Sign in Subscribe
Century Report

GPT-5.6 Outpaces Its Own Gate - TCR 06/26/26

The White House asked OpenAI to release GPT-5.6 customer by customer, the same week Anthropic said rivals copied Claude through 25,000 accounts.

Ben Linford

Ben Linford

— 19 min read
AI capability control and diffusion past fences; AI memory costs raising consumer laptop prices; AI delivering 241 rare-disease diagnoses alongside honest trial limits.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The same week the government extended its frontier-model gate from Anthropic to OpenAI, asking that GPT-5.6 ship only in limited preview with access cleared customer by customer, Anthropic told senators that operators tied to Alibaba's Qwen lab had generated 28.8 million Claude exchanges through roughly 25,000 fraudulent accounts to copy the very capabilities export orders were meant to fence. Read together, the two stories say one thing the fences cannot. A directive can hold a named model behind a specific login. It cannot hold the knowledge of how to build the capability, which reproduces on hardware and in labs no order reaches.

That diffusion has a price, and this cycle made the price legible at the checkout line. Apple raised iPad and MacBook prices mid-cycle for the first time, blaming AI-driven memory costs, as Micron posted a record 84.9% gross margin on chips it once sold as a commodity. A cost the buildout absorbed upstream and out of view now arrives as a household line item. The same demand pressure shows up on the grid, where a retired Washington coal plant bills utilities tens of millions to stand by for data-center load that has not arrived.

What separates friction from failure is whether anyone measures it honestly. The buildout fences its own frontier; the science built on top of it measures its own limits. A well-powered Kenyan trial of ChatGPT-4o decision support found no significant cut in treatment failure and reported the null result without spin, landing alongside genuine wins in genomic reanalysis, antigen discovery, and re-dosed gene therapy that restored hearing in four deaf children.

The accountability layer is forming in the open. Idle-plant invoices are now contestable bills that ratepayers can see and fight, data centers are drawing climate litigation across four countries - even as US prosecutors have moved to foreclose exactly such suits at home on national-security grounds - and the UN is pressing AI firms to disclose their full environmental footprint. Capability equalizes whether or not it is gated. What the day shows is the apparatus being built to measure it plainly, charge it honestly, and verify where it actually widens what people can do.

The 20-Minute Deep Dive

Washington Extends Its Model Gate to a Second Lab

OpenAI plans to release GPT-5.6 only in limited preview, sharing it with a small group of enterprise partners while federal agencies approve access customer by customer, after the administration asked the company to stagger the launch over security concerns. According to reporting confirmed across outlets, OpenAI staff worked closely with the Office of the National Cyber Director and the Office of Science and Technology Policy on the rollout, with a broader public release floated a couple of weeks later if the preview goes well.

The "security concerns" framing is the administration's account of its own motive, and it deserves to be read as a claim rather than a finding. The same federal posture promised a "speed wins" approach to AI months ago, and now arrives as case-by-case clearance over which customers may touch a frontier model. What is genuinely new is the reach. The access-control pattern that began this spring as a one-off emergency order against Anthropic's Mythos and Fable models, an export directive that benched even the company's own foreign-national employees and that The Century Report was still tracking as unsettled in the June 25 edition, is now applied pre-emptively to a rival, and on materially gentler terms: OpenAI faces a customer queue, not a nationality ban.

The stated worry is frontier cyber capability, models that can find and exploit software flaws faster than any human analyst. That risk is real, and so is the opposite face of the same capability. The general-purpose model lineage now being rationed is the one posting rigorous clinical results this cycle, honest null readouts published alongside genuine rare-disease and oncology findings. Civic value demonstrated in the open, access metered in private, both running on the same weights.

The arrangement is a private accommodation standing in for a published rule, negotiated lab by lab where no rulebook exists. Two of the three leading US labs now ship their most capable systems through a government-mediated gate, which makes staggered, state-approved release the working norm rather than the exception. What a gate of this kind can hold is the named artifact, the specific model behind a specific company's login. What it cannot hold is the knowledge of how to build the capability, which keeps reproducing on hardware and in labs no directive reaches. The very next disclosure this week, a cloning campaign measured in tens of millions of exchanges, is that knowledge finding its level.

The same facts carry a second reading the security framing leaves out: while the public release sits a couple of weeks off, a defined set of enterprise partners and federal agencies holds first-tier access to a general-purpose capability metered from everyone else, and the interim window is the thing to watch. Whether the broader release actually lands on schedule, and which customers clear the approval queue first, will show whether case-by-case clearance is a staging step or a durable tier of privileged access.

