The AI industry pumps out so much hype that you’d be forgiven for simply shrugging at the recent release and sudden withdrawal of Anthropic’s Fable 5 model. Set against the Elon Musk vs. Sam Altman trial and Anthropic locking antlers with the Department of War, the Fable episode could read like just another week in AI.

This one is worth paying attention to. This is really the first time the government has stepped in to regulate a specific model release on the grounds that its capabilities could pose a national security risk in the wrong hands. Whatever happens next, the line has been drawn: a frontier model in general release can be taken off the board because Washington decides it’s too dangerous to leave widely available.

For anyone building AI into their daily work, that shifts the calculation in a real way. The intelligence available to you isn’t only a function of price anymore. It’s also a function of policy, geography, the terms you’re willing to accept on your data, and whether the vendor or the government leaves the model running at all.

The story behind the freeze

For readers who don’t track model releases closely, here’s the short version. Fable 5 is the first generally available model in what the company is calling its “Mythos-class” models, a tier above Opus that Anthropic says has crossed a meaningful risk threshold in cybersecurity and biology. Fable 5 is the consumer-safe version, built on the same underlying Mythos 5 model but wrapped in extra guardrails designed to block or downgrade certain cyber, biology, chemistry, and model-development queries. It also jumps Anthropic’s core model number, signaling a generational step forward from Opus 4.8, Sonnet 4.6, and Haiku 4.5.

Then, on June 12, three days after launch, the government ordered Anthropic to block Fable 5 and Mythos 5 from every foreign national, including foreign-national employees working inside the United States. Anthropic said it could not reliably enforce that distinction and disabled both models globally. The trigger, by most accounts, was a suspected jailbreak that punched through Fable’s cybersecurity guardrails. Anthropic disputed the severity of the finding, saying the demonstration uncovered only minor, previously known vulnerabilities that other public models could also identify.

That fight is still going on. Cybersecurity leaders have urged the government to reverse the order, arguing that defenders need access to the same capabilities and that comparable tools are already available from American and Chinese competitors. Anthropic is working to get Fable back online, and rival labs will almost certainly ship something comparable in short order (some are already claiming to have done so).

The specific dispute may resolve in days or weeks. The precedent will outlast it. A model can be released, integrated into workflows, and then disappear because a government draws a line around who may use it. For anyone building around a single model or vendor (and “building” might simply be leveraging it in crucial, strategic use cases), availability is now part of the risk calculation.

What people saw before the lights went out

Early users got just enough time with Fable 5 to confirm Anthropic’s claims about it. Despite controversies over how Anthropic chose to limit how Fable 5 deals with queries the company deems risky (more on that in a minute), users are seeing the power of the model. Fable 5 is designed for agentic work, meaning it can work autonomously on tasks for a long time, sometimes hours or days, without losing context. The advice that came out of those early sessions was consistent: stop using frontier AI like a fancy autocomplete. The best way to use it, many say, is not to ask it to perform straightforward in-and-out tasks like writing an essay or telling you the best parts of a lengthy report, but to give it broader goals about what you’re trying to achieve, let it build the plan, then execute, however long it takes.

That window was short, but it counted. It showed this level of intelligence is no longer a slide in a research deck. The model was pulled back, but the capability threshold remains crossed.

A big part of what makes Fable work is that it grades its own homework. If you’re a regular user of Anthropic’s models, you’ll notice there’s no “Thinking” mode for Fable 5. That’s because adaptive thinking is always on: The model decides when and how much to reason on every request, and at higher effort levels it can reflect on and validate its own work. Tasks turn into loops. As it works to achieve the goal, it can try things, evaluate the results, change course as needed, and try again. And it can do so autonomously.

For media and marketing teams, the practical shift is in scope. Instead of, say, assigning it to design a specific email campaign, or help format your newsletter, you can zoom out and tell Fable 5 to conceive and build an entire marketing strategy around your newsletter. That might involve reformatting your templates, building new landing pages, adjusting the publishing schedule, building a social campaign, and more. Theoretically, with the right access, it could then build all of that for you. Your job is to grade the output. Over time, less of that grading happens mid-process and more of it happens at the end.

That’s the promise anyway. The danger is that organizations may begin designing around that promise before access, cost, and governance are stable enough to support it.

Fable 5 is the first model that puts real agency on the table. Right now, working with agents, while powerful, involves a lot of management: ensuring the plan the agent builds is correct, clearing up barriers that it encounters as it performs the task, and then guiding it to the best output, usually through multiple iterations on the task itself. In theory, a model strong enough to evaluate its own intermediate work shouldn’t need that hand-holding.

That gap between theory and practice is the real story of the freeze. For a few days, users could test a different relationship with AI; then the capability vanished. We crossed the threshold in the lab and lost it in the market on the same week.

The three walls between you and frontier intelligence

Fable 5 and the models that will follow it stand to change how we work with AI, and arguably how we work, full stop. However, using Fable 5 to its full potential was never just a matter of selecting it in your model picker or calling the API and letting it cook. The pullback put a sharper point on a problem that was already there: the most capable models are also the hardest to actually deploy. I see three walls in the way, with a fourth that just got built.

1. Access and context. For an organization to use Fable 5 to its full potential, it would require a large amount of access to the right context (the org’s information and data). Here, Fable’s strength tripped over itself. Because Anthropic fears the model could be misused, it requires prompts and outputs from Mythos-class models to be retained for at least 30 days for safety monitoring, including in enterprise environments that would otherwise use zero data retention. Anthropic says the data will not be used to train models and that, on some third-party platforms, it remains inside the customer’s cloud environment. But companies cannot use Fable 5 under a true zero-retention arrangement.
   
   That retention requirement, plus the restricted categories where Fable 5 quietly throttles down to Opus 4.8, has set off real friction with enterprise buyers. Many companies will be reluctant to cede control over how their own data is retained and reviewed. Microsoft reportedly limited employee access while its legal teams assessed the implications for confidential and customer data.
   
   And on top of all that sits the new wall. Even if a company accepts the privacy terms, secures the integrations, and builds the right internal controls, the model can still disappear because of a government order or vendor decision. Serious agentic systems will need fallback models, portability across vendors, and a plan for what happens when the most capable model is suddenly unavailable.
2. Compute. Fable 5 is not cheap. Anthropic priced it at $10 per million input tokens and $50 per million output tokens, twice the price of Opus 4.8. I’ve written before about how the agent era is squeezing compute budgets at every layer, and with AI hardening into a political wedge issue, expect compute pressure to stay tight for months and probably years.
   
   The premium price doesn’t automatically kill the math. Some early users argued that it could solve hard tasks in fewer turns than weaker models, potentially lowering the total cost of completing the work. Still, that argument only holds if the work was worth doing with a frontier model in the first place.
   
   If Fable 5 and its peers are going to act as the brains at the top of a company’s AI stack, the deployment question is going to need actual rigor. Organizations will need to be very selective of how to deploy it: which tasks to assign to it, who should have access, and what guidelines, rules, and restrictions there need to be on usage.
   
   And there’s an awkward irony in talking about allocation right now. Intelligence can be technically achievable and commercially valuable while still being unavailable.
3. Task imagination. I became aware of the term “task imagination” through the AI Daily Brief podcast, which references a video by the AI strategist Nate B. Jones. In his take on the Fable 5 release, Jones makes the simple observation that not many knowledge workers think about their work in terms of tasks that might take days to do. It requires a certain level of strategic thinking that may not actually apply to many roles. Put bluntly: a model can run for two days, but most workers have never been asked to define a goal worth two days of machine effort.
   
   For media practitioners, that’s the part worth sitting with. An editor might call on the model to develop more granular editorial guidelines and style guides based on different article types (news, features, evergreen explainers, etc.). Reporters might build investigative agents that don’t just surface data in document troves, but develop research plans based on leads and then execute on them by mining remote databases, filing FOIA requests, and other complex touchpoints that typically require human involvement.
   
   The catch is that most jobs aren’t scoped that way. Many jobs have narrow definitions of what the work is, and there’s little motivation to go beyond that. A model that can do days of work isn’t very useful if the work it’s given is still measured in minutes. That puts pressure on workers to imagine more ambitious tasks or risk being left behind.

The paradox of a pause

The Fable 5 pause comes wrapped in a paradox. A pause gives organizations time to build the governance, data practices, and strategic habits needed to use this level of intelligence responsibly. The trouble is, task imagination only develops with hands on the model. Without access, people cannot discover which long-running assignments are worth the money, where agents fail, or how their own roles could expand around them. The pause buys time while taking away the main way to use that time well.

Step back, and a clearer picture forms. A future where we’re working alongside agents will encounter serious barriers beyond just capability (and political freak-outs over that capability). We restrict access to context so neither the tool nor its creators knows too much. We limit how much we spend on models because we’re unsure of the return we’ll get. And many of us throttle our ambition with AI since our jobs simply don’t have a rich enough canvas for a model like Fable 5 to fill in.

A fourth restriction now sits on top of those three: the model itself may simply not be available.

For media leaders trying to make the ROI case, that’s a problem. The strongest demonstrations depend on giving capable models real work, real context, and enough time to execute. When the most capable models are pricey, hemmed in, or suddenly absent, teams drift to safer pilots that are easier to approve and unlikely to move the underlying economics.

None of these walls fall just because someone ships a smarter model. While advancements in security, infrastructure, and work redefinition will help us get past them, those are inherently slower than the rapid advancement of AI.

We pushed past one threshold and walked straight into several walls. I suspect the story of Fable 5 will be looked back on not primarily as a step up in power, but as the moment where the implications of that power pushed the limits of the systems meant to use it. Agentic AI is clearly where this is going. The systems around it need a beat to catch up.

The pause is useful, but it isn’t free. Experimentation is how organizations learn what this intelligence is actually for. For now, AI leaders are about to discover that running frontier AI at full strength is harder than proving the strength exists. The pure experimentation phase is over. The reality check phase has started, and access, cost, control, and utility now matter every bit as much as raw intelligence does.

A version of this column appears in Fast Company.

The post The Fable 5 pullback turns AI availability into a planning problem appeared first on The Media Copilot.

The free-spending era for AI inside corporate America is ending.

According to The Wall Street Journal, executives at companies including Uber, Meta, Microsoft, Salesforce and DoorDash have launched cost-cutting campaigns this year after seeing their AI bills double or triple—or blow through annual budgets in just three months. The culprits: the soaring price of tokens, the basic unit of AI computing, as model providers like OpenAI and Anthropic seek to balance supply and demand.

The result is a notable shift in corporate AI strategy. Where last year the goal was to flood the organization with AI tools and encourage experimentation, leaders are now scrambling to ration access, steer workers toward cheaper homegrown alternatives, and sharpen employee skills to wring better returns from the technology.

“The free-money period for AI is definitively over,” said one senior technology executive at a major financial firm, speaking on condition of anonymity to discuss internal cost pressures.

The cooling, if it holds, could complicate the growth trajectories of AI heavyweights racing toward public listings. Anthropic closed a $65 billion funding round this week valuing the startup at $965 billion, while OpenAI is also moving toward a potential IPO. AI critics have pointed to corporate cost-management efforts as evidence that the ultrafast pace of AI expansion may be unsustainable.

Budgets burned in months

Corporate spending on AI took off in 2024 and 2025 as companies encouraged broad experimentation, eager to signal to Wall Street that they wouldn’t be left behind in the disruption wave. But many enterprises underestimated how quickly costs would accumulate—particularly as employees without specialized training sent inefficient prompts, ran excessive queries, or used premium-tier models for simple tasks that could have been handled by cheaper, internally built tools.

Some companies burned through their entire annual AI budget in the first quarter. Others saw line items in technology budgets that previously seemed large enough suddenly look inadequate. The problem was compounded for organizations that signed multi-year contracts with AI providers before understanding their actual usage patterns.

“Most companies didn’t have visibility into what AI was actually costing them on a per-team or per-use basis,” said an AI strategy consultant who works with Fortune 500 firms. “They just saw a giant bill at the end of the quarter.”

The rationing begins

At Uber, Meta, Microsoft, Salesforce and DoorDash, technical executives have implemented some combination of the same playbook: tiered access to AI tools based on role and need, mandatory efficiency reviews for high-cost teams, and investment in internal AI infrastructure that costs less per query than commercial models.

Some companies have quietly restricted access to certain premium AI features for non-technical employees. Others have introduced internal dashboards that show employees the real-time cost of their AI queries—designed to encourage more efficient prompting habits.

The shift mirrors what happened in cloud computing’s early years, when companies initially over-provisioned infrastructure before learning to optimize.

