SPEAKER_00 0:00

Welcome back to Sumo Growth AI Insights. I'm walking you through this week's AI news the way an operator actually uses it, not as a pile of headlines. Lead story. Anthropic shipped Claude Fable 5.1 and Claude Mythos 5.1. Same underlying model, two safeguard regimes. Fable 5.1 is generally available across Claude, the API under Claude-Fable-5-1, Amazon Bedrock, Google Cloud, and Microsoft Foundry. Mythos 5.1 stays limited to vetted cybersecurity and life sciences trusted access programs. On the benchmarks, Anthropic published, 52.6% on Terminal Bench Science 0.1 versus Fable 5's 24.7, 55.8% on Terminal Bench 4.0 with Mythos at 60.9%, and 31.4% on Automation Bench. The pricing move that matters for agencies is the cash read cut. Cash reads drop 75% from $1 to 25 cents per million tokens. Anthropic says that works out to roughly 25% lower effective cost on typical workloads and up to about 45% on highly agentic ones. List prices stay $10 per million input and 50 per million output. They also added built-in watermarks with a detection API and private preview, enterprise frontier safeguards rolling out this fall, and fewer false safety blocks, about 60% fewer cyber false positives, and about 85% fewer benign biology fallbacks. Plus story, OpenAI said on September 1st that its upcoming model Astra is the first to meet the critical cybersecurity capability threshold under its preparedness framework. That means with the right tools and access, it can find previously unknown security flaws and develop ways to exploit them across many well-protected systems without a person guiding each step. OpenAI still plans to make Astra available soon, but the most advanced cyber capabilities start limited. Vetted testers first, then defensive expansion via daybreak blue. They also delayed parts of Astra's development while strengthening safeguards after the hugging face incident. Evaluations included perfect exploit bench scores and discovery of two zero days now being disclosed to maintainers. Three more headlines worth your attention. First, the Pentagon put ChatGPT MIL and Starshield AI's Groc for Government on Genai.mil alongside Google's Gemini. That portal already has about 1.7 million users across a roughly 3 million person military and civilian workforce with IL5 access for controlled unclassified information. Anthropic's Claude is still absent from the portal amid the supply chain fight. Second, the Wall Street Journal reported Anthropic signed a roughly $35 billion cloud deal with NVIDIA-backed Lambda for a HUT-8 developed campus in Nueces County, Texas. Nvidia supplies the chips and holds the lease. That lands on top of last week's roughly $45 billion N-scale West Virginia capacity buy. Third, Google is preparing Gemini 3.8 Flash, internally called Skimaki, for release as soon as today. The journal says employees preferred it to Anthropic's Opus in head-to-head tests inside Google's Jet Ski Coding Tool. Another bid to close Gemini's coding gap versus Anthropic and OpenAI. Tools of the day. Browser at tribrowser.com puts technical content on autopilot. Connect a repo and it auto-generates self-healing docs, guides, and blogs. Monad at Monid.ai is open router for agent tools. One key to discover, run, and pay for 1800 plus APIs at runtime. Loqua at theloqua.ai turns natural speech and what is on your screen into polished writing and voice-driven workflows. Business AI tip of the day price the agent loop, not just the model name. Fable 5.1's headline is not another clawed number. It is a 75% cheaper cash read that only pays off if your go-high level and agency workflows keep long context warm across many tool calls. Astra's Critical Cyber Label is the other side of the same coin. Frontier models are getting powerful enough that clients will ask who can call what, with which logs, and under which kill switch. This week, before you reprice retainers or swap every client onto Fable, rerun your three heaviest agent jobs, lead research, content regen, and support triage on Fable 5.1 and measure cash hit rate plus total dollars per successful run. Add a one-line model and tool wallet note in every SOP. Which frontier model, which tool router Monid style, and who approves cyber sensitive prompts. For any client that touches security, finance, or government adjacent data, write the Astral rule in plain English. No unattended exploit style probing, dual provider failover, exportable prompts. And if you sell DevRel or product content retainers, trial browser on one GitHub repo this week so docs heal on merge instead of rotting after every ship. A cheaper agent that still burns cold context on every turn is not a margin win. It is the same invoice with a new model string. Pro tip screenshot one before and after cost and cash hit chart from a real client workflow and send it with your Fable Mythos update email. Agencies that show the math close upgrades. Agencies that only paste Anthropics blog get ignored. That is your Sumo Growth AI Insights briefing. Run the numbers on the agent loop this week, and I will catch you on the next issue.