Hey there 👋 I read two model posts back to back this morning, and they pointed in opposite directions. Anthropic shipped Fable 5.1 and loosened the security brakes: fewer false blocks, and the model can now point at a vulnerability. A few hours earlier OpenAI rated its Astra model “Critical” on cyber and locked the sharp part behind a vetted coalition. Same capability, same week, two labs betting the opposite way on who gets to hold it. If I had to bet, both are right for their own book, and you are the one who has to pick a lane. Let’s get into it.
The Big ThingFable 5.1: the agent tax just dropped, and the model stopped crying wolfAnthropic released Claude Fable 5.1 on Tuesday, and the interesting part is the bill. Anthropic says typical workloads run about 25% cheaper than Fable 5, and highly agentic or coding-heavy jobs land closer to 45% cheaper. The sticker price per token did not move, it stayed at $10 in and $50 out per million. The savings come from the plumbing: prompt cache reads dropped 75%, down to $0.25 per million tokens, and long agent runs lean on cache constantly. For anyone running agents that loop, retrieve, and re-read the same context all day, that cache cut is most of your bill. Swapping Fable 5 for 5.1 is a config change, not a rebuild, and the a16z line that agents burn roughly five times the tokens a human would is why a quarter off the total matters more than another benchmark point. The second change is quieter and, for regulated shops, bigger. Anthropic reports Claude Code users see about 60% fewer security-related interruptions per session. If your coding agent kept refusing to touch its own auth code, that tax just eased. The model can now flag vulnerabilities, though Anthropic drew the line at writing exploits. There is also a data story. New Enterprise Frontier Safeguards let business customers keep data on their own cloud infrastructure instead of Anthropic’s servers, while still letting Anthropic watch for misuse. That is the zero-data-retention path finance and health teams have asked for since Fable 5 required a 30-day window. It rolls out in phases starting later this fall, so confirm your date before you promise your compliance team anything. One caveat worth saying out loud: the cost and coding numbers are Anthropic’s own, measured on Anthropic’s workloads. Believe the mechanism, the cache-read cut is real and you can check it on the pricing page. Grade the percentages after your own bill comes in. Ship it? Deploy now if you already run Fable. The cache-read cut alone pays for the swap, and you can A/B it against 5 on a single workload in an afternoon. Sources: VentureBeat, TechCrunch, MarkTechPost (Sep 1, 2026).
Tour de Headlines🔒 OpenAI rated Astra “Critical” on cyber, then gated the good part. Astra is the first model OpenAI has pushed past the “Critical” line in its Preparedness Framework, meaning it can find and chain unknown flaws without a human holding its hand. It scored 100% on ExploitBench, and on an internal set of 20 high-severity bugs it discovered and used two zero-days in one exploit chain. The advanced cyber tier goes only to a vetted group OpenAI calls Daybreak; a broader release is coming “soon” with more guardrails and no firm date. The builder takeaway is boring and useful: if the strongest models start living behind defender-only tiers and safety halts, checkpoint your long agent runs so a mid-job gate fails gracefully instead of torching the work. (Sep 1.) 🎛️ Genesys shipped a governance layer for contact-center agents. At Xperience 2026 on Wednesday, Genesys unveiled four pieces of an agentic stack, and two of them, Contextual Intelligence and the AI Control Plane, are available now. The Control Plane centralizes discovery, identity, policy, and observability: where an agent can act, what it can touch, and when a human has to sign off. The flashier bits, Cloud Navigator and Cloud Orchestrator, are on a promise (Navigator by fiscal Q4, Orchestrator not until fiscal Q1 2028). So the fence is what you can deploy today. The autonomy is still a roadmap. For any CX or RevOps lead being told to “govern the agents,” this is a concrete shape to point at. (Sep 2.) 🔌 Tealium opened a governed MCP for its customer-data platform. The new Configuration MCP lets an agent change CDP settings in plain language from inside Tealium Studio, with the platform’s governance still in the loop, and Tealium shipped fresh APIs alongside it. This is the interesting drift in MCP right now: it started as “read my data” and it is turning into “change my platform settings, safely.” A server that exposes configuration to an agent without handing over the keys is the pattern more vendors will copy. (Aug 31.)
Sponsor Agents get clearance. Your reps get coaching. This week two labs argued over what an agent is allowed to do. RapportScore measures how your people communicate on real calls, then coaches the behavior that builds trust. Deterministic signals you can coach against. See where your team stands. See your team’s score → |
Tool of the Day🍳 PR Sous ChefA scheduled agent that only nudges the pull requests that stalled. This one is worth stealing for the shape it uses. PR Sous Chef (from GitHub’s Agentic Workflows) runs on a 15-minute cron plus a /souschef slash command, reads the state of every open pull request, and only nudges Copilot when a PR has gone quiet. Most cycles it finds nothing and moves on, so it keeps the review queue alive without spraying comments on PRs that are fine. That is the agent pattern that survives in production: read-only, scheduled, model invoked only when a rule trips, with a visible audit trail. Copy the structure into your own repo and point it at your slowest queue. Read the writeup
Worth a Click- Where agent money is landing: Axle AI raised a reported $17.5M (Series A, led by Base10) to build API-native agents for insurance back-office work: verifying coverage, issuing endorsements, reconciling claims. The unglamorous verticals keep being where deployable agent revenue shows up.
- Fork an agent instead of writing one: githubnext/agentics is the open sample pack behind PR Sous Chef and the rest of GitHub’s Agent of the Day series. Ready-made repo workflows you can point at your own project this afternoon.
Same week, same capability, opposite calls. Anthropic eased Fable’s cyber brakes so the model will help you find holes; OpenAI walled Astra’s cyber tier off because it finds them too well. “Can it hunt vulnerabilities” is no longer a question about how good the model is. It is a purchasing decision about which tier you are cleared for. Pick the lane that matches what you are willing to be responsible for. See you Monday.
— Ron |
