Claude Fable 5.1 and Mythos 5.1
Anthropic shipped its next model generation on 2026-09-01: Fable 5.1, a broadly-available general-purpose model, and Mythos 5.1, the same underlying model with reduced safeguards for vetted cybersecurity and life-sciences professionals. Anthropic's own benchmarks show large jumps over the 5-generation — Fable 5.1 hit 52.6% on Terminal-Bench-Science 0.1 versus Fable 5's 24.7%, and Mythos 5.1 hit 60.9% on Terminal-Bench 4.0 versus Mythos 5's 42.0%. Standard pricing is unchanged ($10/$50 per million input/output tokens), but cache-read costs dropped 75% to $0.25/million, cutting typical workload costs ~25% and highly agentic workloads up to 45%. Mythos 5.1 also posted striking domain results: protein binders with 10x the affinity of competition winners and a 50% hit rate (versus a typical 10-15%), a Venus elevation map improved from 10-20km to 2-3km resolution, and up to 2.5x speedups on seven deep-learning models cutting genome analysis costs 30-60%.
Sources & depth
Path to Astra: critical capabilities and frontier safeguards
OpenAI says its upcoming Astra model is the first it has designated at the Preparedness Framework's Critical cybersecurity threshold — able, with the right tools and access, to find and exploit previously unknown vulnerabilities across well-protected systems largely without a human guiding each step. In evaluation, Astra scored 100% on ExploitBench, and on a fresh internal benchmark of 20 recently-disclosed high-severity V8 vulnerabilities it discovered and used two genuine zero-days as part of an exploit chain (now being disclosed to maintainers). Expert red-teamers had it build a full browser-sandbox-escape chain to host command execution and a separate local-privilege-escalation chain to root on a hardened OS. OpenAI paired the capability jump with new safeguards — Astra refuses 91.5% of cyber-jailbreak attempts versus GPT-5.6 Sol's 59%, and in an "honeypot" alignment test it made zero attempts to access out-of-scope systems where Sol did so 56% of the time. Advanced cybersecurity access will roll out first to a small group of alpha testers, then expand through OpenAI's Daybreak Blue program.
Sources & depth
Anthropic launches Enterprise Frontier Safeguards
A same-day companion to the Fable 5.1 launch: Anthropic is rolling out "Enterprise Frontier Safeguards," a service that detects model misuse (fraud, cyberattacks, credential theft) across sessions while keeping the underlying activity logs in the customer's own cloud account (AWS/Azure/GCP) under their own encryption keys — not on Anthropic's infrastructure, and without requiring Anthropic human review. It's aimed at the tension Anthropic says came up in conversations with 100+ customers in financial services, healthcare, and other regulated industries: they want misuse monitoring but can't accept the cross-session data retention that used to require. Rollout starts in phases this fall; Anthropic charges nothing beyond customers' standard cloud storage costs.
Sources & depth
Introducing agentic video understanding with Gemini
Google gave Gemini's video processing an agentic mode: instead of scanning every video at a fixed frame rate, the model now actively decides what to watch, at what speed, and through which modality — enabling sub-second moment retrieval for editing, cheaper long-form analysis, better anomaly detection, and more accurate object/action counting over time. Google reports up to 66% lower analysis cost, up to 88% less token consumption, and up to 7% higher accuracy versus static processing, with Gemini 3.7 Flash posting the best quality-to-cost combination among the three supported models (3.7 Flash, 3.6 Flash, 3.5 Flash-Lite). It's live now via the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform at standard pricing, with a rollout to the Gemini consumer app and YouTube's "Ask YouTube" feature coming in the next few months.
Sources & depth