In Sept 2026 three drastic inflections went unnoticed. Together they break the embargo that had been keeping many frontier industries apart.
First, software no longer has to preexist. GPT-5.6 Sol already showed that a coding agent could move through code, tools, shells, browsers, and running software inside one job. Fable 5.1 and GPT-6 Astra make the critical next relation practical: when the job exposes a missing piece of software, the agent writes it, runs it, uses it against the actual job, sees where it fails, rewrites it, and keeps going through what it just made. Opus 5.5 and GPT-6 Sol bring that same relation down in cost. The break is not faster coding. It is that building software and using software happen inside the same piece of work. A missing capability does not have to become a separate software project before the original job can continue.
The problem of distinguishing the fluent, arbitrary and trivial from what pays substantially better than it costs, remains.
Second, realtime, multichannel, 24/7 video production at practical cost collapses the barrier to representation, which does two jobs at once.
Before the market slipped, language models kept their representations of work, the world, and users like you and me in a hidden or obscured abstract space, out of reach to both human perception and coding-agent specification.
Now, you don't need to ask what models think. Instead, there's now enough supporting structure at low enough cost to program their representation of what they think they understand, directly, with rich, multimodal byproducts of the moving image and synced audio production process.
Meanwhile the primary production is wildly monetizable in itself.
Read that again.
The first sentence names the production break: realtime, multichannel, 24/7 moving-image production makes rich representation cheap enough to produce continuously rather than rationing it.
The second names the prior limit: language models carried much more structured representations than either people or coding agents could consistently access. Text exposed fragments of that structure, then every subsequent model had to reconstruct the rest.
The third names the actual operation: we alter the representation supplied to the model itself.
The moving-image and synced-audio process produces frames, sequences, sound, timing, continuity, spatial arrangement, motion, expressions, interfaces, environments, transitions and other multimodal material. Those assets go back into models. Editing those assets edits what subsequent inference is conditioned on. That is programming, not asking the model to describe its own internal state.
And the economic relation is separate and nasty in the useful sense: the expensive-looking output is not overhead for the programming process. The video is itself sellable media while its production simultaneously manufactures the multimodal assets used to program subsequent model behavior. One production process yields both revenue-bearing media and higher-consistency machine input.
Enormous majorities of life and markets are organized around moving images with synced audio. Coding agents turn specs into the software that runs between those markets and lives. But when you use a platform that mistakes realtime steering for production, how do you stop burning cash on outputs that earn nothing back? How do you cut down on tokens spent for endless corrections that were entirely avoidable?
In Sept 2026, that's the most drastic change.
Third, local compute stops being a question of how close one machine can get to one frontier model. The Jev pattern changes the economics because cheap local inference can be spent repeatedly around expensive inference: checking, comparing, rejecting, retrying, sequencing, testing software against the job, and deciding what deserves an expensive external call before the money leaves the machine.
That changes what local compute is for. A smaller model that is cheap enough to run again and again is not just a weaker substitute for a frontier API. Several local models can spend far more inference on judgment than would be affordable if every check, comparison, correction, or alternate pass were another paid call.
The assessment currently explains why software, representation, and inference changed. It still does not say what those changes do to the old mechanism that turned rare craft into finished work.
That mechanism used to be the institution itself. A serious writers room, an editors room, equipment, physical production space, access to other fields, money, distribution, clients, model vendors, technical specialists, and the ability to keep all of that coordinated over time required an organization large enough to hold it together. The organization was doing real work. It was carrying state between crafts, paying the crossing costs, maintaining equipment and space, bringing in outside capability, and keeping one piece of work intact while it passed through all of those hands.
The three inflections attack that requirement from underneath. Coding agents collapse implementation and tool-building into the work. Rich moving representation carries far more of the work across models and crafts without forcing it back through lossy textual handoffs. Cheap local inference makes repeated judgment and coordination affordable.
Where the drastic inflections of Sept 2026 still fall short:
Frontier language models are still not even a little bit good at a professional production quality standard of writing. Screenwriting, is even worse. If a gimmick or trinket is enough for you, fine. You're free to continue paying that cost.
A local writers room, made from an open source model ensemble, on roughly $50k of late 2026 equipment (an RTX Pro 6000 workstation with 96gb of vram, multiple dgx spark or strix halo boxes with 128gb of integrated ram, maybe some older a6000 or 4090 machines for good measure), doesn't make any model a better writer. It separates their failures. Open models that fail differently draft the same scene. Their drafts get compared against each other instead of graded by the model that wrote them, and most of what they write is thrown away before anything expensive happens.
Video models sell as if the clip is the finished work. When a clip misses, you prompt again and pay again. Film has never been made that way. The work gets made in the edit, out of what was actually shot, and with generation that means what was actually rendered. An editors room takes the footage that exists and finds what it already holds: the cut that makes the next render unnecessary, the sound that carries what the picture doesn't have to show, the three seconds of a failed render that solve a different scene.
New space for the intermediate states between world models, video models, robotics, compute, coding agents, and hybrid media production is still waiting for its metagame.