Eight dollars in, forty dollars out. That is Anthropic’s new price per million input and output tokens for Fable 5.2 and Mythos 5.2—not a cheap conversation, but a lower entry price for its most demanding technical work. The consequential part of AI news today is the combination: Anthropic says outputs are faster, while access to the less-restricted Mythos remains gated. If you build with these models, the decision is about cost per completed, checked task, not simply the newest name in a menu. Monday’s other developments make that distinction sharper. Google is tailoring TPU systems for confidential inference and long context; OpenAI has described a centrally enforced stop mechanism for research workloads; NIST wants feedback on an agent-focused attack taxonomy; and Meta is expanding AI age estimation globally. Better capabilities are arriving alongside more specific controls. Whether either side works on your workload still requires testing.

AI news today: Anthropic lowers the price of ambitious technical work

Anthropic released Fable 5.2 and Mythos 5.2 on November 9, saying output generation is up to 30% faster than the 5.1 versions. Optional fast mode doubles the price. These are provider measurements, not guarantees for every application.

The models share an underlying base, but not the same access policy. Fable includes cybersecurity and biology safeguards; organizations seeking Mythos for legitimate research must apply through verification programs. Anthropic recommends starting most everyday work with Opus 5.5 rather than automatically choosing its largest option.

The Verge’s launch coverage places the release in the competition for long-running technical projects. For engineering leaders, the useful experiment is a fixed migration or debugging task with a known acceptance test. Compare completion, review effort, latency and total spend. Faster tokens do not necessarily mean faster approved software.

Google expands TPUs through packaging and confidential inference

Google announced TPU 8tx and TPU 8i Max, bringing its eighth-generation portfolio to four systems. Importantly, these are new system configurations, not an entirely new silicon generation. The 8tx pairs existing TPU 8t silicon with fifth-generation AMD EPYC hosts and hardware protection for sensitive data and model weights during processing.

The 8i Max puts two TPU 8i chips in one package. Google says it doubles compute, SRAM and high-bandwidth-memory bandwidth relative to 8i, targeting large-context workloads and higher rack density. Those comparisons should not be read as a promise to halve your bill.

For buyers, the split makes requirements more explicit: confidentiality, memory bandwidth and context size can call for different infrastructure. Request workload-specific pricing, availability and measured latency before migrating. Our recent AMD data-center briefing provides context on the wider race to turn hardware announcements into sustained customer demand.

OpenAI describes a 35-second stop for research model output

OpenAI’s Control Room centrally tracks inference workloads across its research infrastructure. The company says it can halt model output within 35 seconds after a decision to intervene. Engines check whether a job is permitted to generate output, and the system is designed to fail closed if Control Room goes down. OpenAI says that fail-closed protection covers frontier research workloads, with the remaining research rollout expected over the following days.

That timing starts after the intervention decision—not when suspicious behavior begins. Nor does stopping token generation automatically reverse a tool action already completed. Those distinctions matter when evaluating any claim about an AI “kill switch.”

OpenAI says the system is already deployed internally and works with misalignment monitoring; this is not an announcement of a generally available customer product. Builders can borrow the architectural lesson: enforce authorization outside the agent, connect every workload to an accountable job identity, and test the stopping path when other systems fail. Our previous brief on AI controls and infrastructure explains why operational evidence is becoming as important as capability benchmarks.

NIST opens comments on agent-focused adversarial machine learning

NIST released an initial public draft of its 2026 adversarial-machine-learning taxonomy on November 9, with comments due December 18. The draft focuses on attacks against agentic systems and corresponding mitigations. It organizes threats by factors including lifecycle stage, attacker goals, capabilities and knowledge.

This is a draft terminology and taxonomy document, not a new binding compliance requirement or a certification that any product is safe. Its practical value is a common language: procurement, engineering and security teams need to describe the same attack without silently meaning different things.

A sensible next step is to map an existing agent threat model to the taxonomy, record ambiguities, and submit concrete feedback before the deadline. Pay particular attention to risks introduced by tools, external content and delegated permissions. A longer checklist helps only if it changes a test, an access boundary or a deployment decision.

Meta takes AI age estimation global on Instagram

Meta says it is expanding AI age estimation globally to move suspected teenagers into Instagram’s Teen Account protections, including people who entered adult birthdays. It expects millions more accounts to receive those protections. The company also says it has begun using Apple’s optional age-range signal in Australia and plans similar age-estimation technology globally on Threads later this year.

The announcement establishes Meta’s rollout claim, not independent proof of accuracy or harm reduction. Age estimation is an inference about a person, so false positives, missed teenagers, appeals and data minimization belong in the evaluation alongside coverage.

For parents and platform operators, the next meaningful numbers are error rates and correction outcomes, not simply accounts reclassified. Meta’s broader argument—that app stores should provide a consistent age signal—also raises a governance question: who verifies, who receives the result, and how much personal information must move between them?

Watch & Learn

Watch “What is Confidential Computing?” — IBM Technology (8 minutes 45 seconds). Alex Greer uses a lightboard to explain trusted execution environments and why protecting stored or transmitted data does not cover processing. It is an accessible primer for buyers evaluating Google’s 8tx announcement. The concepts remain useful, though the product examples are from 2020. The public video is provided as a text link rather than an untested WordPress embed.

AI, Translated: confidential computing

Confidential computing protects data while a computer is processing it, typically inside a hardware-backed trusted execution environment. For example, a model could analyze sensitive records inside an isolated environment rather than exposing them to the surrounding host system. It complements encryption for stored data and network traffic; it does not make model answers correct or eliminate every attack. You should care because “encrypted” alone says little about what happens during inference. IBM explains the underlying security model.

Try This Today: build a model acceptance test with Grok

Goal: spend ten minutes defining what a model must do before you pay to deploy it. Use fictional tasks, not customer records.

1. Open a Grok chat and use the + button to attach a small TXT or CSV containing five sample tasks.
2. Ask for a pass/fail test matrix with expected evidence, acceptable latency and human-review triggers.
3. Deliberately add one impossible or ambiguous task, then check that the matrix requires escalation rather than a confident guess.

Copy-ready prompt: “Turn these fictional tasks into an acceptance-test matrix. Include expected result, evidence required, failure condition, latency budget, review owner and stop condition. Do not invent benchmark results or prices. Mark missing requirements as unresolved. Include one test for refusing an unauthorized action.”

Grok’s official upload guide confirms TXT and CSV support on web and mobile; limits vary by platform or subscription. Manually inspect the resulting tests before using them to judge another model.

One thing to remember

A cheaper, faster model is valuable only when it completes the right task inside the right boundaries. Price the verified outcome—and test the controls with the same seriousness as the benchmark.