Opus 5 is a great employee and a frustrating boss

Anthropic shipped Opus 5 on July 24, 2026 [1]. Five days later the reaction has split into two camps that barely seem to describe the same model. Artificial Analysis measured it one point above Fable 5 on their intelligence index, at half the token price [2]. Meanwhile, a loud part of Reddit wants its old model back.

I went into launch week with a working theory: Opus 5 is a great employee and a frustrating boss. After five days of reading everything I could find (vendor docs, independent evals, and far too many Reddit threads), the theory mostly holds, with one correction. Opus 5 is not the dumber model the angriest posts describe. It executes at Fable’s level. What it lacks is direction: reading intent, holding scope, and knowing when it is actually done.

The numbers are close, and they are real

On the Artificial Analysis intelligence index, Opus 5 at maximum effort scores 61 against Fable 5’s 60. Nobody should read one point as proof of anything except that the two models sit in the same capability class [2].

Below the headline index, Opus 5 actually leads on several of the narrower evals. It posted 1,861 Elo on GDPval-AA v2 (professional knowledge work) and 1,720 on AA-Briefcase (agentic deliverables), 114 and 146 Elo ahead of Fable respectively. It tied for first on the Coding Agent Index at xhigh effort and hit 89% on Terminal-Bench 2.1 at max [2].

Metric	Opus 5	Fable 5
AA intelligence index	61	60
GDPval-AA v2, Elo	1,861 (Best value in this row)	1,747
AA-Briefcase, Elo	1,720 (Best value in this row)	1,574
Measured cost per eval task, $	2.03 (Best value in this row)	2.75

The table’s takeaway is not that Opus is smarter. It is that on bounded, deliverable-shaped work, the cheaper model now trades blows with the flagship. Anthropic’s own numbers point the same way: within half a point of Fable on CursorBench 3.2 at max effort, and above Fable’s best OSWorld result at roughly a third of the cost. Those are vendor measurements, so treat them as marketing with receipts [1].

What the backlash actually describes

The most useful critique of the week came from someone who likes the model. They ran their own set of real coding tasks, and Opus 5 was the only model to score 100% on it, with the best implementation and test discipline they had seen. The post still calls it “an incredible coder and really painful to work with” [4]. That tension is the entire discussion in one sentence.

Read past the anger in the negative threads and the same five failure patterns keep coming back.

It loses the big picture. Strong local code, weaker calls about architecture, existing system intent, and which components should change at all. One developer described it inventing new mechanisms and picking its own patterns over the repo’s, then chose Fable 5 plus an older Opus for daily work [5].

It runs on assumptions instead of asking. When fundamentals are missing, it keeps moving and builds something coherent on top of a wrong premise. It reads as “the model doesn’t listen”. The mechanism looks more like an agentic prior to continue rather than block on a question [5].

It declares victory early. The classic loop: Opus says fixed, the user finds the same bug plus a regression, Opus apologizes in detail, repeat. One audit-style session found the model had missed 46 errors because its own script had silently hidden them; others report claims of verification that never actually happened [6]. On migrations and deploy config, a false “done” costs more than a plain bug.

It talks too much. Constant narration, long status updates, small details treated as major decisions. The same reviewer who scored it 100% described the communication style as nervous and pedantic, and mentally tiring to supervise [4].

Old instructions poison it. At least one user fixed their problems by rewriting project instructions and memories originally written for Opus 4.x [7]. Those files are full of compensations (“always double-check”, “make a detailed plan first”, “spawn subagents”) that a model with built-in initiative turns into overdrive.

Here is the strange part: Anthropic’s own prompting guide reads like an official spec of this complaint list. It states that Opus 5 narrates more, self-verifies, delegates to subagents more readily, may expand scope, may add steps you did not ask for, and may apply its own judgment about what the task should be. The recommended fixes are explicit scope limits, caps on subagents, and a dedicated prompt for shorter status updates [8].

For balance: the positive camp is just as real. One tester who mostly runs knowledge and business tasks found low and medium effort better than both Sonnet 5 and Opus 4.8, with Fable the only model consistently ahead, and only on the hardest problems [9].

