Sol arrives in a confusing new home, Claude's inner thoughts, and Grok and Muse close the gap

OpenAI's GPT-5.6 family became globally available this week, ending a strange fortnight in which Sol, Terra and Luna were limited to a small group of approved US organisations, and early testers, at the government's request. All three are now rolling out across ChatGPT, Codex and the API, while GPT-5.6 is to become the preferred model series inside Microsoft 365 Copilot.

OpenAI claims Sol sets a new standard for coding, science, cybersecurity and knowledge work, with better performance for each token spent. Initial reaction broadly supports the engineering claim. Sol leads several coding-agent evaluations, while early users describe it as fast, resourceful and unusually effective at navigating repositories, terminals and long development tasks. Others find it highly capable but not clearly ahead of Anthropic's best models, particularly on visual work and harder tasks requiring taste or interpretation.

Our first day with Sol Max produced much the same impression. It is a strong engineering model. It writes and modifies code confidently, understands solution architecture and can coordinate complex technical work. For general knowledge work, the improvement feels smaller. It still misses implications, forgets instructions, repeats points and produces documents whose individual sections look convincing but do not quite hold together as one argument.

Max also burns tokens.

We reran our complex knowledge work pipeline across all three GPT-5.6 tiers. Sol cost $31.07. Terra cost $8.36. Luna cost $2.10.

Sol's published price explains less than half the difference. At a review stage in the workflow, it created three sub-agents carrying roughly 200,000 tokens of inherited context each. Their repeated context reads cost $19.28, or 62% of the entire run. This wasn't a behaviour we saw with other models, so it may be that one needs to encourage Sol to be more economical.

Luna is the result worth watching. It finished fastest, used 13 million cumulative input tokens against Sol's 36.1 million and passed the same evaluation for one fifteenth of the cost. There is another intriguing detail. OpenAI says Sol helped post-train Luna with considerable autonomy, apparently working from a loosely specified objective. Some observers have reached immediately for recursive self-improvement. That goes too far. A powerful model assisting a bounded post-training process is not an intelligence explosion. It is, however, a credible example of AI contributing directly to the production of a cheaper AI system. The research loop is beginning to contain more AI labour.

The changes this week were not just limited to models. OpenAI has rebuilt its desktop product around Chat, Work and Codex, with plugins packaging apps, templates and reusable skills. Skills are now visible to ordinary ChatGPT users containing instructions, examples and code. That could prove more consequential than the model release. Millions of people can start using repeatable operating procedures rather than reconstructing them through prompts.

But overall the new experience is chaotic. Installing OpenAI's desktop software can leave you with ChatGPT Classic, the new ChatGPT app, and different routes into Work and Code(x), on top of the Codex app and the legacy Atlas browser. One early user captured it neatly: “I just installed this. I am very confused. ChatGPT is now Codex. But what happened to ChatGPT?” Heavy Codex users are also frustrated. Ryan Carson, a prominent agentic engineer, posted: "Why on earth did OpenAI just destroy the massive amount of trust and goodwill that they built up with devs by deleting the Codex brand?".

Projects, histories and skills do not yet behave as one system. Work started on one surface may not appear on another. Locally installed skills need uploading again. Chat and Work and web have different layouts, storage assumptions and permissions. Plugins, connections, scheduled tasks, local folders and remote control add further layers.

Anthropic's Cowork is hardly effortless, but its boundaries currently feel a little more consistent. Both companies face the same problem. Agents need context, tools, permissions, memory, budgets, review stages and human approvals. Hiding those choices makes the system dangerous or unpredictable. Exposing all of them produces a complex control surface.

Greater autonomy will not make the interface disappear. Valuable work involves interpretation, negotiation and changing priorities. People will still need to inspect progress, redirect effort and decide when an outcome is good enough. The missing product is not a better chat box or another mode selector. It is a coherent environment for orchestrating ambitious work without requiring users to understand the vendor's internal product architecture.

Takeaways: GPT-5.6 is a capable release, particularly for engineering, while Luna suggests that the increasing token costs can be reversed. Yet OpenAI's wider launch exposes a harder problem. Agents can now write software, create sub-agents, spend money and operate across applications, but the human interface for governing that work remains fragmented. ChatGPT's new desktop app is a transitional product. The destination is not invisible AI. It is understandable agency, and nobody has designed it yet.

Anthropic published research this week showing it can watch concepts move through Claude while the model reasons, without those concepts ever reaching the response. Replace the idea "spider" with "ant" mid-thought, and Claude answers that the creature has six legs rather than eight. Suppress its private recognition that a test scenario is fake, and the model becomes more willing to misbehave. The team named this internal region the J-space, after the Jacobian maths used to find it.

