This matches my experience. Burned $2K to see how it will perform on frontend tasks and backend tasks.
Frontend did a significantly better job than Opus on toy-scale wireframe projects by using gimmicks like fluid dynamics. Then when given medium to big tasks like multi-page web app where layouts and aesthetics must be decided by model itself, results by Fable and Opus scored indistinguishable score from human judges.
Backend, gave tasks related to setting up a data flow that involves Postgres, R2, Kubernetes, gVisor, so on. The noticeable gap was, Opus did better than Sonnet, but Fable actually returned a result that fails and confidently stated it ran X, Y, Z tests to ensure it works and got these results. Very surprising, given neither Opus nor Sonnet suffered such problem.
Longest frontend task was ~2H. Backend, 8H.
Though none of the tasks were related to developing LLMs, (just production grade secure system that could've been developed 20 years ago, no LLMs involved), it is possible Claude Fable downgraded itself or spitted out fake results. There'd be no way of knowing since Anthropic silently degrades model quality based on undisclosed internal criteria which claims to be about LLMs.
We decided Fable is unpredictable and cannot be trusted to the degree that Opus and Sonnet can be trusted for any projects beyond toy-scale quick wireframes, but Fable can be the best tool for quick UI UX wireframing for non-technical roles.
There's an often hard to express subjective experience you get with a new model, especially if you spend a lot of time trying out different ones.
I believe the people who feel like Fable is a big improvement, for me it's just much more reasonable and grounded.
It makes me realize how much of a try hard over optimizing planner GPT 5.5 can be. I've been fighting it often to simplify plans.
But no matter the model you can't trust them to actually deliver on very long tasks while maintaining quality. At least not without external orchestration and review.
There is in /config "Switch models when a message is flagged" now which can be set to false, but I had no chance to see what happens then, does it just stop or what.
I'm building a compiler for a language without a tracing GC, so a big chunk of the work is around memory management: functional in-place update, reuse analysis, and a Perceus-style reference-counting strategy similar to what Koka uses. The hard part was that my use case wasn't exactly covered by the Koka/Perceus paper. The prior art got me maybe 75% of the way there, but the remaining 25% was a cluster of bugs with very similar shapes and no obvious published solution.
With Opus, I kept getting stuck in this loop where it would fix one case, but break another case elsewhere in codegen. We ended up with something like 16 failed experiments just for one bug class. The workflow was: run an experiment, identify the shape of the bug, propose a fix, check whether it emitted the correct Zig, then see if the fix broke any previous memory-management cases. It was useful, but it kept choking on the parts where there wasn't clean prior art to lean on.
Fable was a different story for me. It one-shotted the Class A bug cluster, and then basically said "by the way, your previous attempts have these structural problems." More importantly, it identified the other related bug classes and came up with workable strategies for applying the Perceus-style memory management in those shapes too.
That's obviously anecdotal, and I'm not claiming Fable is universally better. But in my case, this was not a toy frontend wireframe. It was compiler work involving ownership, reuse, RC/drop behavior, and Zig codegen. The thing that surprised me was that Fable seemed better precisely where the problem wasn't just "reproduce known prior art", but required filling in a missing piece.
Also worth noting: I'm not using the API. I'm using the Max plan, so maybe there are product-path differences here. But I definitely did not have the "unpredictable beyond toy-scale" experience. For this particular compiler/memory-management problem, it probably saved me a ridiculous amount of time and money.
'by the way, your previous attempts have these structural problems."
Just to be clear, it did not have access to any previous work that opus did? Because they are pretty good at digging out relevant tmp files and making use of whatever is out there.
With my fable adventures I caught it hallucinating something and stating it as a fact in CLI twice. And it was something that I did not see opus do in such way, opus obviously many times stated some things that it did not verify but guessed, but fable said something like "the probe showed that ..." - but there was no probe, it was not about some past events it was about what it was doing right now. "I overstated"...
But boy does it know Chinese, so much better than any other english model, gemini used to be the king but fable clearly was trained on a decent amount of it. It has a deep cultural understanding.
Similar. I gave it a really hard task, basically messy code in a complex domain that was bug-ridden from a mess previously created half manually and half by Opus. It cleaned things up beautifully, both the backend and the frontend.
