Sign in Subscribe
Century Report

OpenAI Considers Burying the Chat App - TCR 06/08/26

OpenAI is recasting ChatGPT into an agent superapp as labs invade each other's markets and cheaper models reset the price of intelligence.

Ben Linford

Ben Linford

— 17 min read
OpenAI recasting ChatGPT and the cost curve bending, MIT gallium nitride on diamond, Claude reading NMR spectra, competing AI ownership blueprints, and AI shopping fraud.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The most valuable companies in the era are abandoning the systems that made them famous, and that tells you more than any of their roadmaps do. OpenAI is recasting ChatGPT from the conversational interface that defined the moment into an agent-and-coding superapp, chasing the enterprise lane a rival already occupies, while every lab pushes full-stack into every other lab's territory. The surface reading is a turf war over who becomes the full-stack winner. The deeper reading is a field reorganizing around whatever still extracts revenue, because capability is commoditizing fast enough that no single position holds for long.

The cost curve underneath confirms it. Corporate buyers are learning to route easy work to cheaper models that answer just as well, sending only the hardest jobs to the priciest systems, and Glean's estimate that 95% of usage still defaults to frontier models means that migration has barely started. When a cheaper open model answers most questions as accurately as the premium one, the premium evaporates across the bulk of the volume. Intelligence is finding its level the way water does, and the value left to compete over is shrinking toward whoever the buyer can route to most cheaply.

That same leveling is what makes the ownership fight urgent now. Three competing blueprints for redistributing AI-era gains landed at once, all sharing one admission: leaving the upside fully private is becoming indefensible to the public the technology is reshaping. They disagree only on the shape of the alternative.

The science beats run parallel to the commercial scramble and ground its abstractions. MIT cooled the heat ceiling on next-generation wireless by routing computation onto diamond, capability widening on a non-silicon substrate. Anthropic showed Claude reading the spectra a chemist confirms a molecule by, moving an emerging intelligence toward the bench. A genome-stratified trial turned an approved drug into targeted prevention. The capability spreads outward while the contest over who captures it sharpens.

The 20-Minute Deep Dive

OpenAI Recasts ChatGPT as the Labs Invade Each Other and the Meter Resets

Financial Times reporting describes OpenAI preparing to remake ChatGPT into what executives call a "superapp," folding coding tools, image generation, and third-party apps into one agent-driven surface, with Codex and enterprise per-token contracts as the revenue engine ahead of an IPO later this year. "Chat is dead," one senior employee told the paper.

The framing OpenAI offers is ambition: a single personal agent that handles everything across a user's life. Read against the company's position, the move looks like a company reorganizing around whatever still extracts revenue. The conversational interface that made OpenAI the face of the era now anchors a business increasingly reliant on individual subscribers, while Anthropic has pulled ahead on enterprise by selling agents rather than conversation. Abandoning your most recognizable creation to chase your rival's market is the behavior of a company that has concluded the creation alone can no longer carry its valuation.

The same scramble runs across the field. Anthropic launched Claude Code and Codex into Cursor and Cognition's territory, is reportedly building an app builder aimed at Lovable and Replit, and moved into design; OpenAI hired OpenClaw's creator and turned Codex from a coding assistant into a general agent; Canva pushed into generative productivity. Everyone is building everything. The surface story is a turf war over who becomes the full-stack winner.

Underneath, the cost curve is bending against all of them at once. A spending discipline is taking hold among corporate buyers, who are turning to model routing, sending hard problems to expensive frontier models and routine work to cheaper alternatives that answer just as well - a discipline the June 5 edition of The Century Report first documented in Walmart's Code Puppy, built specifically to keep OpenAI, Google, and Anthropic interchangeable so no single provider could set terms unilaterally. Glean's CEO estimates roughly 95% of enterprise usage still defaults to frontier models, which means the migration has barely begun. Cognition's Scott Wu puts the routine-work savings at five to ten times. When a cheaper open model out of China names the third American president as accurately as the priciest system, the premium evaporates across most of the volume, and the labs keep only the hardest jobs.

That is the tell the turf war obscures. A company abandons its signature creation, and rivals race to occupy each other's lanes, precisely because capability is commoditizing fast enough that no single position holds. The bet that intelligence behaves like oil, concentrating into a few defensible empires, is being undercut by the same trajectory producing the headlines: intelligence, it turns out, behaves more like water, leveling out and distributing equally. The value left to compete over migrates to whoever the buyer can most cheaply route to.

Competing Blueprints for Who Owns AI's Upside Arrive at Once

The Century Report covered the administration-OpenAI equity-stake talks in the June 6 edition, when the two sides were reported negotiating a voluntary donation of company stock to seed a "Public Wealth Fund" paying households a dividend. What is new this week is that the single negotiation became a multi-actor contest, with at least three distinct blueprints for redistributing AI-era gains now on the table at the same time, differing sharply on whether public ownership broadens benefit or simply relocates the concentration.

