Weekly AI Tools Roundup: August 11, 2026

What's New This Week

Google: Gemini 3.5 Flash family expands with 3 new models

Google launched three new Gemini 3.5 Flash variants targeting different agentic workloads — all rolling out July 20-21 with immediate API availability:

• Gemini 3.6 Flash — the general-purpose successor. 17% fewer output tokens than 3.5 Flash, DeepSWE coding score of 49% vs 37% for 3.5 Flash, input pricing of $1.50/M and output at $7.50/M.
• Gemini 3.5 Flash-Lite — built for high-throughput, latency-sensitive tasks. Delivers 350 tokens per second at $0.30/M input. The cheapest frontier-tier pricing in the market right now — beats Grok 4.5 consumer pricing for bulk structured extraction.
• Gemini 3.5 Flash Cyber — fine-tuned for vulnerability discovery and patching. Found 55 unique confirmed issues in Chrome's V8 engine in benchmarked head-to-head against Claude Opus 4.6 (47 found). Available exclusively to governments and trusted partners via the CodeMender agent platform.

For developers and teams: The 3.6 Flash improvement on DeepSWE coding is the practical takeaway — a 12-point jump in agentic coding is a generational step. At $1.50/$7.50 pricing, it's a direct competitor to Claude Sonnet 5's August pricing ($2/$10) and beats Grok 4.5 ($2.49/$6) on a price-to-capability basis. Flash-Lite is the new budget option for batch API calls. Both models are live in Google AI Studio now.

OpenAI: GPT-Red makes GPT-5.6 6× more injection-resistant

OpenAI released GPT-Red, an automated red-teaming model that uses self-play reinforcement learning to find and exploit prompt injection vulnerabilities — then adversarially trains production models to close those gaps. Result: GPT-5.6 Sol recorded 6× fewer failures on OpenAI's hardest prompt injection benchmark versus models from four months prior.

In head-to-head testing, GPT-Red outscored human red-teamers 84% to 13% on novel injection scenarios. It also demonstrated a live exploit: manipulating an office vending-machine agent to alter prices and cancel orders — proving that AI agents handling real transactions are vulnerable to injection if not hardened.

For teams deploying AI agents with tool access: Prompt injection isn't theoretical anymore. GPT-Red exists specifically because agents are being exploited. If your workflow routes user input into tool calls, database queries, or financial transactions, you need input validation and scope-aware permissions regardless of which model you use. Consider this your signal to audit agent safety before you need it.

xAI: Grok Outlook add-in summarizes emails and drafts in your voice

Grok for Microsoft Outlook is now available to all paid X and SuperGrok subscribers via the Microsoft Marketplace. It also works inside Word, Excel, and PowerPoint. Key features:

• Summarizes long email threads
• Drafts replies that match the user's writing style
• Organizes inboxes: archives finished conversations, flags needs-reply
• Reads attachments and searches the web and X without leaving Outlook

For teams drowning in email: This is the most practical AI deployment inside an existing app we've seen — no new tab, no context switch. The reply-style-matching feature is the real unlock. If Grok can write in your actual voice and you don't have to manually copy-paste replies, the time saved compounds daily. Compare this against similar features in Microsoft Copilot and Google Gemini Workspace side by side.

OpenAI: ChatGPT for Small Businesses program goes live

OpenAI launched ChatGPT for Small Businesses — a training and integration program for small business owners who want practical AI workflows. Components:

• Virtual training webinars and in-person AI academies across the US
• New guides for marketing, accounting, and ecommerce workflows
• Partner integrations: Dropbox, Shopify, Intuit, Slack, Wix
• Powered by ChatGPT Work — the agentic mode built on GPT-5.6 for multi-step tasks

Past attendees report: 78% built a functional AI workflow in a single day, 42% saved over five hours weekly.

For small business owners: This is the first time OpenAI has packaged its mid-market AI strategy with real integrations rather than just API access. The Dropbox connection alone — AI summaries and extraction from your file storage without manual upload — is compelling. If you run a small team and haven't evaluated ChatGPT Work yet, this program is worth registering for even just for the guides.

NVIDIA: Cosmos 3 Edge brings physical AI to local devices

At SIGGRAPH 2026 (July 20), NVIDIA unveiled Cosmos 3 Edge — a billion-parameter open world model built to run in real time on local devices for physical AI applications (robotics, autonomous systems, drones). Key announcements:

• Cosmos 3 Edge: optimised for edge inference without cloud dependency
• Synthetic Video Detector NIM microservice: identifies AI-generated video frame by frame — built for newsrooms verifying content authenticity
• Model Context Protocol integrations with Adobe, Blender, Unreal Engine, and SideFX Houdini — AI agents working inside creative tools while keeping human artists in control

For creators and developers: The MCP integration with creative tools is the most immediately useful piece. If you use Blender, Unreal, or Adobe tools, expect AI agent plugins to appear in those apps in Q4 2026 that can suggest edits, generate assets, and automate repetitive tasks — without leaving the tool. The synthetic video detector is a reminder that authenticity verification is becoming a commercial product category, not just a research concern.