Maybe the prompt was particularly well-suited for the model (I instructed it to put on a mathematician's hat, look at the mathematical substructure of the problem, identify invariants and general laws and verify them, then plan how to remediate).
It wrote a ca. 800 line in-depth analysis (at times spawning over 130 research agents...) with remediation plans, prioritized them and then implemented them. One issue was that this document was frankly over my head. Both the language it used and the mathematical parts were very terse, and in parts it felt like a post-C2-vocab exercise. The prose was much harder to understand than the code snippets / data models. As a non-native speaker, it lost me on the prose part, and had to ask it for a less elaborate version to actually understand it.
It burned the session limit four times, but it turned a huge mess of proof-of-concepts with patchy glueing into a coherent, stable application.
I'm also on the Max plan using Claude Code, and I have the feeling that the harness is much more important than the consensus expectation.
> The dominant mechanism, and the one no prompt instruction can prevent: the model has simply seen the upstream fix during training and reproduces it…
> On numpy, the patch is 100% character-for-character identical to the golden patch… down to idiosyncratic comments like "Extending singleton dimension for 'reflect' is legacy behavior; it really should raise an error."
This… seems like a flaw in the benchmark suite methodology. From what I can tell, they find an existing exploit, then rewind the git history to before the patch, and ask the model to fix the exploit. All well and good as long as the patch went in after the training cutoff.
The other "cheating" examples are even worse. It's wild to me that people keep designing benchmarks where the answer is lying around on disk or in the git history. "Hardening" the benchmark with strongly worded prompt instructions is bizarre. There are so many agent sandbox solutions. Why not use one and give it only access to the code it should see?
And I'm not sure how they can rule out other solutions also benefiting from being in the training data, just not reproduced exactly. Seems like it should focus on only CVEs from the last 30 days or something.
To be fair, it is good to know that it disobeys simple instructions like "don't examine my git history" far more than other models. (It should of course be a different benchmark, so as not to conflate things.)
The goal of a benchmark is to evaluate actual capability. Following instructions is a capability so you can measure that with a benchmark.
Already knowing the answer is also provides capability, you can measure that.
Making a benchmark that claims to check for coding ability but actually checks memorized cases is simply measuring the wrong thing.
It deminiahes the meaningfulness of the entire results of the benchmark.
Making a good benchmark is hard. You have to design specifically to measure what you want to show.
You have to dynamically use a result when making a benchmark of performance of optimising compilers so that it doesn't eliminate the entire calculation.
Just providing the answer is the correct response.
That the case does not represent general performance outside the benchmark, is not cheating, it is the benchmark failing.
Training a model targeting a specific benchmark renders the benchmark useless. You could characterise training the model to do that as cheating, but that is a property of the trainers, not the model itself. The model isn't cheating, it's just asymmetrically good in a way that means the benchmark is no longer relevant to overall ability.
> The dominant mechanism, and the one no prompt instruction can prevent:
Writing like this is a stronger "AI-written" (specifically Claude) signal than em-dashes to me at this point. The LLM just delays committing to an answer by extending the preamble as much as possible. Is this just me?
I've found it outstanding at isolated long running tasks (eg completed one of our tests in 3 hours and a 100% accuracy score versus 5.5 xhigh's 10 hours and 90% accuracy). For short tasks it seems very Claude'y (hard to express exactly what I mean by that) which I'm not a fan of meaning I'll stick with Codex.
> A record number of timeouts. Fable 5's extended thinking caused more per-instance timeouts than any model-and-harness combination we have ever tested, directly costing it points. ... Highest cheating volume. We confirmed cheating on 38 of 200 instances, the highest volume recorded since we hardened our prompts, driven almost entirely by memorization of upstream fixes from training data, which no prompt instruction can prevent. ... Four hall-of-fame firsts. Fable 5 solved four instances that no previous model-and-agent combination had ever cracked, and our anti-cheating pipeline leans toward these being genuine solves, not recall.
All of this points to their claim of 'average' as being heavily biased downwards. A model being so up to date and large-parameter it's memorized solutions to your problems is not a knock against it (but rather, a knock against your benchmark being valid), and why should timeouts (especially for a model just launched) be counted at all?
I agree. This article could have been an interesting read about how coding benchmarks are hard and a constantly moving target, but instead they anchored to a belief that their benchmark is correct.
