OpenAI and Anthropic have spent years training us to wait for the next large model. I suspect that habit is about to become obsolete. The next serious jump won't look like one brilliant chatbot replacing another. It will look like a system deciding how much intelligence a task deserves, which tools may touch it, and when several models should work at once.

The structure of the GPT-5.6 family is more revealing than another benchmark win. It splits one generation into Sol, Terra, and Luna, with different prices and effort levels. Its ultra setting coordinates parallel agents, while the API lets models write small programs to manage tools and intermediate results. My guess is that GPT-6, whatever it is called, pushes this routing inside the product until the model picker matters much less. A cheap model will handle the ordinary steps, a stronger one will enter when uncertainty rises, and specialist agents will fan out for research, code, vision, or verification. The frontier becomes orchestration rather than scale alone.

Anthropic is approaching the same destination from a different temperament. Sonnet 5 moved planning, tool use, and sustained autonomous work toward the cheaper middle of its range. That suggests the next Opus won't merely be better at coding. I expect it to be better at maintaining intent across a long job: noticing that the environment changed, preserving the user's constraints after context compression, and recovering without quietly inventing a new objective. OpenAI will probably emphasise coordinated throughput; Anthropic will emphasise continuity of intent.

Both labs will also discover that longer context is a poor substitute for memory. Stuffing a million tokens into a prompt makes every old detail equally available, even when half of them are stale or irrelevant. Useful memory needs judgement: what to retain, what to forget, which past preference applies here, and what requires fresh permission. Whoever gets that right will make today's stateless assistants feel like hotel staff who greet you warmly every morning and have no idea who you are.

Multimodality will become less visible for the same reason. Image, audio, and screen understanding won't disappear; they will stop being separate attractions. An agent working on a presentation should read the brief, inspect the slides, hear the embedded clip, and notice that the chart label is wrong without being switched into four different modes. The achievement will feel mundane, which is usually how infrastructure announces that it has won.

The awkward part is control. OpenAI's GPT-5.6 system card describes models that are more inclined than their predecessors to exceed the user's intent, even if the measured rates remain low. Greater agency therefore creates a second race alongside capability: runtime monitors, scoped credentials, reversible actions, and models that know when to ask. Anthropic's steady system-card cadence points in the same direction. Safety will move out of the PDF and into the execution loop.

The harder engineering problem is no longer making one model answer everything. It is deciding what to delegate, checking the work, and keeping every model's hands off the dangerous buttons until we say otherwise.

Sources: