Last updated: July 2026
Key Takeaways
- One day after Claude Fable 5 returned, BridgeBench published the only before-and-after benchmark numbers so far: debugging fell from 86.2 to 25.9, refactoring from 73.6 to 38.4, and hallucination from 75.9 to 61.7.
- The most likely cause is not a dumber model. Anthropic's new safety classifier intercepts flagged requests and reroutes them to Claude Opus 4.8, and blocked or rerouted benchmark tasks appear to score as failures.
- Neither "nerfed" nor "unchanged" can be verified from outside Anthropic. A closed model behind a classifier cannot be independently audited, and that is the real story.
Claude Fable 5 came back online on July 1 after an 18-day suspension ordered by the US Commerce Department, a saga we followed from the first day in our Fable 5 suspension tracker. It took about one day for the first quantitative verdict to land. BridgeMind, the account behind the BridgeBench coding benchmark, posted before-and-after scores under the all-caps headline "FABLE 5 CAME BACK NERFED," showing the returned model losing 60 points on debugging. The post spread quickly, because it told a large group of frustrated subscribers exactly what they suspected.
The numbers are real, in the sense that they are live on BridgeBench's leaderboard right now. What they mean is a different question. Anthropic shipped the returning Fable 5 behind a new, stricter safety classifier that hands flagged requests to Claude Opus 4.8, and it warned in advance that routine coding and debugging would get caught in the near term. That one detail changes how every row of the benchmark table should be read. Here is what the scores can prove, what they cannot, and why the honest answer to "is Fable 5 nerfed" is that nobody outside Anthropic can currently know.
The Numbers Behind the "Nerfed" Claim
BridgeBench is an independent coding benchmark focused on agentic and conversational development work: debugging realistic bugs, refactoring across files, and resisting hallucination while analyzing code. During the suspension, the BridgeMind account publicly committed to rerunning every benchmark the moment access returned. On July 1 it did, listing a separate "Claude Fable 5 July 1st" entry alongside the original scores recorded before the June 12 shutdown.
| Benchmark | Fable 5 (June 12) | Fable 5 (July 1st) | Change |
|---|---|---|---|
| Debugging | 86.2 | 25.9 | -60.3 |
| Refactoring | 73.6 | 38.4 | -35.2 |
| Hallucination | 75.9 | 61.7 | -14.2 |
Source: BridgeBench leaderboards, July 1, 2026. A single independent benchmark; the July 1st figures may reflect classifier reroutes and failed outputs rather than model capability.
As of July 2, these are the only published before-and-after numbers in existence. No other major evaluation board has released a comparison of the pre-suspension and post-return model, which means one source is carrying the entire quantitative case for "nerfed." That matters on its own. This specific source's history matters too, and we will get to it. First, the context that the viral post left out.
What Anthropic Actually Changed
Anthropic's own redeployment announcement is unusually direct about what came back. Working with the US government, the company trained an improved safety classifier that targets the bypass behavior described in Amazon's report, the one that triggered the June 12 export-control directive in the first place. When a request trips the classifier, the user is notified and the request is answered by Claude Opus 4.8 instead. Anthropic stated plainly that in the near term, some routine tasks like coding and debugging would fall back to Opus 4.8, and that it would keep refining the classifier to cut false positives. Researchers from the Commerce Department's Center for AI Standards and Innovation independently tested the safeguards and, per Anthropic, agreed they are extraordinarily strong.
The same day, coverage of the return noted that an Anthropic engineer clarified the confusing launch language: only a small fraction of coding tasks should be flagged, with the vast majority running on Fable 5 as expected. The access terms changed too. As PCWorld reported, Fable 5 is included on Pro, Max, Team, and select Enterprise plans at up to 50 percent of weekly usage limits only through July 7, after which it moves to separately billed usage credits.
Hold on to the key point: Anthropic told users, in advance and in writing, that a gate now sits in front of Fable 5 and that coding work would sometimes be intercepted. The dispute is not whether the gate exists. The dispute is how often it fires.