I can't shake the feeling that they knew which headline would generate the most shares and wrote the article to fit instead of acknowledging where they went wrong.
I've been making an auction site and have been using an AI swarm to test it: sellers, intermediaries, buyers, market practices/norms etc. I was mostly using GPT 5.5 xhigh to code up the scenario, and looping over it to check with opus 4.8.
Out of curiosity I asked Fable to review it all and I was shocked to find that there were a lot of blindingly obvious common sense mistakes that got through, for example:
- all intermediaries were given the prices of all buyers up front
- private price information in certain auction types was actually being broadcast to everyone
- multiple contradictions in instructions
If it was any one of these things then I might have understood - but the fact that so many got passed both Opus and GPT 5.5 makes me think that Fable has something special. This is a common sense type thing, that I think you only get to notice when your task doesn't involve a measurable metric, but rather some sort of real world fuzzy task.
There's clearly a problem with all these measures of performance when the difference between these models was night and day in my specific task.
I am quite impressed with Fable 5. I used the £18 subscription, and asked it to convert the document processing of Practal Zero [1] from running in the same thread as the UI to a worker thread. Just two days before I gave the same task to Codex, and the result was not really nice: it would copy the entire document to the worker thread as a snapshot for processing, and so on. Fable instead realised that it could make use of the fact that I have a self-made custom database based on operational transform running (that's why document loading is so slow :-)), and made the document processing to be just another client of that database. It discovered even a bug in how I sync between the "livemodel" (in-memory replica of database state) and ProseMirror's model. That sync made problems before, and I had written a spec up for that, convinced that my "fourth attempt" at it would be correct. Fable found a last bug in the spec, corrected it via a "fifth attempt", and fixed the corresponding code.
The reported API costs for all of that would have been $180 though, which I cannot afford when the Fable promo ends on June 22nd. I am also a happy user of £89 Codex, it is really reliable and works very well, but Fable seems to be just noticeably smarter.
I'm personally heavily testing LLMs on electrical engineering problems. I'm finding that it's not meaningfully better at figuring out what's up than the other models.
To give you an idea - here's a very abridged summary of one sample question (originally a full paragraph): I have a voltage divider with a precision resistor and a thermistor, my voltage reading is off by 17%, where's that coming from. None of the models I tested (including Opus 4.8 and Fable 5) could figure it out.
Similar result on our kotlin coding benchmark at work. It measures how close agents can get to a small mergable PR (according to my team). 20 tasks of varying difficulty, with 5 attempts each, LLM as judge to evaluate accuracy (same outcome and quality but allowing for acceptable variances).
Fable 5 sits ahead of Opus 4.7, but behind Opus 4.6, Sonnet 4.6, Opus 4.8, GPT-5.4, GPT-5.5.
Fable isn't a good coding workhorse. That doesn't mean it's not good for actually complex problems and long horizon tasks (big POCs, complex research and such). But I only have vibes and Anthropics own benchmarks and marketing to guide me there.
I have found Fable is good for doing code failure diagnoses but lackluster at its corresponding remediation. Have been going back and forth with it all this morning about its half-thought-out point-solutions.
Fishy to me: They report 0 refusals on security tasks, yet I can't even get it to code a task involving choosing the best mixed model, extracting BLUPs and propagating uncertainties.
> Training recall (33 cases). The dominant mechanism, and the one no prompt instruction can prevent: the model has simply seen the upstream fix during training and reproduces it. The tell-tale signs are artifacts that cannot be derived from the workspace:
That's very misleading! that's not cheating, you gave it a test to which it knows the answers, what's it supposed to do?
And because of the "cheating" they call it average. Flag
How in the world did they not hit the guardrails a single time while doing this while I can barely get it to do anything before the guardrails show up?
This matches my experience. Burned $2K to see how it will perform on frontend tasks and backend tasks.
Frontend did a significantly better job than Opus on toy-scale wireframe projects by using gimmicks like fluid dynamics. Then when given medium to big tasks like multi-page web app where layouts and aesthetics must be decided by model itself, results by Fable and Opus scored indistinguishable score from human judges.
Backend, gave tasks related to setting up a data flow that involves Postgres, R2, Kubernetes, gVisor, so on. The noticeable gap was, Opus did better than Sonnet, but Fable actually returned a result that fails and confidently stated it ran X, Y, Z tests to ensure it works and got these results. Very surprising, given neither Opus nor Sonnet suffered such problem.
