0:07
GPT-5.6 Sol nearly matches Fable 5 on aggregated benchmarks at one-third the cost
OpenAI's GPT-5.6 Sol scores 59 points on the Artificial Analysis Intelligence Index, just one point behind Claude Fable 5. At $1.04 per task, it costs a third of what Anthropic's top model charges. In agentic coding, Sol beats every competitor, adding even more pricing pressure on Anthropic.
The article GPT-5.6 Sol nearly matches Fable 5 on aggregated benchmarks at one-third the cost appeared first on The Decoder.
0:32
Anthropic found a hidden space where Claude puzzles over concepts
The AI firm Anthropic has developed a technique that has given it the clearest glimpse yet at what’s really going on inside large language models as they answer questions or carry out tasks. What they found ranges from the mundane to the unnerving.
Researchers at the company built a tool called the Jacobian lens (or J-lens) and used it to uncover a hidden area, which they named the J-space, inside Claude Opus 4.6, a version of Anthropic’s flagship LLM released in February.
The J-space contains individual words that are related to the words and phrases that the model is most likely to spit out in a response in the near future. If Claude were a person (which it is not), you might say that these hidden words can reveal what’s on its mind before it actually speaks.
Anthropic found that what an LLM is actually doing can often be different from what it says it is doing. The company claims that monitoring words that pop up in the J-space gives it a new way to understand and control its models.
The company shared its results in a paper posted on its website this week. It has also teamed up with Neuronpedia, an open-source platform that lets you poke around inside LLMs yourself, to make a hands-on demo that anyone can try.
“It’s very good and interesting work,” says Tom McGrath, chief scientist and cofounder at Goodfire, a startup that also builds tools to understand and control LLMs.
For the last couple of years, Anthropic has been pushing the envelope in a field of research known as mechanistic interpretability, which involves probing the internal workings of LLMs to see how they tick. (MIT Technology Review picked mechanistic interpretability as one of this year’s top breakthrough technologies.) The new technique builds on previous work from Anthropic and others to expose a deeper level inside LLMs that researchers had not seen before.
Picture an LLM as a stack of books. Each book is a layer of basic computational units known as neurons, with each neuron in one layer passing information to the neurons in the layers above. The books at the bottom of the stack are the input layers, which process the text coming into the model. The books at the top are the output layers, which prepare the text that the model is about to produce. Much of what goes on in these input and output layers is housekeeping.
But in the middle of the stack, you get the layers that do the heavy lifting, churning through the complex math that turns prompts into responses one word at a time. That’s where the really clever—and mysterious—stuff happens.
To peer deeper into those middle layers, Anthropic adapted an existing tool called a logit lens. A logit lens can be used to look inside an LLM to identify the words that it is likely to produce next. Moving the lens down the stack of books reveals what words the LLM is focusing on at that particular point in its number crunching.
Anthropic’s J-lens works in a similar way but picks out words that an LLM is likely to say at some point in the near future, not necessarily straight away. What that reveals in practice are words that are related to the response an LLM is working on but that might not actually end up being part of that response by the time the math in the middle layers has run its course.
“When a model is operating, it’s not only trying to predict the next token,” says McGrath. “It’s also computing a lot of other things that might be useful for tokens that happen in the future.”
Again, if Claude were a person (it’s not), you might say that the J-lens gives clues about what it is thinking about at different levels of the book stack but not saying out loud.
“A lot of the time the contents of the J-space are fairly mundane,” says McGrath, who has tried out Anthropic’s J-lens himself. “But sometimes it produces quite surprising things that seem to be, like, sort of internal themes or thought processes.”
Anthropic gives a number of examples of what it found. Sometimes the J-lens exposed the steps that Claude took when it was working through a problem. For example, when it was asked to calculate (4+7)*2+7, its J-space contained the word “math” and numbers representing the intermediate results “21” (for 4+7) and “42” (for 21*2).
In other cases, the J-lens revealed how Claude recognized different inputs. For example, the prompt “What is this? MSKGEELFTGVVPILVELDGDVNGHKFSVS” triggered the words “protein,” “fluor” (the first token in the word “fluorescent”), and “green.” (Which makes sense: the string of letters represents the first 30 amino acids in the green fluorescent protein found in a particular type of jellyfish.)
And when Claude was shown an ASCII face—
—the “o” triggered the word “eye,” the “^” triggered the words “nose” and ”face,” and the “—” triggered the word “smile.”
Anthropic also found that the J-space can sometimes give remarkable insights into an LLM’s decision-making. In one striking example, researchers testing Claude Opus 4.6 asked the model to find a bug in a large code base.
4:28
Instagram users: Here’s how to stop Meta’s AI from using your photos
Muse Image allows users to generate AI images using photos from public Instagram accounts. As long as a person's profile is public, another user can tag that account and use their images as part of an AI-generated creation.