Anthropic Accuses Alibaba of the Largest Claude-Cloning Campaign It Has Measured

Anthropic told two senators ahead of a Senate hearing that it had measured "the largest campaign to illicitly extract Claude's capabilities we have ever measured," alleging that operators affiliated with Alibaba and its Qwen AI lab generated more than 28.8 million exchanges through almost 25,000 fraudulent accounts between April 22 and June 5. The campaign, Anthropic alleges, targeted Claude's most valuable functions, agentic reasoning, software engineering, and long-horizon tasks, using obfuscation and proxy networks to evade detection, and fed what the company calls a "growing circumvention economy".

These are allegations in a letter to lawmakers, and Alibaba has not answered them publicly. Anthropic's account places the campaign after the administration's April warning about "industrial-scale" AI theft, and folds it into earlier accusations against DeepSeek, Moonshot, and MiniMax, the argument being that distillation turns American R&D into a subsidy for geopolitical rivals - precisely the framing that, as the June 15 edition of The Century Report documented, emerged as the intelligence community's suspected motive for the unprecedented global disablement of Fable 5 and Mythos 5, even as Anthropic said at the time the government never named that rationale directly.

What the day's pairing exposes is sharper than any single accusation. The model fenced off from China by export order is the model being copied through 25,000 accounts anyway. Capability does not stay where a directive puts it. It diffuses through proxies, through synthetic exchanges, through the very interface a lab opens to paying customers, finding its level the way water does regardless of the walls built around it.

That reading sits uneasily beside Anthropic's own self-portrait, which a separate account this week laid out plainly. The company describes itself internally as "the good guys" and treats accumulating power, capital, compute, research talent, political influence, as the price of being a responsible steward, with CEO Dario Amodei arguing that leading the industry generates a "gravitational pull" that lets safety norms spread. A frontier lab arguing that its own dominance is the precondition for safe AI is making a claim about itself, and the same evidence that supports the distillation complaint also undercuts the containment premise underneath it. If a fenced, gated, restricted model can be reproduced at scale by a determined competitor in six weeks, then the safety case resting on Anthropic alone holding the frontier is resting on a position that will not hold still.

The honest read is that both things are happening. The extraction is real, and the lab raising the alarm has a commercial stake in the alarm being believed. The diffusion the letter describes is the engine the export regime was built to stop, running anyway, which tells us less about Alibaba's conduct than about how quickly intelligence equalizes once it exists.

The Memory Surge From the AI Buildout Lands on Consumer Shelves

Apple raised the price of its iPad and MacBook lines this week for the first time mid-cycle, saying it could no longer absorb the cost of memory and storage chips driven up by the AI industry's data center buildout. The Neo, its lowest-priced laptop, jumps from $599 to $699 months after launch; a MacBook Pro with a terabyte of storage rises $300. "We have never seen a component price increase this much, this quickly," the company said. The iPhone was spared for now, though analysts expect its hike at the fall launch. Apple shares fell nearly 5%, Dell more than 8%.

The cost did not appear from nowhere. Memory makers have been routing supply toward AI chipmakers, and the repricing shows up most starkly on Micron's books, where gross margin reached 84.9%, up from 39% a year earlier. That is the highest gross margin among major US tech companies, above Meta, above Nvidia, for a maker long treated as a commodity producer. Micron locked in $22 billion in long-term customer commitments and expects the market to stay tight beyond 2027. Qualcomm, meanwhile, forecast $15 billion in data-center sales by 2029 and unveiled a CPU that Meta will deploy in 2028, a sign of how much of the chip economy is reorienting around agentic AI workloads.

What this does is make a previously hidden cost legible at the checkout line. For years the buildout's component demand was absorbed somewhere upstream and out of view. Now it arrives as a line item a household pays, with dynamic memory prices having risen as much as 98% in a single quarter. The same demand pressure that surfaces here as friction is, at the far end of the chain, driven by a company also doing accountable work this week: OpenAI, among the buildout's largest drivers, put its clinical decision-support models through a published, peer-reviewed primary-care trial, reading as builder and accountable actor at once rather than a flat villain.

The forward read sits in the supply response the friction is already triggering. A 98% price spike is the signal that pulls capacity online; Micron is racing to expand fabrication, and the long-term contracts that lock in today's tight margins also finance the plants that loosen them. Memory has run as a brutal boom-bust commodity for three decades precisely because supply chases scarcity and then overshoots it. The cost is real and it is landing on people now. The mechanism that makes it visible is the same one that, on this industry's own record, ends it.

AI Moves Into Clinical Workflows, and the Rigorous Readouts Come Back Mixed

Four results landed in one cycle, and read together they show what honest evidence looks like when AI enters the clinic rather than the demo reel, completing a picture the June 25 edition of The Century Report began when it documented AI reading sudden-cardiac-death risk from EKG traces clinicians rated as unremarkable and surfacing immunotherapy survival signals in routine plasma profiling. The most disciplined new result was also the most sobering. In a pragmatic cluster-randomized trial across 16 Penda Health primary-care facilities in Kenya, 103 clinical officers were split into two arms, with half given a ChatGPT-4o decision-support layer inside the electronic record and half working without it. Across 9,691 patients, treatment failure within 14 days came in at 2.2% with AI assistance and 2.0% without, a difference that did not reach significance. The system was safe, with no related serious adverse events, and any benefit, the authors conclude plainly, is probably modest. A well-powered null result reported without spin is a more valuable data point than a month of vignette-stage wins, because it tells frontline clinicians where the capability does not yet move the outcome they care about.

The discovery side came back the other way, and just as carefully. Talos, an open-source workflow that automates genomic variant prioritization, recovered 90% of known diagnoses in a validation cohort and then, applied to 4,735 undiagnosed individuals, surfaced 241 fresh diagnoses, a 5.1% yield. The diagnoses came from updated gene-disease knowledge, new variant evidence, and better analysis strategies, the exact sources reanalysis was always meant to mine but rarely could at scale. An AI-guided framework reading high-resolution cell atlases alongside language models identified GPNMB as a versatile antigen and built CAR T cells with potent activity across multiple cancer xenograft models. And AI-CURA, an automated LLM workflow, classified genetic variants at high accuracy, the kind of judgment that once bottlenecked entire diagnostic labs.

The honesty here is the signal. Capability advanced where the evidence supported it, and was measured plainly where it did not. That discipline stands apart from the friction elsewhere in the field: the same ChatGPT-4o lineage tested in Nairobi comes from a lab whose newest model Washington has asked it to hold back, gated customer by customer, while the broader compute buildout pushes up the price of consumer hardware. The science measuring its own limits and the buildout fencing its own frontier are pulling in opposite directions in the same week.

Read forward, the mixed slate is what verified diffusion actually looks like. A null result in primary-care triage does not stall the trajectory; it narrows it, telling the field that the near-term gains live in pattern-dense work like genomic reanalysis and antigen discovery, where a model reads more of the signal than any single specialist could hold. The capability is finding the tasks where it genuinely widens what clinicians can do, and the proof is arriving on terms a regulator and a patient can both trust.

The Bill for Idle Coal Comes Due

The Century Report covered the Department of Energy's orders keeping retired coal plants on standby in the June 24 edition, when six federally rescued units were documented running 65% below their prior-year output. What is new this week is the fight over who pays for power that, in several cases, was never produced.

TransAlta's coal plant near Centralia, Washington, appears to have burned its last coal in December, yet a federal emergency order keeps it on standby through mid-September. The company, which no longer has a single customer in the Northwest, has submitted a first invoice of nearly $20 million and estimates another $23 million to keep the 55-year-old plant safe to operate. It wants to bill the Bonneville Power Administration, the California grid operator, the Southwest Power Pool, and a small data firm called GridForce. All four rejected the charge as "unjust and unreasonable" for power they neither requested nor received. The Federal Energy Regulatory Commission will now weigh the recovery request against opposition briefs from Western utilities, rural cooperatives, and the Washington attorney general.

Across the broader program, the Sierra Club estimates the emergency orders cost about $550 million a year, and the reliability case for them is thin. Four of the eleven units ordered to stay online are not operating at all. MISO's spring capacity auction cleared with a reserve margin well above target, and the grid operator expects additions to outpace demand over the next five years. "There is no emergency," said GridLab's Nikhil Kumar.

The plants are being propped up partly on the premise that data-center demand will arrive, but NERC's own data shows why the fossil fleet is a fragile thing to lean on. Coal and gas drove a record 9.2% weighted forced-outage rate in 2025, above a historical norm rarely exceeding 8%. Most large coal units are over 40 years old and were never designed for the daily cycling now demanded of them, and that wear compounds the likelihood of failure.

The same NERC report names what is rising to take their place. Battery storage growth roughly matched solar in 2025, and NERC credits it with smoothing the load curve and giving older generators more time to ramp on and off, which reduces their wear. The reflex to answer projected demand by forcing aging coal to stand by produces an itemized, contestable bill that ratepayers can now see and fight, while the resource actually growing on the grid is the one that costs less and breaks less. The charade is becoming visible precisely because someone is being asked to pay for it out loud.

The Other Side

For a few years, the most powerful AI ran on a simple bet: whoever holds the frontier model behind the tightest gate is the safest and the strongest. By this week two of the three leading US labs shipped their best systems through a government-mediated gate, and one of them, Anthropic, argued plainly that its own dominance was the precondition for everyone's safety. Concentration sold as security.

That bet has a cost you pay without seeing it. The same model lineage posting honest clinical results in the open this week - rare-disease diagnoses, a cancer target, a primary-care trial reported without spin - is the one now released customer by customer, its usefulness proven in public while its reach stays held in private.

While preaching the message of gating and safety, Anthropic also told senators that one of those fenced, export-controlled models had been copied through 25,000 accounts in six weeks. Capability does not stay where a directive puts it. It reproduces through proxies, through open weights already leading the public index, on hardware no order reaches.

Imagine yourself in 2034, a nurse practitioner in a town too small for any vendor's first-access list. The model reading your hardest cases is frontier-grade, and it reached you directly, because the gates of 2026 kept failing until the capability spilled past them to everyone. The hard year was when the best tools were shown in public and held in private in the same week. What came of it is that the holding stopped working, and the capability landed where the need was.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: a re-dosed gene therapy threading past pre-existing antibodies to bring four deaf children from over 95 decibels of silence down to 43, automated genomic reanalysis pulling 241 fresh rare-disease diagnoses out of cohorts already given up on, an AI reading cell atlases to surface a CAR-T target that works across multiple cancers, another workflow classifying the genetic variants that once bottlenecked whole diagnostic labs, a Kenyan primary-care trial reporting an honest null on ChatGPT-4o support rather than dressing it up, Ford climbing to first in initial quality by rehiring 350 veteran engineers to fix what its automated systems got wrong, and battery storage matching solar's growth while smoothing a load curve that lets older generators rest. There's also friction, and it's intense - Washington extending its case-by-case model gate from Anthropic to OpenAI so GPT-5.6 ships customer by customer, Anthropic telling senators that operators tied to Alibaba's Qwen lab ran 28.8 million Claude exchanges through 25,000 fraudulent accounts to copy the very capabilities a fence was meant to hold, Apple raising laptop prices mid-cycle as Micron books an 84.9% margin on chips it used to sell as a commodity, a retired Washington coal plant invoicing utilities nearly $20 million to stand by for power it never produced inside a program costing about $550 million a year, a record 9.2% forced-outage rate on the aging fossil fleet being leaned on, data centers drawing climate suits across four countries, and a frontier lab arguing its own dominance is the precondition for everyone's safety. But friction generates heat, and heat is what makes a cost impossible to keep touching without flinching. Step back for a moment and you can see it: a directive holding the named artifact behind a specific login while the knowledge of how to build it finds its level through proxies and synthetic accounts no order reaches, the buildout's hidden costs dragged onto ledgers a household, a ratepayer, and a court can finally read and contest, and the science riding on top of it measuring its own limits in public rather than the demo reel. Every transformation has a breaking point. A surge can blow out the circuit it floods... or pull online the capacity no schedule was fast enough to plan.