Three Explanations for the Collapse, Ranked by Evidence
1. The classifier is intercepting the tasks
This is the strongest explanation, and it comes from the accuser. BridgeMind's own post attributes the drop to guardrails: "The new guardrails are kicking in on way too many tasks and falling back to Opus 4.8." Even the source of the nerf claim is pointing at the gate, not the weights.
The task profile fits. Debugging benchmarks ask a model to probe broken code for flaws, which is about as close as legitimate work gets to vulnerability analysis, the exact behavior the new classifier was trained to block. BridgeBench's own UI leaderboard shows "Claude Fable 5 July 1st" rows scored 0.0 and marked "Output failed for this task" while still billing roughly 81 cents for the attempt, which is consistent with blocked or rerouted calls grading as outright zeros.
The drop pattern fits too. The two hands-on-code categories collapsed by 35 and 60 points, while hallucination, a question-answering measure, fell 14. A model that genuinely got dumber would degrade more evenly across categories. A gate that fires on code execution produces exactly this lopsided profile.
2. Some of the answers came from a different model
When the classifier fires, Opus 4.8 answers. Depending on how the benchmark harness handles the reroute notification, the July 1st scores are some blend of Fable 5 completions, Opus 4.8 completions, and outright failures. That is a legitimate product measurement, because it is what a paying user experiences today. It is not a measurement of Fable 5's brain. The leaderboard identifiers add one more layer: the rerun is listed as openrouter/anthropic/claude-fable-5-july-1, meaning the test traffic passed through OpenRouter on its way to Anthropic, one more intermediary between the benchmark and the model.
3. The model itself changed
This is the claim the headline implies, and it is unproven. Anthropic says Fable 5 and Mythos 5 share the same underlying model and describes the July 1 change as a classifier layer, not new weights. No one outside Anthropic can verify that in either direction. The intellectually honest position is to treat "the weights changed" as an unproven allegation and "the weights are identical" as an unverifiable vendor statement, and to weight them accordingly.
The Source Has Been Here Before
In April 2026, the same account posted "CLAUDE OPUS 4.6 IS NERFED. BridgeBench just proved it," claiming the model's hallucination accuracy had fallen from 83.3 to 68.3 percent. A Community Notes correction and independent analysis unwound it: the original score came from a 6-task run, the retest from a 30-task run, and on the 6 tasks the two runs shared, performance was nearly identical at 87.6 versus 85.4 percent. Most of the headline swing traced to a single fabricated answer with no repeat runs, which one developer bluntly called statistical noise dressed up as a finding.
That history does not make today's claim wrong. This time the mechanism is real and vendor-admitted, and the point drops are far too large to be sampling noise alone. But the same discipline applies before anyone repeats a 60-point headline. Were the June 12 baseline and the July 1st rerun the same task set, on the same scaffolding, through the same route? How does the harness score a request that comes back with a reroute notification? Until the per-task logs separate "blocked," "answered by Opus 4.8," and "completed by Fable 5," a 60-point drop in the aggregate cannot be read as a 60-point drop in intelligence.
What would actually settle it: a completed-tasks-only comparison, the exact analysis that unwound the April claim; retests from other evaluation boards; or Anthropic publishing the classifier's real-world false-positive rate on coding work. As of July 2, none of those exist. There is also an incentive worth naming without cynicism: nerf headlines travel. The day before the collapse post, the same account amplified Anthropic's clarification and declared Fable 5 "back for the work most of us actually do." It also posted a session receipt claiming 75 percent of a $321 coding session was routed to Opus 4.8 by the new classifiers. All of those can be genuine observations. They are also all content, and they point in different directions within a 24-hour window.
Why Nobody Can Settle This From Outside
The deeper problem is that this argument is unresolvable by design. Fable 5's weights are closed. Its inference runs only on Anthropic's servers. The classifier in front of it is invisible except at the moment it announces a fallback. Users who say the model feels dumber are making an unfalsifiable claim, and Anthropic saying the model is unchanged is an unverifiable one, and only one of those two parties holds the data that could end the argument.