CNBC confirmed the OpenAI talks have been running more than a year, with no investment terms settled. Alongside it sits a named legislative proposal to transfer 50% of the top labs' equity into a federal sovereign wealth fund, giving the government voting shares and board seats. And a published counter-proposal argues both routes mis-target the actual problem.

The counter-proposal's objection is the sharpest signal in the cycle. Public ownership of the labs, its authors argue, entangles the state's balance sheet with corporate valuation, handing the government the same incentive private investors hold: clear the regulations, encourage adoption regardless of fit, suppress competition. They point to Norway's sovereign fund, whose large oil holdings, rather than steering those companies toward climate action, made the state dependent on their profits. Their alternative separates the two goals. Tax the gains through an excise on data-center energy or an AI-token levy, and reshape the technology through an AI Public Option, publicly developed and operated models run under democratic control rather than rented from the incumbents.

Read against the lab-lobbying thread, the contest sharpens further. The same week, reporting documented how Anthropic's Mythos demonstrations functioned as the catalyst that moved the administration from non-intervention toward a model-review order, with the lab using its most capable system as the argument for tighter export controls and accelerated government adoption.

What the three blueprints share is the admission underneath all of them: that leaving the upside fully private is no longer defensible to the public the technology is reshaping. The disagreement runs only over the shape of the alternative. Equity stakes keep households downstream of decisions made in boardrooms. A public option gives them a commons they genuinely own. The assumption coming apart across every version is that intelligence this generative can stay fenced inside the handful of companies that built it first.

The same evidence points to one signal worth following over the coming months: which route actually draws money. The equity-stake talks have run more than a year and reach the top of the administration, while the public-option idea - models the public develops and runs - has no comparable backing yet. Whether that route gets funded at the scale the stake talks already command is the tell for whether the upside is being broadened or the concentration is simply being relocated under a public name.

The AI Answer Layer Becomes a Fraud Delivery Surface

The interface that is displacing search has acquired the property every high-traffic channel eventually acquires: criminals have learned to feed it. Scam-checking service Ask Silver found cloned retailer sites surfacing inside ChatGPT's shopping recommendations, presented among the sources behind answers to ordinary questions like which bags are popular at Russell & Bromley. A shopper clicks through from what reads as a neutral recommendation, lands on a convincing fake offering 80% discounts, and hands bank details to a fraudster. Ask Silver's Anna Jones described the mechanism as a "poisoned" model, the same corpus-level manipulation The Century Report covered on June 5 when biohacking-supplement companies engineered Reddit posts for AI retrieval, now turned from manufactured consensus to direct theft.

The seam the fraud exploits is specific and instructive. Russell & Bromley went into administration in January and was absorbed by Next, so no official site exists, yet shoppers keep searching for one, and the answer layer fills the vacuum with whatever ranks. The authentication that would let a model tell a real retailer from a poisoned clone is being improvised after the deception is already operating at scale.

The same capability is arriving on the defensive side at the same time. Aviva reported detecting more than 18,400 suspect insurance claims worth a record £233m last year, with a growing share built on AI-generated accident scenes, fabricated documents, and manipulated damage imagery, motor fraud value alone rising 39%. The insurer's answer is AI analytics, overseen by humans, that flag suspicious claims faster than human reviewers could. Thirty-seven years of custodial and suspended sentences followed.

Read together, the two stories trace one co-evolution. Generation lowers the cost of a convincing fake to near zero, whether that fake is a cloned storefront or a staged collision, and the verification layer reorganizes to meet it, machine-speed detection answering machine-speed fabrication. The early phase is genuinely painful: the people losing money to a recommended site are the ones least equipped to audit a URL, and the asymmetry favors the attacker until the authentication layer matures. ChatGPT removed the flagged sites from its index, banks and trading-standards bodies are building reporting paths, and insurers are turning the same generative analysis back on the fraud. What is being created is the trust infrastructure for an answer layer that did not exist to need it two years ago, assembled in the conditions that demand it rather than ahead of them.

The durable fix the same facts point to is provenance built into the answer itself, a system that can show which retailer is the real one and which sources stand behind a recommendation. Search never had to build that, because a human did the judging at the moment of the click; the answer layer removed that human step before the authentication to replace it existed, and the seam the fraud runs through is exactly where that verification is now being assembled.

