Claude Fable 5 arrives as investors question AI's costs

Fable 5 lands with caveats attached

Anthropic's new flagship is pitched as Mythos-class and reportedly beats recent frontier models by 10 to 15 percent on Hex's data-analysis evals. AWS made Fable 5 available on Bedrock with built-in safeguards, aiming it at enterprise workloads that need compliance controls as well as capability. Endor Labs' assessment is more mixed, pairing Mythos-grade hype and record cheating with a few hall-of-fame entries, so the headline numbers deserve some caution. A separate forecast argues that Mythos-class capabilities will diffuse through open weights by 2029. If it is right, teams committing to Fable 5 now may want to weigh lock-in against a future where comparable capability runs on self-hosted models.

Anthropic also backtracked on a policy researchers said sabotaged their work, following pushback from the research community. The episode shows a recurring tension for frontier labs: restrictions meant to reduce misuse can also limit the outside scrutiny that helps catch safety problems.

Agents run longer and start paying

Xiaomi's MiMo Code harness reportedly beats Claude Code on 200+ step tasks, the long chains where context management and planning tend to break down. That is a narrow result, but it tests sustained autonomy rather than one-shot code generation, and it comes from an open-source project engineers can run themselves. Visa will give AI agents their own cards, with OpenAI as launch partner, supplying its network, credentialing and security infrastructure for agent-initiated payments. Further out, a piece on first steps toward automated AI research describes systems handling hypothesis generation, experiment design and evaluation, not only implementation. With less human oversight in the loop, how labs verify those results becomes more consequential.

Cutting compute waste before and during runs

Two pieces focus on the cost of training and inference rather than new capabilities. Goodfire's predictive data debugging research describes tooling that reveals and shapes what a model will learn from a dataset before training, so data quality and bias issues surface before an expensive run. Lower in the stack, Hugging Face's guide to fusing PyTorch MLP layers into single kernels shows how to reduce memory traffic and kernel-launch overhead, the GPU fundamentals that still largely determine real-world cost and latency.

AI spending reaches budgets and balance sheets

EE Times reports that an AI-driven memory shortage is upending IT budgets, as rising DRAM and HBM costs and allocation priority for AI workloads crowd out conventional enterprise compute purchases. IT leaders are rethinking procurement timing and capacity planning as a result. Investors are applying similar pressure to vendors. Oracle beat on revenue, but its AI infrastructure spending worried investors. Adobe topped earnings targets yet sank to a seven-year low as its CFO departed, amid skepticism about its position in the AI era. In both cases, a strong quarter did not settle doubts about AI strategy.