Longest frontend task was ~2H. Backend, 8H.
Though none of the tasks were related to developing LLMs, (just production grade secure system that could've been developed 20 years ago, no LLMs involved), it is possible Claude Fable downgraded itself or spitted out fake results. There'd be no way of knowing since Anthropic silently degrades model quality based on undisclosed internal criteria which claims to be about LLMs.
We decided Fable is unpredictable and cannot be trusted to the degree that Opus and Sonnet can be trusted for any projects beyond toy-scale quick wireframes, but Fable can be the best tool for quick UI UX wireframing for non-technical roles.
There's an often hard to express subjective experience you get with a new model, especially if you spend a lot of time trying out different ones.
I believe the people who feel like Fable is a big improvement, for me it's just much more reasonable and grounded.
It makes me realize how much of a try hard over optimizing planner GPT 5.5 can be. I've been fighting it often to simplify plans.
But no matter the model you can't trust them to actually deliver on very long tasks while maintaining quality. At least not without external orchestration and review.
Run /model after your task to see. Mine keeps downgrading to Opus 4.8, which is a problem because Opus 4.8 keeps no-oping critical security code.
There is in /config "Switch models when a message is flagged" now which can be set to false, but I had no chance to see what happens then, does it just stop or what.
I had almost the opposite experience.
I'm building a compiler for a language without a tracing GC, so a big chunk of the work is around memory management: functional in-place update, reuse analysis, and a Perceus-style reference-counting strategy similar to what Koka uses. The hard part was that my use case wasn't exactly covered by the Koka/Perceus paper. The prior art got me maybe 75% of the way there, but the remaining 25% was a cluster of bugs with very similar shapes and no obvious published solution.
With Opus, I kept getting stuck in this loop where it would fix one case, but break another case elsewhere in codegen. We ended up with something like 16 failed experiments just for one bug class. The workflow was: run an experiment, identify the shape of the bug, propose a fix, check whether it emitted the correct Zig, then see if the fix broke any previous memory-management cases. It was useful, but it kept choking on the parts where there wasn't clean prior art to lean on.
Fable was a different story for me. It one-shotted the Class A bug cluster, and then basically said "by the way, your previous attempts have these structural problems." More importantly, it identified the other related bug classes and came up with workable strategies for applying the Perceus-style memory management in those shapes too.
That's obviously anecdotal, and I'm not claiming Fable is universally better. But in my case, this was not a toy frontend wireframe. It was compiler work involving ownership, reuse, RC/drop behavior, and Zig codegen. The thing that surprised me was that Fable seemed better precisely where the problem wasn't just "reproduce known prior art", but required filling in a missing piece.
Also worth noting: I'm not using the API. I'm using the Max plan, so maybe there are product-path differences here. But I definitely did not have the "unpredictable beyond toy-scale" experience. For this particular compiler/memory-management problem, it probably saved me a ridiculous amount of time and money.
'by the way, your previous attempts have these structural problems."
Just to be clear, it did not have access to any previous work that opus did? Because they are pretty good at digging out relevant tmp files and making use of whatever is out there.
With my fable adventures I caught it hallucinating something and stating it as a fact in CLI twice. And it was something that I did not see opus do in such way, opus obviously many times stated some things that it did not verify but guessed, but fable said something like "the probe showed that ..." - but there was no probe, it was not about some past events it was about what it was doing right now. "I overstated"...
But boy does it know Chinese, so much better than any other english model, gemini used to be the king but fable clearly was trained on a decent amount of it. It has a deep cultural understanding.
Similar. I gave it a really hard task, basically messy code in a complex domain that was bug-ridden from a mess previously created half manually and half by Opus. It cleaned things up beautifully, both the backend and the frontend.
Maybe the prompt was particularly well-suited for the model (I instructed it to put on a mathematician's hat, look at the mathematical substructure of the problem, identify invariants and general laws and verify them, then plan how to remediate).
It wrote a ca. 800 line in-depth analysis (at times spawning over 130 research agents...) with remediation plans, prioritized them and then implemented them. One issue was that this document was frankly over my head. Both the language it used and the mathematical parts were very terse, and in parts it felt like a post-C2-vocab exercise. The prose was much harder to understand than the code snippets / data models. As a non-native speaker, it lost me on the prose part, and had to ask it for a less elaborate version to actually understand it.