4:44
Google will now disclose which ads are made with AI
While Google prohibits misleading and deceptive ads, an ad can still leverage AI to create some type of synthetic or digitally altered content. Until now, that's something Google only required election ads to disclose.
4:58
Meta enters the crowded AI coding battle with Muse Spark 1.1
I just can't stop thinking about this, bro. So Meta's been quietly working on this AI coding tool called Muse Spark 1.1, and it's basically a powerhouse for handling massive coding projects. They're talking about large agentic workloads, which is just a fancy way of saying it can take on a ton of complex tasks and still keep up. It's also got the ability to fix bugs and help with these huge code migrations that companies are always struggling with. The automation potential here is huge, and I think it's a big deal that Meta's jumping into this space.
5:29
Fidji Simo steps down from OpenAI’s no. 2 role
So, I was digging into this news about Fidji Simo stepping down from her role at OpenAI, and what caught my attention is that her medical leave ended up being a lot longer than expected. That's already a pretty significant development, but what's even more interesting is that this is happening right when OpenAI is trying to make some big moves - they're considering an IPO and they're in a heated competition with Anthropic to dominate the enterprise market.
It's worth noting that Fidji Simo was in the No. 2 spot at OpenAI, so her departure is definitely going to leave a leadership vacuum. The question is, who's going to fill that gap and how will it impact the company's trajectory. I'm curious to see how this plays out, especially given the timing.
The IPO plans and the competition with Anthropic are already high-stakes, and now you've got a key executive leaving the company. It's like a domino effect - every move OpenAI makes now is going to be under a microscope. I'm guessing this is going to be a pretty interesting few months for the company.
6:22
Ways to think about token pricing
There are only two things you can say with certainty about token prices: we’re in a supply crunch, and this is unstable. All of the variables are in play, and the market will get shaken out over the next few years to arrive at a new equilibrium. Right now we have a lot of frantic analysis of ‘time to power’, but the question at the end of that remains whether the foundation models have sustainable pricing power, strategic leverage and value capture, or whether they become low-margin commodity infrastructure providers. At the moment, I think every dynamic we can see points to the latter.
Clearly, the situation today is transitory. On the supply side, a trillion dollars or more of data centre capex is coming down the pipe (and plenty more semiconductor capex behind that), inference efficiency continues to improve very quickly, and new models are far more (or far less!) efficient in their token use. On the demand side, although the market has been capacity-constrained since 2022, the crunch in the first half of this year has been driven by sudden product-market fit in really just one use case, software development, and that’s actually a pretty small field (imagine if we had product-market fit for a consumer use case with hundreds of millions of DAUs - today’s infrastructure couldn’t support it at any price). We don’t know what the next use-cases to scale will be, nor when that would be, nor what their token needs would be.
Going up a level, it’s been pretty widely reported that inference today has 40-50% gross margins: this includes deprecation of the associated server costs (or the cost of renting them), but we don’t really know the asset life (five years? Seven years?) and obviously this doesn’t include the cost of training the next model a couple of times a year, which is currently far larger than revenue. In principle, inference is a marginal cost and training is a fixed cost, so with high enough revenue you can reach profitability, but we don’t know how training costs will change. On the other side of the table, it’s unclear how much of the surge in use in the last few months has an ROI (or at least has an ROI that can be quantified to a CFO), let alone any future use cases, and hence what prices people might be prepared to pay for them.
So, all the variables will move all over the place over the next 12 months, and move again over the next three to five years. How could we suggest where this will settle? How and where will supply, demand, price, capacity and capex get back into equilibrium?
In theory, you can model this bottom-up. You can make some assumptions around each of the variables I suggested above, and then try to model out how many chips there are now, how many chips with what performance TSMC and the rest of the semis industry might be able to deliver when, and how fast all of that can be brought online in data centers and how fast those can be powered. Then you can wonder about price discipline, and make some guesses about use cases. This will get you a number, but it will be rather like trying to build a five-year forecast for the broadband market in 1998: the spreadsheet will be very pretty, and you might even get close to the right number for this year, but there are too many unknown variables to make a useful forecast of a longer-term market structure.
In other words, we can say that token price is a function of supply and demand at a level between the sellers’ marginal cost and the buyers’ ROI, but we don’t actually know what supply, demand, marginal cost or ROI will be.
The other approach is to look from the top down: how do things like this tend to play out? What are the building blocks, and where can they go? Most of this conversation depends on what happens to this curve.
First, how many people will pay to be at the top right of the curve - to be at the frontier? At one extreme there are already use cases that already work just fine with a small, old, perhaps open source model that runs for ‘free’ on-prem or on your phone; at the other extreme there will be some that get better results from the latest, most expensive frontier model, consuming lots of tokens for lots of money; and then there will be many that are somewhere in between.