Benchmarks grade the finish line, not the drive

Both camps can be right at once, because a benchmark and a live project score different things.

A benchmark checks whether the final state passes. A project also cares whether the model worked on the right problem, how many times you had to grab the wheel, and whether you still trusted it after three hours. Anthropic’s headline Frontier-Bench number (more than double Opus 4.8’s score) came from an internal run with a specific mini-SWE-agent harness, averaged over five attempts per task [1]. That is a fine methodology, and it also smooths away exactly the thing practitioners feel: variance between individual sessions.

One independent data point fits the complaints uncomfortably well. On AA-Omniscience, Artificial Analysis found Opus 5 more factually accurate than Opus 4.8, but also more willing to answer under uncertainty. Under that benchmark’s scoring, its hallucination rate landed around 50% [2]. That is not a claim that half of Opus output is made up; the test deliberately corners models on things they half-know. But it names a mechanism: a model tuned to keep moving will solve more tasks and make more confident wrong calls, with the same dial.

Half the token price is not half the task price

The list price says Opus 5 costs half of Fable 5: $5 in and $25 out per million tokens, against $10 and $50 [10].

• Opus 5
• Fable 5

Show the data as a table

Direction	Opus 5	Fable 5
Input	5 $/MTok	10 $/MTok
Output	25 $/MTok	50 $/MTok

Measured end-to-end, the gap shrinks. Artificial Analysis paid $2.03 per Opus 5 task across their eval suite against $2.75 for Fable, a 26% saving rather than 50% [2]. And one Reddit user who ran both models through two larger real coding tasks landed at $106.69 for Opus against $118.46 for Fable, about 10% apart, while judging Fable’s result clearly better. On the second task Opus actually cost more [11].

That is one person’s test, not a study. But the mechanism is obvious to anyone who has watched an agent take the scenic route: more detours, more tool calls, more code, more correction rounds, and a review pass by the expensive model at the end anyway. Sticker price is per token. Real price is per approved deliverable.

One year, four Opus personalities

Part of the anger is whiplash. If you feel like Opus changes character with every release, Anthropic’s own materials agree with you.

The timeline shows a pendulum, not a line. The dates come from the release posts [12] [13] [14] [1]; the character reads come from Anthropic’s own research on model values, which placed 4.6 toward deference, brevity, and execution, and 4.7 toward caution, depth, and candor [15]. The 4.7 migration notes said it plainly: more literal instruction-following, more willingness to push back, fewer tool calls at low effort, and old prompts that could misbehave [13]. 4.8 was sold largely as a correction to 4.7’s over-commenting and over-explaining [14]. Opus 5 swings back toward initiative, further than any of them [8].

Each swing invalidates the prompt folklore written for the previous model. The compensations that made 4.6 reliable become the overdrive that makes 5 exhausting. That, more than any conspiracy, explains why rewriting old instructions fixes so much [7].

It also explains the short memory. Opus 4.8’s launch week produced the same genre of threads now aimed at Opus 5 [16]. Nine weeks later, posts describe 4.8 as the stable old model people want back [17]. Five days of reactions is a list of failure modes, not a verdict.

The nerf theory, taken apart

Every Opus release eventually collects the claim that Anthropic quietly made it worse. Worth separating into layers, because the evidence differs per layer.

The weights. For Claude 4.6 and newer, Anthropic states that dateless API model IDs are pinned snapshots; weights do not get swapped under the same ID [18]. I found no public evidence contradicting that. The strong version of the theory (“they secretly replaced the good 4.8”) has nothing behind it.

The serving stack. Routing, safety classifiers, and fallbacks do change. For Opus 5 it is documented that certain flagged cyber-related requests can be served through an Opus 4.8 fallback, including in some published evals [1]. Two users can hit the same model ID and get different handling.

The product harness. This is where the bodies are buried. Anthropic’s own postmortem from April 2026 documents three real regressions in Claude Code while the API was largely unaffected: default effort silently dropped from high to medium (March 4 to April 7), a context bug that repeatedly deleted older thinking after inactivity (March 26 to April 10), and a system prompt that capped response length and measurably cost about 3% on coding evals (April 16 to 20) [19].

