Redeploying Fable 5
anthropic.com
anthropic.com
> The export control directive on June 12 came after the government became aware of a report in which Amazon researchers had found a method of bypassing Fable 5’s safeguards: prompting it so that it identified a number of software vulnerabilities. In one case, the model produced code demonstrating how the relevant vulnerability could be exploited. Over the past two weeks, we have worked closely with the government and other partners, including Amazon, to review the report and evidence.
> Our testing confirmed that many less capable models—including Claude Opus 4.8, GPT-5.5, and Kimi K2.7—could identify the same vulnerabilities as Fable 5 did in the report. When it came to the demonstration of how to exploit the single vulnerability, every model we tested could produce the same demonstration as Fable 5 (including Claude Haiku 4.5, Sonnet 4.6, Opus 4.6, Opus 4.7, Opus 4.8, GPT-5.4, GPT-5.5, and Kimi K2.7).
This indicates three things:
1) WTF was Amazon thinking? Didn't their researches try the same thing in other models too before telling the CEO to tell the government it was dangerous (!?)
2) Anthropic - in particular Dario - really needs to learn government relations better. Most of the problems Anthropic has had with the government seem to stem from Dario's attitude rather than actual facts. (Eg, the DoD debacle seems to have ended up with OpenAI signing almost the same contract Anthropic already had, just worded differently)
3) The administration decision making is just wacky. In a normal administration they'd have actual policy documents you could look at to understand under what circumstances they think models have a problem. With this they just seem to make it up as they go, and the tools they use make no sense at all. If it is dangerous for cyber security reasons why would export controls make sense to use?
This may be in Anthropic's benefit. Autonomous killer drones will be widely deployed by the US military, and at that time, Anthropic can point to them and say "see, we were trying to avoid that"
All that matters is the killer drones use someone else's AI first.
Tautological
LLMs only really took off almost entirely under this administration. No one knows quite how to regulate these things. It's not surprising that they would flail a bit. It seems to have been DoD people who did the flailing here, and apart from Hegseth, there might have been others who did the overreacting, possibly before him.
The party will be short-lived.
And yes, after the IPO.
I think the 5x subscription is here to stay - I'd bet they make money on that from lots of people not using it.
The 20x is already unavailable in Teams plans.
GLM 5.2 comes close on agentic tasks, but doesn't code as well.
Kimi 2.6 and Deepseek v4 Pro write great code but lose track when doing agentic workflows. They were better than Sonnet 4.6 but not as good as Opus. I haven't compared them to Sonnet 5 yet.
> Kimi 2.6 and Deepseek v4 Pro write great code but lose track when doing agentic workflows.
Here is an idea: can GLM 5.2 and Kimi 2.7 be combined in some way? Maybe GLM doing planning and Kimi doing coding.
This isn't said out of naiveté or the idea that companies won't cheap out, but at some point – if access to models for defense is broadly available enough – we have to take a step back and say, "Aggressively insure your code against attack with AI on your side, because after <date> the other side will just _have_ AI."
I feel like something lost in a lot of the discussion around mythos and fable is that computer security absolutely has a substantial defender's advantage. It is indeed possible to ship e.g. surfaces that would be super resilient to attack (e.g. no unnecessary open surfaces, etc.) modulo category-shift attacks like RowHammer, etc.
Besides, just making sure that more people in the world actually have access to non-lobotomized models, this is _necessary_ if open source is (hopefully) going to continue to progress and if jailbreaks aren't totally vanquished.
Any exact times when we can expect it? The beginning of workday Pacific time, maybe?
Is it just select users, or does es nobody have access?
I have the sense that this limited usage model will persist, Opus at its current token price will gradually be phased out, and eventually the cost of these models will be unsustainable and impractical.
I’d rather not play this game anymore and seek out more reliable providers which may provide less capable models but aren’t going to gradually kneecap my plan or rug-pull me.
Frankly I’m much less concerned with sheer model capability these days and more interested in harnesses. Claude Code is not the best on the market, and I think there’s a lot of exploring and learning to do in that arena.
The whole Fable debacle has dramatically changed my perspective on the market and what I want from it/who I should support.
This is generally true, but there are some providers offering inference at relatively stable prices. They aren’t Opus-tier models, but some appear to be at or close to Sonnet 4.5 or so. For much of the work I do, this is fine.
Essentially if you aren’t at the frontier, you can find cheaper tokens that aren’t about to be rug-pulled or decommissioned on a whim.
for now. Once they get enough customers, why will they not jack up prices is my point.
All businesses will act in that manner because it serves their interests. That is not going to change if you're OpenAI or ChineseAI or BelgiumAI.
Again, this is different from the export control issue because it's a fundamental business thing.
You will always want more money from customers not less and there is no noble business that exists just to make your tokens cheap while they give up potential profits unless they have something to gain from tokens being cheap.
They will gradually up their prices as well and squeeze as much as they can like most first-gen AI providers are gradually doing since they got a lot of people hooked already.
It’s so expensive, though. Useful interactively in coding agent but boy if you have a kind of task that Fable can do in an agentic loop that Opus can’t then you’re in a good place. Doubtless the frontier will move forward. These kinds of “problems of the future” are great to have solutions for.
And for an interactive agentic loop, not using Fable is just missing out on something. There’s no lock-in there.
What is the best harness right now?
It bothers me that the Claude Code leak (from what I read, didn't look) a few months ago showed a lot of careful consideration for context handling. I assume this means that each harness performance/correctness would differ drastically just from this alone.
All I can say with certainty is that Claude Code isn’t a panacea, and it’s quite buggy and clunky. I think this is evident just from regular usage, but becomes even clearer when you use harnesses that get in the way less and demonstrate better implementations across the board. Sometimes I think Claude Code feels like software that happened while other harnesses feel like they were engineered intentionally.
I bet it'll continue to be messy at the frontier for the foreseeable future as society gradually wakes up to the consequences of strong AI.
That is nice advertising for Kimi, huh?
When it's re-enabled the fact the quotas are halved supports that theory.
disappointed but not surprised