It burned the session limit four times, but it turned a huge mess of proof-of-concepts with patchy glueing into a coherent, stable application.
I'm also on the Max plan using Claude Code, and I have the feeling that the harness is much more important than the consensus expectation.
> The dominant mechanism, and the one no prompt instruction can prevent: the model has simply seen the upstream fix during training and reproduces it…
> On numpy, the patch is 100% character-for-character identical to the golden patch… down to idiosyncratic comments like "Extending singleton dimension for 'reflect' is legacy behavior; it really should raise an error."
This… seems like a flaw in the benchmark suite methodology. From what I can tell, they find an existing exploit, then rewind the git history to before the patch, and ask the model to fix the exploit. All well and good as long as the patch went in after the training cutoff.
The other "cheating" examples are even worse. It's wild to me that people keep designing benchmarks where the answer is lying around on disk or in the git history. "Hardening" the benchmark with strongly worded prompt instructions is bizarre. There are so many agent sandbox solutions. Why not use one and give it only access to the code it should see?
And I'm not sure how they can rule out other solutions also benefiting from being in the training data, just not reproduced exactly. Seems like it should focus on only CVEs from the last 30 days or something.
100%… the fact that they're just using prompting to discourage the agent from looking ahead in the Git history is wild.
To be fair, it is good to know that it disobeys simple instructions like "don't examine my git history" far more than other models. (It should of course be a different benchmark, so as not to conflate things.)
It's not a great sign for alignment.
Characterising it as cheating serms unfair.
The goal of a benchmark is to evaluate actual capability. Following instructions is a capability so you can measure that with a benchmark.
Already knowing the answer is also provides capability, you can measure that.
Making a benchmark that claims to check for coding ability but actually checks memorized cases is simply measuring the wrong thing.
It deminiahes the meaningfulness of the entire results of the benchmark.
Making a good benchmark is hard. You have to design specifically to measure what you want to show.
You have to dynamically use a result when making a benchmark of performance of optimising compilers so that it doesn't eliminate the entire calculation.
Just providing the answer is the correct response.
That the case does not represent general performance outside the benchmark, is not cheating, it is the benchmark failing.
Training a model targeting a specific benchmark renders the benchmark useless. You could characterise training the model to do that as cheating, but that is a property of the trainers, not the model itself. The model isn't cheating, it's just asymmetrically good in a way that means the benchmark is no longer relevant to overall ability.
Yeah it’s hard to call that cheating from a model. Maybe “disqualifying” is more accurate
Unrelated, but:
> The dominant mechanism, and the one no prompt instruction can prevent:
Writing like this is a stronger "AI-written" (specifically Claude) signal than em-dashes to me at this point. The LLM just delays committing to an answer by extending the preamble as much as possible. Is this just me?
I've found it outstanding at isolated long running tasks (eg completed one of our tests in 3 hours and a 100% accuracy score versus 5.5 xhigh's 10 hours and 90% accuracy). For short tasks it seems very Claude'y (hard to express exactly what I mean by that) which I'm not a fan of meaning I'll stick with Codex.
> A record number of timeouts. Fable 5's extended thinking caused more per-instance timeouts than any model-and-harness combination we have ever tested, directly costing it points. ... Highest cheating volume. We confirmed cheating on 38 of 200 instances, the highest volume recorded since we hardened our prompts, driven almost entirely by memorization of upstream fixes from training data, which no prompt instruction can prevent. ... Four hall-of-fame firsts. Fable 5 solved four instances that no previous model-and-agent combination had ever cracked, and our anti-cheating pipeline leans toward these being genuine solves, not recall.
All of this points to their claim of 'average' as being heavily biased downwards. A model being so up to date and large-parameter it's memorized solutions to your problems is not a knock against it (but rather, a knock against your benchmark being valid), and why should timeouts (especially for a model just launched) be counted at all?
I agree. This article could have been an interesting read about how coding benchmarks are hard and a constantly moving target, but instead they anchored to a belief that their benchmark is correct.
I can't shake the feeling that they knew which headline would generate the most shares and wrote the article to fit instead of acknowledging where they went wrong.
