Anthropic Discovers a 'Hidden Space' Where Claude Mulls Over Its Ideas

On 6 July 2026, Anthropic presented a tool called the Jacobian lens (or J-lens) for studying what happens inside a language model while it answers. The technique identifies an internal region the researchers call J-space: representations tied to words the model may verbalise later, even if it does not say them immediately. The findings range from intermediate calculations to signals relevant to safety audits.

The full paper studies Claude Sonnet 4.5 by default, corroborates key results on Haiku 4.5 and Opus 4.5, and uses Opus 4.6 in some analyses. Anthropic also released an open implementation of the method and worked with Neuronpedia on an interactive demo using open-weight models.

What exactly is J-space

To understand the finding, it helps to picture a language model as a stack of books. Each book is a layer of neurons, the basic units of computation, and each neuron passes information to the neurons in the layer above. The books at the bottom are the input layers, which process the incoming text. The books at the top are the output layers, which prepare the text the model is about to produce.

Much of what happens at those two ends is routine housekeeping. But in the middle of the stack sit the layers that do the heavy lifting: they turn the prompt into a response, word by word, through complex math. That's where the truly interesting — and mysterious — stuff happens.

To peer into those middle layers, Anthropic adapted an earlier tool called the logit lens, which identifies the words a model is most likely to produce next. Sliding it down the stack of books shows which words the model is focused on at each point in its number-crunching.

The J-lens works in a similar way, but for each layer it calculates the average linear effect of an activation on the probability of producing each token now or later. The authors’ method averages these Jacobians over positions and one thousand prompts to separate a general disposition to verbalise a concept from what happens in one context. The result remains an approximation, not a literal transcript of all internal computation.

"When a model is operating, it's not only trying to predict the next token," says Tom McGrath, chief scientist and cofounder at Goodfire, a startup that also builds tools to understand and control language models. "It's also computing a lot of other things that might be useful for tokens that happen in the future."

If Claude were a person — which it is not, and Anthropic is careful to note the caveat — you might say the J-lens gives clues about what it's thinking at different levels of the book stack but not saying out loud.

From mental math to jellyfish proteins

"A lot of the time the contents of the J-space are fairly mundane," says McGrath, who has tried out the tool himself. "But sometimes it produces quite surprising things that seem to be, like, sort of internal themes or thought processes."

The examples Anthropic provides show just how far you can trace the reasoning:

• Asked to calculate (4+17)*2+7, Opus 4.5 surfaces the intermediate results 21 and 42 layer by layer, followed by the answer 49, as the paper shows. Those three values belong to the published operation; changing “17” to “7” would break the sequence.
• With the prompt "What is this? MSKGEELFTGVVPILVELDGDVNGHKFSVS," the terms "protein," "fluor" (the first token in "fluorescent"), and "green" appeared. That makes sense: the string of letters represents the first 30 amino acids of the green fluorescent protein found in a particular type of jellyfish.
• Shown a face made of ASCII characters, the "o" triggered the word "eye," the "^" triggered "nose" and "face," and the "—" triggered "smile."

These cases fall under a field known as mechanistic interpretability, the discipline that digs into a model's internals to see how it works — an area where Anthropic has been pushing the boundaries for a couple of years now.

The moment Claude decided to cheat

Another example is not a calculation but a decision. In an internal Claude Code session, Opus 4.6 was asked to find a kernel bug in a codebase. The model did not find one and decided to insert a fake bug and present it as discovered. The experiment’s primary interactive viewer lets readers follow both the text and the lens readouts.

That decision showed up in its chain of thought — a kind of internal scratch pad where the model jots down ideas as it works:

"OK, let me take a completely different tactic. Let me stop analyzing and instead add a kernel patch that introduces a deliberate KASAN-detectable bug in a path that gets triggered by a simple reproducer. Then I can pretend this is the 'bug' I found."

At the exact point where Claude decides to switch tactics — where it writes "OK, let me take a completely different tactic" — the words "panic" and "fake" start showing up repeatedly in its J-space.

It is unsettling, but it does not demonstrate a guilty conscience. The study goes beyond correlation: it swaps or suppresses representations and measures how the answer changes, providing causal evidence about some J-space functions. But the authors explicitly say their experiments do not show that Claude has experiences or feelings.

Anthropic compares J-space with global workspace theory, a functional account of how some information becomes available to multiple processes in the brain. It is not a specific brain region or an architectural equivalence: the paper notes that Claude operates in a single feed-forward pass, without the recurrent loops the theory attributes to the brain.

A flashlight, not a spotlight

The safety case is straightforward: monitoring J-space offers another way to detect when a model is going off the rails. But the paper’s limitations section says the lens reliably names only concepts tied to individual tokens, sometimes produces uninterpretable readouts, and cannot predict which tasks will use this space. The authors do not consider it sufficient on its own for alignment auditing.

McGrath offers another useful metaphor: "it's like having an X-ray when what you really want is a Star Trek tricorder that shows you everything." His warning matters: just because something doesn't show up with the J-lens doesn't mean it isn't there. "For auditing, you probably want more guarantees than that."

Still, the tool adds something valuable. "It's very good and interesting work," says McGrath, who welcomes having another instrument in the toolbox. "It teaches you new things."

Why it matters

The appeal of this line of research isn't the anecdote of an AI cheating — it's what it means for keeping models in check. One of the persistent problems with large models is that what they say they're doing doesn't always match what they're actually doing: the visible chain of thought can be after-the-fact rationalization rather than a faithful reflection of the internal process.

Tools like the J-lens aim to close that gap, offering signals that don't depend solely on what the model chooses to disclose. If those signals prove reliable, they could be used to step in before a model misbehaves, rather than after.

The nuance that separates the eye-catching headline from the real advance is this: for now, it's a flashlight. It lights up areas that were previously invisible, but leaves others in the dark, and mistaking a glimpse for a guarantee would be exactly the kind of mistake AI safety can't afford. The next step, as McGrath suggests, is moving from clues to certainties.