This observation covers verified technical changes formally released around the period from 8:20 am on 2 September to 8:20 am on 3 September 2026 in Australia/Sydney. Two items are included.
Google releases Gemini 3.8 Flash and restricts access to Flash Cyber
Google released Gemini 3.8 Flash and Gemini 3.8 Flash Cyber on 2 September. The general model is now available through the Gemini API, Google AI Studio, Android Studio, Antigravity, Stitch and Gemini Enterprise, as well as to Google AI Pro and Ultra subscribers. Its introductory per-token price matches 3.7 Flash at US$0.75 per million input tokens and US$3.75 per million output tokens. Google also says 3.8 uses additional reasoning steps, tool calls and output tokens on complex tasks, so the total cost of an individual task can exceed that of 3.7. Gemini 3.7 Flash remains supported for workloads that prioritise token efficiency.
Google reports a 54.9% result on HLE-Verified and improvements over 3.7 Flash on software-engineering and professional-agent evaluations. These are developer-reported results, not complete independent reproductions. The Verge independently confirmed the release, access channels and pricing, and cited an early Artificial Analysis measurement in which per-task cost was about 40% higher, primarily because of greater output-token use and more agent turns. That early measurement does not establish a uniform cost difference across all workloads.
Flash Cyber uses the same foundational intelligence but different safety conditions. It is available only through the new Fairwind Program to reviewed government authorities, critical-infrastructure operators and software maintainers. Google reports a success rate above 70% on an internal vulnerability-discovery evaluation spanning 20 programming languages, a 47.2% pass@1 result on CWE-Bench, and results from Chrome patching and an internal Wiz evaluation. These figures come from Google and its partners and have not been fully reproduced publicly. The general Flash model includes safeguards for CBRN and cyber-offence misuse. The Cyber variant uses more permissive cybersecurity mitigations and is therefore not generally available.
Meta releases Muse Spark 1.3 and begins API rollout
Meta released Muse Spark 1.3 on 2 September and began rolling it out in Muse Code and the Meta Model API. Compared with 1.2, the update focuses on how coding and long-running agents operate: it can handle several workflows in a single long thread, ask questions when instructions are ambiguous, request user help when blocked, and confirm before taking consequential actions. Previously available reasoning modes are live. The “max reasoning” mode still requires additional safety testing and has not yet been released.
In internal comparisons by Meta engineers, 1.3 used about 20% fewer tool calls and 25% fewer tokens than 1.2. These are developer evaluations without a published cross-platform independent reproduction. Meta also reports stronger resistance to prompt injection and adversarial inputs, but the announcement does not provide complete test data for independent review. Axios confirmed the 2 September release, the Muse Code and API rollout, and pricing unchanged from the previous model. Muse Spark 1.3 is currently a proprietary service. Meta lists open weights only as a future plan and did not release model weights with this update.
Sources
Google, “Introducing Gemini 3.8 Flash and 3.8 Flash Cyber”, 2 September 2026: https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/
The Verge, “Google says its new Gemini 3.8 Flash model ‘works harder’ but might cost more”, 2 September 2026: https://www.theverge.com/ai-artificial-intelligence/988742/google-gemini-3-8-flash
Meta AI Research, “Introducing Muse Spark 1.3”, 2 September 2026: https://research.meta.ai/blog/introducing-muse-spark-1-3
Axios, “Meta debuts Muse Spark 1.3 as personal agent work continues”, 2 September 2026: https://www.axios.com/2026/09/02/meta-debuts-muse-spark-13-as-personal-agent-work-continues
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.