Fable 5.2 - Cost-Effective Banger; Meta's Frugal Agent; OpenAI Plays Your Games

Today's AI Outlook: 🌩️

AI’s Power Hunt Hits the Salvage Yard

Data center developers are hunting for used gas turbines and retired jet engines as demand for AI computing strains the supply of new power equipment. The reporting puts the average U.S. wait for a new turbine at six years, with delivery discussions stretching into 2030 and 2031. Builders are refurbishing idle equipment and adapting aircraft engines to generate electricity, while utilities face the same crowded shopping list.

Why It Matters

Reliable electricity and equipment delivery dates constrain how quickly data centers can open. Competition for turbines also affects utilities buying capacity for communities that have little involvement in the AI boom.

The Deets

• Reported global turbine orders exceed 100 gigawatts, compared with factory capacity of 60 to 70 gigawatts.
• Well-maintained used turbines reportedly cost $1.5M to $1.7M per megawatt, versus $2M to $2.5M for new equipment.
• The Tennessee Valley Authority reports rising turbine prices over the past two years, while rural electric cooperatives are considering refurbished units.

Key Takeaway

Compute expansion depends on industrial supply chains whose timelines are measured in years.

🧩 Jargon Buster - Gas Turbine: An engine that burns fuel to spin machinery, which can drive an electrical generator.

Gmail’s AI Privacy Promise Gets Fine Print

Google has added a qualification to an earlier promise that content connected through Gmail, Docs and Drive would not be seen by human reviewers. According to today’s reporting, linked Workspace data can be used for model training and human review in limited circumstances, and Gemini’s help documentation describes the use of email excerpts, potentially including an entire short email. The exact circumstances matter when deciding what to connect.

Why It Matters

Connecting an inbox can bring private correspondence into an AI interaction. A broad privacy promise is less useful than a clear account of which settings and circumstances permit training or review.

The Deets

• The reporting identifies a note added last month to Google’s 2023 announcement.
• It describes human reviewers, including contractors, grading responses and supplying improved answers.
• The linked clarification pages could not be retrieved for this edition, so the precise account types, exceptions and settings were not independently confirmed.

Key Takeaway

The reported exception deserves attention, but it should not be read as evidence that every connected mailbox is used for training.

🧩 Jargon Buster - Human Review: People examining AI interactions or answers to assess quality and help improve a system.

♟️ Power Plays

Google Gives Its Chips New Job Descriptions

Google has announced two specialized AI chips, the TPU 8tx and 8i Max, expected to reach customers in 2027. One is designed to keep sensitive data encrypted while a model processes it; the other combines two chips to handle workloads that need more processing power and memory bandwidth, such as agents working across large codebases or long videos.

Why It Matters

Specialized chips give businesses more options for running demanding AI workloads. The encryption feature is particularly relevant to organizations handling sensitive information, while extra memory bandwidth can help agents work with larger inputs.

The Deets

• The 8tx targets governments and regulated industries by protecting data during processing.
• The 8i Max joins two chips into one unit, doubling processing power and memory bandwidth, according to the reporting.
• Google reportedly has nearly 100 TPU customers, including Apple and Anthropic, with available supply already committed.

Key Takeaway

These chips are a 2027 prospect; near-term purchasing decisions still depend on available capacity.

🧩 Jargon Buster - TPU: A processor Google designs specifically to accelerate AI calculations.

🛠️ Tools & Products

A $200 Handheld for Your AI-Built Games

OpenAI’s merchandise store has begun selling the $200 ModRetro Chromatic: Codex Edition, a retro handheld that plays games users build with AI assistance. A ModRetro plugin lets users describe a game, generate sprites, maps and logic, then test and revise it before writing it to a cartridge. The AI helps make the game; the finished game runs without a model on the device.

Why It Matters

The workflow gives hobbyists a tangible way to play software they create through prompts. The hardware is optional for trying the games, which also run in a browser or on compatible handhelds.

The Deets

• The $200 package includes a blank cartridge but no Codex subscription.
• Users can revise game mechanics and visuals in plain language before writing the finished game to a cartridge.
• The handheld was previously given away at DevDay in September; this release makes it a retail product.