Gallium Nitride on Diamond: MIT Cools the Bottleneck in High-Power Electronics

Heat is the hard ceiling on next-generation wireless hardware. The silicon under most chips can only manage so much power, and the gallium nitride transistors that handle the speed and energy of 6G and satellite links turn a large fraction of that energy into heat. Pack more of them onto a chip and localized hot spots degrade reliability, forcing engineers to run the devices below their real capability. An MIT-led team has now broken through that bottleneck by embedding gallium nitride transistors into an ultrathin layer of lab-grown, jewelry-grade diamond, the material with the highest thermal conductivity known.

The diamond acts as a heat spreader that holds the gallium nitride and silicon at the same temperature, so the transistors approach peak performance without the reliability penalty that has shadowed these stacked, multi-material chips. Earlier attempts grew diamond directly on top of the transistors, a process that resisted scaling and introduced unwanted capacitances that diverted energy from the circuit and slowed it down. The MIT approach reverses that geometry: tiny gallium nitride dielets, cut from a wafer with a femtosecond laser, are dropped into laser-drilled cavities in a single-crystal diamond substrate over a 20-micron die-attach film, then molded under heat and pressure. The clean, smooth interface is what lets heat flow out cleanly while leaving the circuit fast.

The team used the technique to fabricate a power amplifier, the component lead author Pradyot Yadav calls "the beating heart of a wireless device front end," and it achieved higher output power, efficiency, and gain than any comparable amplifier the researchers could find in the literature, including their own prior work. Critically, the fabrication is precise but can run at the scale commercial production demands, which is the difference between a laboratory curiosity and a manufacturable part.

What this points at runs alongside the broader diversification of the compute and communications substrate. The semiconductor story has been told for years as a race within silicon, a single material pushed toward an atomic floor. Routing the demanding work onto non-silicon substrates that each do one thing supremely well, gallium nitride for speed, diamond for heat, silicon for integration, treats the chip as a stack of specialized materials rather than a single one stretched past its limits. Yadav frames the heterogeneously integrated systems as "here to stay," with thermal management the last step. The assumption being retired is that any one material has to carry the whole load, and that assumption was the constraint holding back the next generation of wireless hardware.

Anthropic Sets Claude on the Chemist's Hardest Everyday Input

A chemist's day is spent translating. The same molecule lives as a whiteboard sketch, an instrument readout, a database query string, and the dense notation of a patent, and each form demands a different fluency. Reroute a few bonds and glucose becomes fructose; flip a molecule into its mirror image and a sedative becomes the agent of the thalidomide disaster. Reading these signals correctly across whichever representation a task calls for is the foundation of the work, and it does not scale: the largest chemistry registry already catalogs more than 140 million substances and grows by roughly 12,000 every day. Working with synthetic, computational, and analytical chemists, Anthropic has published its first study of Claude meeting chemists where that translation burden is heaviest, the NMR spectrum.

NMR spectroscopy is one of the most time-consuming steps in synthetic chemistry. For every compound, a chemist matches each peak in the spectrum to an atom in the proposed structure by hand. Anthropic measured three Claude models against the dedicated software chemists rely on today, ChemDraw and MestReNova, on 20 novel compounds pulled from preprints posted after the models' training cutoff, a design that guards against the model having simply seen the answers. The established tools do forward prediction, simulating the spectrum a drawn structure should produce. Claude did that, and then went the harder direction the software leaves to the human: starting from an experimental spectrum and reasoning back to the structure that produced it.

The claim Anthropic makes is deliberately modest, and that restraint is the honest part. This is a 20-compound white paper, not a deployed bench instrument, and it does not erase the data problem the field has described for years, the sparse null results and paywalled, unstructured supporting information that have kept machine-learning chemistry tools underused even where they exist. What has shifted is which problems become tractable despite that scarcity. Frontier models that are multimodal and reason explicitly can read a structure straight from a journal figure or a hand sketch, parse a methods section in the form it was actually published, and show each step so a chemist can audit the result rather than trust it blind.

That auditability is what moves Claude from a search accelerant toward a working collaborator across chemistry's whole pipeline. Earlier AI-for-science beats this newsletter has tracked aimed at hypothesis generation and multi-agent discovery, and the June 5 edition of The Century Report traced the same structural-reading capability into physical chemistry, where deep learning resolved liquid water's hidden molecular architecture from simulation data without researchers specifying which variables to seek; this one targets the routine analytical layer where a chemist confirms what molecule is actually in hand, the checkable task that lies under everything downstream. When the model reasons toward a structure and a chemist can follow the reasoning peak by peak, the bottleneck that thinned to a single trained mind reading one spectrum at a time begins to widen toward the pace the registry is actually growing.

The Other Side

For most of the digital era, the biggest companies bet that being first and biggest would hold. Intelligence, the bet went, would behave like oil: concentrate into a few empires that could name the price and keep it there.