The IPO problem

The corporate reckoning comes at a delicate moment for the AI industry. Both Anthropic and OpenAI are navigating toward public markets, and institutional investors are watching corporate AI spending closely for signs that the technology is generating sustainable returns—or that the boom could go bust.

If major corporate customers begin to pull back on AI spending or demand better pricing terms, it could affect the revenue projections that underpin those anticipated listings. Anthropic’s $965 billion valuation, for context, represents a multiple that assumes continued rapid growth in enterprise demand.

AI critics say the cost backlash was inevitable. Proponents counter that efficiency improvements and competition among AI providers will eventually bring down prices—and that early-stage overspend is normal for transformative technologies.

For now, the corporate AI spendometer is being watched more carefully than ever.

The post Corporate America is starting to ration AI as costs skyrocket appeared first on The Media Copilot.

OpenAI today released GPT-5.5, what it says is its “smartest and most intuitive to use model yet, and the next step toward a new way of getting work done on a computer.”

The company says the model understands what users are trying to do faster, can carry more of the workload itself, and excels at writing and debugging code, researching online, analyzing data, creating documents, operating software, and moving across tools until a task is finished.

The company published performance numbers from Terminal-Bench 2.0, which tests complex command-line workflows requiring planning, iteration, and tool coordination.

OpenAI said GPT-5.5 outperformed its predecessor on every major coding and agent benchmark the company tested, while using fewer tokens and running at the same speed as the older model. On one third-party coding index, it matched leading rivals at about half the cost.

Keeping a larger model that fast required rebuilding inference as a single system rather than a patchwork of tweaks, the company said. GPT-5.5 was designed, trained and served on NVIDIA’s latest hardware, and OpenAI credited its own Codex tool and GPT-5.5 itself with helping hit the efficiency targets.

Early testers told the company the model seems to grasp how a codebase fits together — why something is failing, where the fix belongs and what else the change will touch.

Dan Shipper, Founder and CEO of Every, called GPT-5.5 “the first coding model I’ve used that has serious conceptual clarity.” After launching an app, he spent days debugging a post-launch issue before bringing in one of his best engineers to rewrite part of the system. To test GPT-5.5, he effectively rewound the clock: could the model look at the broken state and produce the same kind of rewrite the engineer eventually decided on? GPT-5.4 could not. GPT-5.5 could.

Pietro Schirano, CEO of MagicPath, saw a similar step change when GPT-5.5 merged a branch with hundreds of frontend and refactor changes into a main branch that had also changed substantially — resolving the work in one shot in about 20 minutes.

One engineer at NVIDIA with early access went as far as to say: “Losing access to GPT-5.5 feels like I’ve had a limb amputated.”

OpenAI is already running the model internally at scale. More than 85% of the company uses Codex every week across functions including software engineering, finance, communications, marketing, data science, and product management. The finance team used GPT-5.5 in Codex to review 24,771 K-1 tax forms totaling 71,637 pages, accelerating the task by two weeks compared to the prior year.

The model also shows gains on scientific and technical research workflows. On GeneBench, a new eval focusing on multi-stage scientific data analysis in genetics and quantitative biology, GPT-5.5 outperforms GPT-5.4 on problems that often correspond to multi-day projects for scientific experts. On BixBench, a benchmark designed around real-world bioinformatics and data analysis, it achieved leading performance among models with published scores.

In a notable example, an internal version of GPT-5.5 with a custom harness helped discover a new proof about Ramsey numbers — one of the central objects in combinatorics — later verified in the Lean proof assistant. The result is a concrete example of GPT-5.5 contributing not just code or explanation, but a novel mathematical argument in a core research area.

OpenAI says GPT-5.5 was released with its strongest safeguards to date, including tighter controls around cybersecurity workflows and repeated misuse patterns. The model was evaluated across the company’s full safety and preparedness frameworks, with input from nearly 200 trusted early-access partners before launch.

GPT-5.5 is available today in ChatGPT and Codex for Plus, Pro, Business, and Enterprise users. GPT-5.5 Pro is rolling out to Pro, Business, and Enterprise tiers. API access, which requires different safeguards, is coming “very soon,” OpenAI said.

The post GPT-5.5 Is ‘Our Smartest Model Yet,’ Says Company With History of Saying That appeared first on The Media Copilot.