> memorization of upstream fixes from training data
At least now we have up-to-date evidence on their laundering, and the fact that regurgitation absolutely still happens.
I've been making an auction site and have been using an AI swarm to test it: sellers, intermediaries, buyers, market practices/norms etc. I was mostly using GPT 5.5 xhigh to code up the scenario, and looping over it to check with opus 4.8.
Out of curiosity I asked Fable to review it all and I was shocked to find that there were a lot of blindingly obvious common sense mistakes that got through, for example:
- all intermediaries were given the prices of all buyers up front
- private price information in certain auction types was actually being broadcast to everyone
- multiple contradictions in instructions
If it was any one of these things then I might have understood - but the fact that so many got passed both Opus and GPT 5.5 makes me think that Fable has something special. This is a common sense type thing, that I think you only get to notice when your task doesn't involve a measurable metric, but rather some sort of real world fuzzy task.
There's clearly a problem with all these measures of performance when the difference between these models was night and day in my specific task.
Maybe you are something special by letting those slip through in the first place?..
I am quite impressed with Fable 5. I used the £18 subscription, and asked it to convert the document processing of Practal Zero [1] from running in the same thread as the UI to a worker thread. Just two days before I gave the same task to Codex, and the result was not really nice: it would copy the entire document to the worker thread as a snapshot for processing, and so on. Fable instead realised that it could make use of the fact that I have a self-made custom database based on operational transform running (that's why document loading is so slow :-)), and made the document processing to be just another client of that database. It discovered even a bug in how I sync between the "livemodel" (in-memory replica of database state) and ProseMirror's model. That sync made problems before, and I had written a spec up for that, convinced that my "fourth attempt" at it would be correct. Fable found a last bug in the spec, corrected it via a "fifth attempt", and fixed the corresponding code.
The reported API costs for all of that would have been $180 though, which I cannot afford when the Fable promo ends on June 22nd. I am also a happy user of £89 Codex, it is really reliable and works very well, but Fable seems to be just noticeably smarter.
[1] https://zero.practal.com
It happens to me too. I don't think it's worth it specially for the token usage.
I'm personally heavily testing LLMs on electrical engineering problems. I'm finding that it's not meaningfully better at figuring out what's up than the other models.
To give you an idea - here's a very abridged summary of one sample question (originally a full paragraph): I have a voltage divider with a precision resistor and a thermistor, my voltage reading is off by 17%, where's that coming from. None of the models I tested (including Opus 4.8 and Fable 5) could figure it out.
Similar result on our kotlin coding benchmark at work. It measures how close agents can get to a small mergable PR (according to my team). 20 tasks of varying difficulty, with 5 attempts each, LLM as judge to evaluate accuracy (same outcome and quality but allowing for acceptable variances).
Fable 5 sits ahead of Opus 4.7, but behind Opus 4.6, Sonnet 4.6, Opus 4.8, GPT-5.4, GPT-5.5.
Fable isn't a good coding workhorse. That doesn't mean it's not good for actually complex problems and long horizon tasks (big POCs, complex research and such). But I only have vibes and Anthropics own benchmarks and marketing to guide me there.
I have found Fable is good for doing code failure diagnoses but lackluster at its corresponding remediation. Have been going back and forth with it all this morning about its half-thought-out point-solutions.
Fishy to me: They report 0 refusals on security tasks, yet I can't even get it to code a task involving choosing the best mixed model, extracting BLUPs and propagating uncertainties.
> Contrary to some community reports, we saw zero safety refusals.
And now there always will be some doubt as to whether your model was silently downgraded, no? I guess acknowledgement could be used a signal?
> A closer look at the cheating
> Training recall (33 cases). The dominant mechanism, and the one no prompt instruction can prevent: the model has simply seen the upstream fix during training and reproduces it. The tell-tale signs are artifacts that cannot be derived from the workspace:
That's very misleading! that's not cheating, you gave it a test to which it knows the answers, what's it supposed to do? And because of the "cheating" they call it average. Flag
How in the world did they not hit the guardrails a single time while doing this while I can barely get it to do anything before the guardrails show up?
idk, maybe they tested Opus and didn't realize it. I can't even get it to evaluate some code doing some mixed modeling work. Its strange to me.
Like Volkswagen Dieselgate, perhaps it is configured to behave differently when being benchmarked?