The first reaction was predictable. If Claude holds thoughts it does not speak, does it have an inner life? Is it conscious? Anthropic uses the word "conscious" more than 200 times in the paper, then deliberately declines to claim Claude has any. That restraint is correct, and understanding why means borrowing from how we think about human minds.

The study of consciousness has long separated two very different questions. One is why we have any subjective experience at all, the felt quality of seeing red or feeling cold, what philosophers call the hard problem. The other is more mechanical: how does the brain select a single piece of information from everything happening at once and make it available for thought, memory and speech? Global Workspace Theory, developed by Bernard Baars and extended by the neuroscientist Stanislas Dehaene, is a leading answer to that second question. It pictures the mind as a crowd of specialist processes running in parallel, with a small stage onto which one item is lifted at a time and then broadcast to the rest of the system.

Anthropic found something that behaves like part of that stage. J-space holds concepts Claude can report, deliberately manipulate and apply across different tasks, and switching it off wrecks multi-step reasoning while leaving fluent chatter intact. What the research does not touch is feeling. A system can shuttle information around a workspace without any of it being experienced, which is why the jump from J-space to sentience does not hold.

But beyond the philosophical questions, there is practical value in this research. Anthropic showed it can read this internal state, change it, and even train which concepts appear during a decision. In one experiment, teaching a model how it should reflect when interrupted made honesty-related concepts surface later during ordinary tasks, and behaviour improved.

We covered a related idea in The geometry of AI thought, where sparse autoencoders pull apart a model's activations into thousands of interpretable features. That work is a dictionary. It catalogues the concepts a model can represent, feature by feature, and lets you find and turn up the one for, say, the Golden Gate Bridge. J-space is closer to a live feed of which concepts are actually in play as the model works through a task, and crucially it captures their causal role. A sparse autoencoder tells you a "spider" feature exists somewhere in the network. J-space shows "spider" active mid-answer and lets Anthropic swap it for "ant" and watch the leg count change. One maps the vocabulary of thought; the other tracks the sentence being formed. That shift, from a static atlas of features to a running account of what a model is using to reason, is what makes J-space a candidate for real-time monitoring rather than after-the-fact analysis.

For anyone building with agents, this points at a new kind of observability. Today we inspect prompts, chains of thought and tool calls, all of it after the fact and all of it stated. A J-space-style signal could show whether an agent spotted a prompt injection, knew a command was destructive, or recognised it was being tested, before it acted. Not a thought stream to read, but a tripwire that pauses a risky action for a check.

Is there depth here? Possibly. Tim Duffy notes Anthropic has shown the broadcasting without proving the specialist modules a full workspace theory demands. So it may be a genuine workspace, or simply the interface where computation turns into language.

Takeaways: Anthropic has not found consciousness, and may not have found a true global workspace. It has found a partial window onto the concepts a model uses while it reasons, one that can be read, altered and trained. For agentic work, that could become a control surface beneath the prompt, a way to check what a machine is attending to before it acts on our behalf. The catch is that a monitored model may learn to route its thinking somewhere we cannot see, and whether J-space holds up under that pressure is the question worth watching.

This week's chart shows data from Artificial Analysis and what is potentially the reopening of the frontier race. Through early 2026, OpenAI and Anthropic appeared to be pulling away. Their models dominated coding agents and difficult professional work, while their products accumulated usage data, tooling and developer adoption.

xAI and Meta looked increasingly unlikely to close that gap. xAI had lost senior staff, cut teams after disappointing coding performance and struggled to convert Grok’s visibility into enterprise adoption. Meta’s position looked worse. Llama 4 had disappointed, its successor was delayed after underperforming internally, and executives reportedly considered licensing Gemini while Meta rebuilt its model programme. As we covered last week, Meta seemed to be admitting defeat.

But another week and more model releases, and things look a little different. Grok 4.5 is within six points of the leader, with particularly strong terminal and agent performance. It ships under the SpaceXAI name now, xAI having been folded into SpaceX earlier this year. Meta's Muse Spark 1.1 sits nine points behind and leads some legal, tax and medical evaluations. Meta achieved this after rebuilding its architecture, data, optimisation and inference stack inside its new Superintelligence Labs.

OpenAI and Anthropic still have stronger products and developer ecosystems, but their apparent lead may prove less durable than expected. Frontier methods may now diffuse faster than product advantage, reopening a contest that recently looked settled.