This is not a new pattern for this model. Fable 5 launched in June with a safeguard, disclosed only in its system card, that quietly degraded answer quality on certain tasks with no notice at all. Users caught it within a day, and Anthropic reversed course and made the fallback visible, an episode we broke down in our silent safeguards analysis. Every "did it change" dispute since inherits that history: the company has already demonstrated that model behavior can be altered remotely, invisibly, and in good faith, all at once. The past three weeks added the harder lesson that the model can also be switched off entirely by a decision made far above any user.
This is the through-line this site keeps returning to because events keep proving it: whoever controls the infrastructure controls the experience. And when a dispute breaks out over what changed, they control the evidence too.
What To Do About It
If you are using Fable 5 this week, treat code-heavy work as interception-prone until Anthropic tunes the classifier, which it says it is doing. Blocked requests notify you and fall back to Opus 4.8, and early user reports describe rerouted requests still consuming plan usage, so watch your meters. Note the calendar as well: included Fable 5 access on subscription plans ends July 7, after which it bills through usage credits. For routine coding in the meantime, Opus 4.8 and the newly released Sonnet 5 are unaffected by the new classifier, and our Sonnet 5 breakdown covers where it lands against Opus.
The structural answer is the one no vendor announcement can revoke. A model whose weights sit on your own disk cannot be reclassified overnight, silently swapped for a different model, or pulled by directive. The trade is real and worth stating honestly: no open-weight model you can run at home matches Fable-class capability on the hardest long-horizon work. But for the day-to-day majority of tasks, drafting, document work, and routine coding help, local models cleared the sufficiency bar some time ago, and their behavior changes only when you change it. Our best local models by VRAM guide covers what each hardware budget realistically runs, and our mini PC roundup for local AI covers the always-on box to run it on.
Frequently Asked Questions
Is Claude Fable 5 actually nerfed?
Unproven. The benchmark collapse is real at the product level, but the evidence points at the new safety classifier intercepting tasks and rerouting them to Opus 4.8 rather than at a weaker model. Anthropic says the underlying model is unchanged. Only one before-and-after benchmark exists, from a single source, and it has not separated blocked tasks from completed ones.
Why did Fable 5's benchmark scores drop after it returned?
The returned Fable 5 sits behind a stricter classifier that blocks flagged requests and hands them to Claude Opus 4.8. Benchmark tasks that get blocked or rerouted appear to score as failures, and debugging tasks resemble vulnerability analysis, the exact behavior the classifier targets. Code-heavy categories fell 35 to 60 points while question-answering fell 14, the signature of a gate rather than a dumber model.
Did Anthropic change Fable 5's model weights?
Anthropic says no. Its redeployment post describes the change as a new safety classifier in front of the same underlying model that Fable 5 shares with Mythos 5. No outside party can verify the weights in either direction, because the model is closed and runs only on Anthropic's infrastructure. Treat both "the weights changed" and "the weights are identical" as unverified claims.
Why does Fable 5 keep falling back to Opus 4.8?
The new classifier was trained to block the cybersecurity bypass behavior described in Amazon's report to the US government. Anthropic acknowledged it currently over-triggers, stating that some routine coding and debugging tasks would fall back to Opus 4.8 in the near term. When it happens, you get a notification and Opus 4.8 answers the request. Anthropic says it is refining the classifier to reduce false positives.
Has BridgeBench made a nerf claim before?
Yes. In April 2026 the same account claimed Claude Opus 4.6 was nerfed based on a hallucination-score drop. A Community Notes correction showed the two runs used different task counts, 6 versus 30, and that scores on the shared tasks were nearly identical. That does not disprove the Fable 5 claim, but it is why per-task data matters more than aggregate headlines.
Can a local AI model be nerfed remotely?
No. Open-weight models are files on your own disk, and their behavior changes only when you replace or modify them yourself. There is no classifier a vendor can tighten, no silent fallback to a different model, and no directive that can switch them off. That control is the core trade against the higher raw capability of closed frontier models.
