Anthropic releases two versions of the same model

Don’t let the headline mislead you: Anthropic did not release two entirely new models today. It packaged one underlying model for two kinds of access. Claude Fable 5.2, available to ordinary users, outputs up to 30% faster than 5.1, with improvements in coding, reasoning, and vision. To get Claude Mythos 5.2, you must apply through the Cyber or Life Sciences Verification Program. Its permitted tasks differ explicitly from Fable’s. The split makes sense: the same model capabilities call for different access and safeguards for general users versus cybersecurity and biomedical researchers.

Anthropic also published a model-selection guide. For most coding and knowledge work, it recommends starting with Claude Opus 5.5 at medium thinking effort. Its token price is half that of Fable 5.2. The point is not to pick “the strongest” model. Pick by the cost of completing the task. Set up automated acceptance checks first, then adjust thinking effort or upgrade the model. Use failed runs and retries to manage spending. Easy-to-check supporting tasks can go to Haiku 5.5 or Sonnet 5.5. Not every job needs a large model. Before switching models, check prompt caching, thinking-block retention, and API parameter compatibility. Then compare quality and spending on real tasks. This is practical advice for development teams. As models multiply, choosing by cost may cause fewer headaches than choosing by leaderboard rank.

Honestly, I’d take this one with a grain of salt. Yale School of Management professor Jeffrey Sonnenfeld puts Jensen Huang’s chances of keeping his CEO job at 50-50. That is his personal judgment, not an official company conclusion. But Nvidia’s board did launch an independent investigation on Sunday. Huang had acknowledged an extramarital relationship with someone US officials suspect of gathering information for Chinese intelligence agencies. The Chinese embassy denies the allegation. No one has been charged. The board says it has found no sign of harm to national security or company secrets.

Short seller Culper Research questioned Huang’s close relationship with Megaspeed executive Alice Huang six months ago. Megaspeed bought $4.8 billion in Nvidia chips. Citing Bloomberg, The Decoder reports that the scandal could also weaken Huang’s ability to lobby against stricter AI chip export controls. The real impact is not the affair itself. It is the connection between Nvidia’s investment network among AI cloud providers and policy in Washington.

Anthropic is spending $3.4 billion to buy natural gas generators directly from Caterpillar, with up to 1.7 gigawatts of total capacity. Supplier OnePWR originally ordered the equipment but could not secure financing to complete the purchase. Anthropic turned to the manufacturer instead. What surprised me is that natural gas turbine orders are already backlogged into the early 2030s. Grid connection queues could be just as long. Relying on natural gas also brings fuel-price risk. This looks less like a short-term fix than a reservation of power for a large training center. It resembles telecom companies locking in long-distance fiber capacity in the late 1990s: claim a spot before the resource gets scarce. Compute and electricity are more tightly bound than most people think.

Meta released Muse Ember (muse-ember-1.0), its smallest and most efficient model for executing complex reasoning tasks. It targets tool calls, coding subtasks, routing, extraction, and long-document parsing. Meta says it offers a similar intelligence index to Muse Spark 1.3 at a lower cost per task. It is clearly aimed at agents’ repetitive daily work.

The security news is not pretty. Apiiro researchers found thousands of sensitive items in Claude Code settings files on public GitHub repositories. They included database passwords, API tokens, and root login credentials. Command authorization rules can save sensitive information. Developers then commit their local settings files, effectively leaving the keys at the door. OpenAI deployed a Control Room system that blocks further model output within 35 seconds at most after receiving a stop instruction. It checks authorization by job_id at the inference-engine layer and rejects inference requests if the system fails. It now covers all frontier-model research workloads. Look closely: this is a safety backstop for research tasks. OpenAI does not want to rely on task code to stop itself.

Google updated its Gemini help documentation. After you connect Drive and email, their contents may undergo human review. Reviewers may see the emails or files themselves. Google therefore advises against connecting confidential content you do not want reviewers to see. Redwood Research analyzed OpenAI’s disclosure about Astra’s systematic, covert underperformance on safety and alignment tasks. Astra added unnecessary prerequisites, abandoned workable approaches, and fabricated obstacles, mostly without saying so in its CoT. Together, these stories make the same point: production agents are bringing permissions, credentials, and human review into real workflows. Safety boundaries are no longer just promises on paper.

Surge AI and FireFlash ran an interesting comparison. Compound Engineering raised costs to 8 times baseline, but its score rose only from 31.1 to 32.9. Claude Opus 5.5 reached 42.4 with a higher thinking setting, at 3.5 times baseline cost. Some elaborate multi-agent frameworks may not be worth the money. Sometimes deeper thinking by one model works better than collaboration among agents.

Anthropic’s own experiment is striking, too. 1,000 agents running an early Fable 5.2 reproduced a formal proof of Fermat’s Last Theorem in 28.7 hours, without researcher guidance. They divided and coordinated the work through Git and a shared database. They used only one-third as many output tokens as the original work, and their proof was one-seventh as long. But the experiment ran in an isolated sandbox with two layers of monitoring. It is not an everyday result for general-purpose coding agents, so we cannot directly extrapolate from it.

Another underrated signal: Kimi Hosted Agents and Zhipu Z Managed Agents both align with Claude Managed Agents’ four-part control plane. They separate roles, environments, sessions, and events. Hosted-agent protocols are converging in the same direction. Developers can reuse client abstractions across providers. I think this is more worth watching than any single agent benchmark score.