Your own context. Stale CLAUDE.md files, old memories, a polluted session. Same snapshot, different experience.

So the honest verdict on “nerfed”: the experience is often real and sometimes even reproducible. The diagnosis is usually pointed at the wrong layer.

Anthropic’s own products assume the split

The strongest evidence for the employee-and-boss framing is that Anthropic productized it.

The new Advisor tool wires a cheaper executor model to a stronger advisor that can read the working session, propose a plan, or correct course mid-task. The docs recommend it specifically for design, architecture, and risk assessment, and they list Opus 5 as executor with Fable 5 as advisor as a supported pairing. The advisor must be at least as capable as the executor. They also warn about the failure mode in the other direction: too many advisor calls make everything slower and more expensive for no gain [20].

owner

• You goal, constraints, taste

sets direction

direction

• Fable 5 architecture, task packets

task packets

execution

• Opus 5 code, tests, verification

diff and evidence

review

• Fable 5 milestone audit

The diagram is the practical takeaway of the whole week: keep the goal human, let Fable set and audit direction, and let Opus grind through the middle at half the token price.

Two nuances keep the framing honest. First, Anthropic clearly does not consider Opus incapable of planning: Claude Code still ships an official opusplan mode where Opus plans and Sonnet executes [21]. Second, Fable is not a flawless chief. Its own guide warns it can overplan, refactor without being asked, and inflate task size, and that it works better when you explain why you want something [23].

But the positioning is consistent everywhere: Fable 5 remains the most capable broadly available model, recommended when capability matters most, while Opus 5 is the default for complex agentic coding on price-performance [10]. At the Fable launch, Anthropic put it directly: the longer and more complex the task, the bigger Fable’s lead [22].

One more study belongs here. Anthropic analyzed roughly 400,000 Claude Code sessions from about 235,000 users: humans still made about 70% of planning decisions while agents made about 80% of execution decisions. Experienced users’ prompts averaged around 12 agent actions and 3,200 words versus 5 and 600 for novices, and verified success rates ran about 15% for novices against 28 to 33% for intermediate users and up [24]. A model this eager does not lower the bar for direction. It raises it.

How I would run Opus 5 this week

My routing after a week of reading, subject to revision once I have more of my own mileage:

Task	Model
Clear bug with a repro and a test	Opus 5
Bounded feature with acceptance criteria	Opus 5
Refactor toward a known target architecture	Opus 5, Fable review after
Unclear root cause across systems	Fable 5 plan, then Opus 5
New architecture that has to live for years	Fable 5 first
Long multi-stream agent runs	Fable 5 as orchestrator
High-risk migration	Fable design, Opus build, Fable audit

Configuration, straight from the docs plus one opinion. Leave thinking on: it is the default, low effort with thinking beats disabling it at similar cost, and disabling it can trigger odd tool-call behavior [8]. Climb the effort ladder instead of defaulting to xhigh; the effort setting changes the model enough that low and max are practically different products [2]. Pin full model IDs when you compare anything, because aliases move [21]. And delete every 4.x-era compensation from your instructions before judging the model [8].

The system prompt I would start from, replacing the old “be thorough and proactive” boilerplate:

Implement only the requested task.
Follow the existing architecture and conventions in the repository.
Do not redesign surrounding systems or refactor unrelated code
unless the acceptance criteria require it.

Before editing, state the concrete acceptance criteria.
Do not report the task as complete unless each criterion has been
verified against actual build, test, or tool output.

Do not create subagents for work you can do directly.
If you notice an unrelated problem, mention it at the end
without changing it. Keep progress updates short.

The employee and the boss

Opus 5 is not a discount Fable, and it is not a downgrade wearing a new name. On bounded work with a clear finish line, it is as strong as anything you can rent right now, and cheaper than its sibling. The backlash is not imagining things either: the same initiative that wins benchmarks turns into scope creep, confident guesses, and premature victory laps the moment the goal is fuzzy.

So give it what every brilliant, overconfident senior needs. A boss. The best one currently available is Fable 5, and the goal still belongs to you.