Key Takeaway

The creative tool is the plugin; the handheld adds a physical way to play and share the results.

🧩 Jargon Buster - Runtime: The period when a finished program is running. These games do not need an AI model during play.

🧠 Research & Models

Meta Puts Agent Busywork on a Budget

Meta has released Muse Ember, its smallest AI model, to handle the repetitive work inside AI agents. The company says Ember matches Claude Haiku 5.5 on a major intelligence benchmark while completing an average test task for $0.14 versus Haiku’s $0.41. Its pitch to developers is straightforward: let a larger model plan the work and give the smaller one the busywork.

Why It Matters

Agent systems may make many model calls to finish a single job, so small efficiency gains can add up. Ember’s reported savings come from using fewer tokens per task, even though its token pricing matches competing models.

The Deets

• Ember and Haiku 5.5 each scored 36 on Artificial Analysis’ Intelligence Index; GPT-6 Luna scored 28.
• Meta positions Ember alongside the larger Muse Wonder, with planning handled by the larger model.
• Meta says it trained Ember using Muse Wonder Full and is developing more efficient models.

Key Takeaway

Compare the cost of a successfully completed task on your own workload; the reported test average is a starting point.

🧩 Jargon Buster - Token: A small unit of text or other information a model processes, commonly used to calculate API charges.

Anthropic Drops Fable 5.2 with a Bang

Anthropic has released Claude Fable 5.2 and Mythos 5.2, upgrades that bring lower prices and faster, incredible performance to its most powerful model family. Fable took first place on Artificial Analysis’ Intelligence Index, with a score of 65, five points ahead of Google’s unreleased Gemini 4 Carbon. Mythos remains restricted to vetted security and life sciences groups, while Fable adds stronger safeguards around cybersecurity and biology.

Why It Matters

The release gives developers stronger performance at a lower price while illustrating how Anthropic is limiting selected capabilities. Its approach leaves powerful general-purpose tools available while reserving the less restricted model for approved specialists.

The Deets

• The reporting puts 5.2 at 20% cheaper and up to 30% faster than 5.1.
• Fable costs $8 per million input tokens and $40 per million output tokens, still twice the reported price of Opus 5.5.
• Anthropic says Fable coordinated more than 1,000 subagents that coded and tested for eight days, producing an operating system with its own kernel and drivers that boots on a laptop. A separate demonstration formalized a mathematical proof in 30 hours. These are demonstrations, not guarantees for everyday projects.
• Some skills that accelerate AI development were held near Opus levels.

Key Takeaway

Fable raises the performance bar, but buyers still need to weigh its premium price against their actual tasks.

🧩 Jargon Buster - Guardrails: Restrictions intended to reduce unsafe or prohibited uses of an AI system.

⚡ Quick Hits

• CoreWeave and Wayve: The cloud provider signed the self-driving startup to a multiyear training deal. Financial terms were not disclosed.
• Instagram age checks: Meta is extending AI-based identification of suspected teen users globally, placing more of them into restricted accounts.
• Disney and Adobe: A multiyear agreement will support custom models trained on Disney characters and visual styles, with work underway across theme parks, film and ESPN.
• Anthropic’s IPO: The Wall Street Journal reports that trading could begin as soon as Thursday next week, despite market unease. That timing remains a reported plan.
• Fireworks breach: The AI hosting startup confirmed a hack exposing billing information and portions of some customers’ models and datasets, ahead of a planned fundraising effort at a valuation above $30B.

🛠️ Tools of the Day

• Claude in Excel: This guide covers cleaning a sales export, fixing totals and building a report. Keep an original copy, review proposed changes and test whether new rows update the chart.
• Instagram Effect Builder: U.S. users can turn text prompts into Stories effects with this free web tool. Effects must pass a safety review before others can use them.
• Seedream 5.5: ByteDance’s image model is highlighted for its editing abilities. Today’s issue does not provide pricing or detailed test results.
• Inkling 2 Small: Thinking Machines Lab’s model handles coding and vision tasks, according to today’s tool listing.
• Stride workflow: A reader-built project combines Garmin running data with weather and elevation, then asks Claude to explain patterns. It is a community example, with performance claims supplied by its creator.

Today’s Sources: The Internet, AI Secret, The Rundown AI