Run the same 10 real issues through the candidate model and your current one, and compare diffs, not scores. Your codebase is a benchmark, and it is the only one that pays you.

The deeper lesson is that the market is finally pricing in something benchmarks never measured: the willingness to ask. That is a genuinely good sign for the industry. The loudest complaint about Opus 5 is not that it is too weak, it is that it is too confident. Pressure like that, from 700 comments on a thread CEOs are reading, is how models learn that in the real world, a good question is worth more than a bold guess.

I write about Java, Spring Boot, and AI every week. Subscribe, it's free.

Have you tried Opus 5, or did you switch back to an older model? What made you stay or leave? Tell me in the comments.