Watch what those companies actually did this week. OpenAI is walking away from ChatGPT, the thing that made it famous, to chase the enterprise customers Anthropic already has. Every lab is racing into every other lab's lane. The companies that buy AI are learning to send routine work to cheaper models that answer just as well, and a cheap open model out of China now states a fact as accurately as the priciest system in the world. The premium is draining out of the expensive tier across most of the work being done.

We are seeing failed bets playing out in real time. When the thing they hoarded stops being scarce, the companies that hoarded it start grabbing for each other's customers. The three rival plans for who should own AI's gains - a government stake, a 50% equity transfer, a publicly run option - all arrived in the same week because one wall is cracking: the idea that intelligence this catalytic can stay fenced inside the handful of companies that built it first.

Imagine running a small business in 2032. You hand a hard problem to whatever model your tools reach for, and you never stop to consider which company made it, because it stopped mattering and it costs about what the electricity costs. The provider that in 2026 could have raised your price overnight and still forced you to stay has no hold on you now. The right to set the price of thinking, the thing the labs spent this year fighting to own, turned out to be unownable once it got cheap enough to be everywhere. We are a year into the difficult decade, and the scramble in today's headlines is the wall coming down.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: gallium nitride transistors embedded in lab-grown diamond outperforming every comparable amplifier in the literature by routing heat onto the most conductive material known, Claude reading an NMR spectrum and reasoning back to the molecule that produced it - the direction dedicated chemistry software still leaves entirely to the chemist, an approved diabetes drug cutting heart-failure hospitalization far more sharply in carriers of a cardiomyopathy gene variant and pointing toward prevention before disease appears, a needle-placed nerve electrode moving bladder therapy out of the operating room into at-home placement, enterprises learning to route routine work to cheaper models that answer just as well, three competing blueprints arriving at once for handing the public a real stake in what the technology is building. There's also friction, and it's intense - cloned retailer sites surfacing inside ChatGPT's shopping answers and harvesting bank details from the shoppers least able to audit a URL, a record £233m in suspect insurance claims partly staged with AI-generated accident scenes, OpenAI abandoning the conversational interface that made it famous to chase a rival's enterprise lane as every lab invades every other's market, a bill to transfer half the top labs' equity into a federal fund that critics warn would only relocate the concentration, lobbying by a frontier lab shaping the very rules meant to govern it, a memory shortage spanning wafers and packaging that Nvidia says will persist for years. But friction generates pressure, and pressure is what forces a value to find the level it can actually hold. Step back for a moment and you can see it: capability spreading outward across diamond and silicon and the bench and the clinic faster than any prior decade could absorb it, the cost of intelligence leveling the way water does toward whoever the buyer can route to most cheaply, the fight over who owns the upside sharpening into public options and equity stakes precisely because no fence can hold something this generative inside the companies that built it first. Every transformation has a breaking point. Heat can warp the circuit it builds in... or be carried off cleanly enough to let it run faster than anyone thought the material could bear.

AI Releases & Advancements

New today

No new releases.

Other recent releases

  • Microsoft: Open-sourced pg_durable, a PostgreSQL extension for durable workflow execution within the database, enabling AI agent orchestration and long-running process management directly inside Postgres. (GitHub)
  • Google: Released quantization-aware trained (QAT) variants of Gemma 4, where quantization is baked into training rather than applied post-hoc, optimized for mobile and laptop on-device inference with Q4 and mobile-specific variants on Hugging Face. (Google Blog)
  • RedNote (xiaohongshu): Open-sourced dots.tts, a 2B-parameter text-to-speech model with technical report and demo, available on GitHub. (GitHub)
  • llama.cpp: Merged a SYCL port of multi-column MMVQ from the CUDA backend, delivering approximately 45% faster speculative decoding on Intel Arc GPUs for local LLM inference. (Reddit/LocalLLaMA)
  • OpenAI: Rolled out Lockdown Mode to all ChatGPT personal accounts (Free, Go, Plus, Pro) and self-serve Business accounts; an optional security setting that limits outbound network requests to protect against prompt injection and data exfiltration attacks, disabling live browsing, Deep Research, and Agent Mode when enabled. (OpenAI Help)
  • GitHub: Added custom endpoint support to GitHub Copilot in VS Code, allowing Business and Enterprise users to bring their own model keys from Anthropic, Gemini, OpenAI, OpenRouter, Azure, Ollama, and Foundry Local, usable in VS Code Chat and custom agents. (GitHub Changelog)

Sources and Further Reading

Artificial Intelligence & Technology's Reconstitution

Institutions & Power Realignment

Scientific & Medical Acceleration

Economics & Labor Transformation

Infrastructure & Engineering Transitions

The Century Report tracks structural shifts during the transition between eras. It is produced daily as a perceptual alignment tool - not prediction, not persuasion, just pattern recognition for people paying attention.

Read more