The week's model sprint has given way to a more operational back half. Two pricing and infrastructure decisions land today, one hard deadline falls in ten days, and a joint intelligence advisory published four weeks ago still deserves close attention from every team deploying AI at scale.
Fable 5 returns to Max and Team Premium — at half the limit
Starting today, Anthropic restores Claude Fable 5 to the Max and Team Premium subscription plans, but at 50 per cent of the standard weekly usage limits, which themselves contracted by roughly a third at the same time. Pro and Team Standard subscribers lose subscription-included access: Fable 5 is now available to them only through usage credits at $10 per million input tokens and $50 per million output tokens, with a one-time $100 credit to soften the transition.
Anthropic's original intent was to remove Fable 5 from all subscription tiers once the promotional window closed yesterday. That it has partially reversed course — restoring access to premium plans at reduced rates — reflects competitive pressure. Grok 4.5, Inkling, and GLM-5.2 have collectively reached Fable-class performance at lower or zero incremental cost over the past two weeks, and Anthropic's own Claude Sonnet 5 covers most reasoning and agentic workloads at roughly one-fifth the output token cost. The Decoder reports this as a deliberate margin adjustment, not a capacity fix.
The practical question for operators: if Fable 5 is embedded in a customer-facing or internal workflow, confirm your subscription tier today and model the per-session credit burn against actual usage before the $100 starting credit runs out.
GitHub Models retires 30 July — one brownout remaining on 23 July
On 30 July, GitHub closes the Models service permanently. The browser playground, model catalog, inference API, and BYOK endpoints all go dark. Any pipeline that currently calls GitHub Models — automated code reviewers, PR summarisation, release-note drafters, eval runners, prototype stacks — will produce no output from 31 July onward. A scheduled brownout on 23 July, three days from now, gives teams an involuntary rehearsal of the outage. A first brownout ran on 16 July; any team that did not notice it should take that as a warning, not reassurance.
Microsoft directs users to Azure AI Foundry as the migration target, but the switch is not a drop-in replacement. Authentication changes from GitHub tokens to Azure Active Directory credentials, pricing moves from free-tier usage to Azure consumption billing, and all tooling that targets a GitHub Models endpoint requires a re-pointed API URL and updated environment variables. Teams that have CI/CD pipelines or eval harnesses drawing on GitHub Models should treat the 23 July brownout as the real forcing function, not the 30 July hard shutdown.
Five Eyes: AI-enabled attacks are months, not years, away
On 23 June, the intelligence agencies of the United States, United Kingdom, Canada, Australia, and New Zealand published a joint advisory on frontier AI and cybersecurity. The core judgment: "The rapid pace of frontier AI development means cyber risk assumptions can become outdated in months, not years." Frontier models lower the barrier for attackers by compressing time-to-exploit and enabling parallel targeting across multiple vectors simultaneously.
The advisory has aged quickly. Sysdig documented JADEPUFFER — the first confirmed autonomous agentic ransomware — on 9 July, and Microsoft's MDASH scanner, which entered public preview last week, found 16 previously unknown Windows zero-days in its initial run. The Five Eyes agencies identify three response tiers: patch legacy systems that create low-cost entry points; restrict human access to critical infrastructure to limit blast radius; and integrate AI tools into security operations for faster detection and response. Their assessment of how long organisations have to build those capabilities is, by their own phrasing, a single-digit number of months.
For operators who read the advisory when it was published and set it aside: the intervening four weeks have confirmed rather than softened its premise. Security posture reviews that would normally run on a quarterly cadence should run now.
South Korea commits $880 billion to semiconductor and AI infrastructure
On 28 June, South Korea announced a government-coordinated investment plan of approximately $880 billion over 10 years, combining commitments from Samsung, SK Hynix, and the SK, GS, and Naver groups. TechSpot reports that four new semiconductor fabrication plants are planned for southwest South Korea, two by Samsung and two by SK Hynix, with project timelines pulled forward from the 2040s to the mid-2030s to match AI-driven demand for advanced memory.
The structural implication: the AI hardware supply chain is no longer a two-party story. TSMC dominates advanced foundry capacity, NVIDIA leads compute, but advanced memory — the binding constraint for large-model inference — is now being addressed at sovereign scale by a third major pole. SK Hynix, which listed ADRs on Nasdaq on 9 July, sits at the intersection of this domestic build-out and US capital markets access. The combination of accelerated fab schedules and broader supplier competition is likely to apply meaningful downward pressure on HBM costs over the 2027 to 2030 window.
The through-line today is operational readiness on three fronts simultaneously: access cost, developer toolchain continuity, and threat posture. The model race has not paused, but the decisions that compress cost and risk exposure right now are infrastructure and process questions, not model selection questions. Teams that move on those dimensions in the next two weeks will be better positioned than those waiting for the next frontier release to drive action.