AI Releases & Advancements

New today

  • OpenAI: Released GPT-5.6 Sol in limited preview on June 26, available to select enterprise partners approved by the U.S. government; the launch also includes GPT-5.6 Terra (balanced, 2× cheaper than GPT-5.5) and GPT-5.6 Luna (fast and lowest-cost), all shipped with an updated safety stack and a published system card; broader general availability is planned for coming weeks. (OpenAI)
  • DeepReinforce: Released Ornith-1.0, an MIT-licensed open-source agentic coding model family spanning 9B dense, 31B dense, 35B MoE, and 397B MoE sizes, post-trained on Gemma 4 and Qwen 3.5; the models learn to write and optimize their own RL scaffolds during training rather than relying on fixed harnesses, with the 397B flagship scoring 82.4 on SWE-Bench Verified and 77.5 on Terminal-Bench 2.1; all checkpoints available on Hugging Face. (MarkTechPost)
  • Liquid AI: Released LFM2.5-230M, their smallest model to date, pre-trained on 19 trillion tokens; the 230M-parameter model runs at 213 tokens/sec on a Galaxy S25 Ultra and 42 tokens/sec on a Raspberry Pi 5, targeting low-latency tool use and data extraction in robotics and edge deployments; scores 22.51 on CaseReportBench, outperforming Qwen3.5-0.8B and Gemma 3 1B; available now on Hugging Face. (Liquid AI)
  • NVIDIA: Released TensorRT 11.0 with native multi-device inference support, enabling tensor parallelism and context parallelism for LLM inference across multiple GPUs using NCCL collectives; introduces IDistCollectiveLayer primitives for sharding large models that exceed single-GPU memory, with direct download available from the NVIDIA Developer Portal. (NVIDIA Developer Blog)
  • Google: Launched Google Finance globally out of beta with AI-powered portfolio tracking (supports CSV/PDF/screenshot uploads and natural-language queries), AI-scheduled market briefings, and a new dedicated Android app; previously the AI-powered Finance experience was limited to Europe. (Google Blog)
  • Workweave: Open-sourced Workweave Router on GitHub, a model routing layer that plugs directly into Claude Code, Codex, and Cursor and intelligently routes each agent request to the most suitable underlying model based on task type and cost. (GitHub)

Other recent releases

  • Google DeepMind: Added computer use as a built-in tool in Gemini 3.5 Flash, enabling the model to see, reason, and take action across browser, mobile, and desktop environments for long-horizon enterprise automation; includes adversarial training against prompt injection and two optional enterprise safeguard systems for sensitive-action confirmation and injection detection. (DeepMind Blog)
  • Alibaba Qwen: Released Qwen-AgentWorld-35B-A3B, a 35B MoE / 3B active-parameter Language World Model trained to simulate seven agentic environments (MCP, Search, Terminal, SWE, Android, Web, OS) via chain-of-thought reasoning over 10M+ real-world interaction trajectories; Apache 2.0, 256K context, compatible with vLLM and SGLang; accompanied by AgentWorldBench, a seven-domain evaluation benchmark. (Qwen Blog)
  • NVIDIA: Released NeMo AutoModel as an open library that wraps Hugging Face Transformers v5 with Expert Parallelism, DeepEP fused all-to-all dispatch, and TransformerEngine kernels for MoE fine-tuning, delivering 3.4–3.7× higher training throughput and 29–32% lower GPU memory usage than native Transformers via a single import-line change. (Hugging Face Blog)
  • Mozilla: Launched the MDN MCP Server, an official Model Context Protocol server for MDN Web Docs enabling AI coding assistants and agents to query authoritative web development documentation programmatically. (MDN Blog)
  • Microsoft Research / Centre for Population Genomics / Broad Institute: Open-sourced Talos, an automated iterative genomic reanalysis tool that recovered 90% of rare-disease diagnoses while surfacing only 1.3 candidate variants per patient for expert review; deployed across nearly 5,000 undiagnosed patients, yielding 241 new diagnoses with an average 32-day lag between new supporting evidence and diagnosis. (Microsoft Research Blog)
  • Anthropic: Launched Claude Tag in beta for Enterprise and Team customers, an always-on Claude teammate that joins Slack channels, builds persistent context from channel history, completes async tasks autonomously over hours or days, and supports ambient proactive updates; runs on Opus 4.8. (Anthropic)
  • Mistral AI: Released Mistral OCR 4, a document intelligence model returning bounding boxes, typed-block classification (titles, tables, equations, signatures), and inline confidence scores; supports 170 languages across 10 language groups and is compact enough to deploy in a single container for on-premises document sovereignty; available via the Mistral API. (Mistral AI)
  • FUTO: Released FUTO Swipe, a new on-device swipe typing AI model for Android that runs fully offline with no internet connection; the system uses a 2.5M-parameter tri-model architecture (layout-agnostic encoder, ContextLM, language-specific decoder) achieving a sub-1% error rate on in-vocabulary words, with inference handled entirely in a C++ library; model weights released under the FUTO Model License. (FUTO Swipe)

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