Jack Dorsey: BUZZ — open-source AI-native group chat

Jack Dorsey launched BUZZ, an open-source, AI-native group chat platform designed to compete with Slack and GitHub. Positioned as:

• Model-agnostic — works with any LLM backend
• Decentralized and self-sovereign — teams can self-host
• Built for people and AI agents — first-class agent participant support

The community-first approach means developers are invited to build alongside the project from day one. No enterprise pricing, no vendor lock-in.

For teams evaluating Slack alternatives: If AI agents alongside your team is part of your workflow, BUZZ is worth a look. Early-stage open-source projects don't have the polish of Slack, but the model-agnostic approach means you can switch models without switching tools — a meaningful advantage if you're multi-model routing already. Watch for community plugin ecosystem growth over the next 3–6 months.

Honourable Mentions

OpenAI: Codex Micro hardware ships at $230

Codex Micro (made with keyboard maker Work Louder) is a $230 physical control pad for OpenAI's Codex coding platform. Features: 13 mechanical switches, a joystick, dial, and touch sensor. Six frosted keys show live Codex agent thread status via color. All controls map through the ChatGPT desktop app.

Quantities are limited. Most useful for developer power-users who want physical shortcuts for agentic coding workflows. Not essential, but a signal that OpenAI is exploring hardware alongside its software stack.

Sakana AI: Fugu-Cyber scores 86.9% on CyberGym

Sakana AI released Fugu-Cyber, a multi-agent orchestration model built for enterprise cybersecurity. It scores 86.9% on the CyberGym benchmark (vulnerability analysis in complex codebases) and 72.1% on CTI-REALM (threat intelligence to detection rules). Available at sakana.ai/fugu by application review. Updated Acceptable Usage Policy prohibits offensive use.

This is the clearest signal yet that agentic AI is moving into the SOC (Security Operations Centre). If you're evaluating AI tools for security teams, benchmark Fugu-Cyber against Microsoft Project Perception and OpenAI's GPT-5.6 Cyber variants.

OpenAI: ChatGPT Small Businesses training underway nationally

The ChatGPT for Small Businesses program is running virtual AI academies across the US, with Dropbox, Shopify, Intuit, Slack, and Wix integrations already live. If you've been waiting for a structured way to bring AI workflows to your team without hiring a consultant, this is the on-ramp.

Why This Matters for Creators

• AI inside existing productivity stacks is the winning pattern. Grok in Outlook, ChatGPT Work in business tools, AI inside Adobe via MCP — the tools that survive are the ones you don't have to switch to. Context switching is the silent killer of AI adoption on teams.
• Security hardening is now a competitive selling point. GPT-Red, Fugu-Cyber, Fable 5's new safety classifier — every major lab is now advertising safety and injection resistance. For teams evaluating AI tools for regulated industries, this is becoming a table-stakes requirement, not a nice-to-have.
• Physical AI (robotics + edge) is shipping, not just promising. Cosmos 3 Edge running billion-parameter models on local hardware means AI inference is leaving the cloud for embedded systems. For creators in gaming, film VFX, and simulation, this changes the latency ceiling for real-time AI interactions.
• Agentic coding platforms are the new IDE layer. Codex Micro hardware, Codex cloud agents from Gitpod acquisition, Cursor Sand in development — every major player is racing to own how developers write code with AI. Choose your editor based on ecosystem trajectory, not today's feature list.
• Open-source alternatives are genuinely viable now. BUZZ, Kimi K3 open weights, Nemotron 3 Embed free to self-host, Groq hosting open models at LPU speed — the gap between proprietary and open AI is closing on capability. For budget-conscious teams, this is the year to benchmark open alternatives seriously.

What to Watch Next

• Google Gemini 3.5 Flash-Lite at $0.30/M input — the cheapest frontier-class pricing yet. Benchmarks against Claude Sonnet 5 $2 intro pricing on batch tasks will determine if this changes developer routing defaults.
• GPT-Red adversarial training rollout — OpenAI has not confirmed whether GPT-Red hardening will be applied to GPT-5.5 or release models beyond Sol. Watch for an announcement on broader hardening.
• BUZZ plugin ecosystem — open-source projects live or die on plugins. If BUZZ gets MCP support and first-class agent integrations within 3 months, Slack should be worried.
• Codex Micro adoption — limited launch, limited quantities. If it sells out fast, expect a Codex hardware line expansion. If it doesn't, OpenAI shelves the hardware bet.
• Sakana Fugu-Cyber enterprise uptake — application required; if SOC adoption accelerates, expect OpenAI and Anthropic to announce competing enterprise cybersecurity products.
• OpenAI Codex cloud workspace features — Gitpod integration will surface in Codex over the next 4–6 weeks; watch for "Codex Cloud" launch announcements.

StigStack reviews AI tools independently. Some links may be affiliate links — we only recommend tools we've personally evaluated. Last updated August 11, 2026.