Indexed — Issue 004
Tuesday, July 28, 2026
Google shipped three new Gemini models this week and skipped the one everyone wanted. Below: what that means for your agent builds, a stack move for image work, and a prompt you can run in the next ten minutes.
The Signal
Google introduced Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber on July 21, all aimed at teams building AI agents at scale. The stated focus was efficiency, latency, and reliability — cheaper, faster models tuned for coding and production workloads, according to Google's official announcement.
What Google did not ship matters as much. Gemini 3.5 Pro was not part of this release; Google says it looks forward to releasing it soon.
Read the gap plainly. Flash-class models prioritize low cost and fast responses for production apps. Pro-class models are Google's high-capability tier for complex reasoning and coding. Google chose to optimize the cheap, fast layer while rivals keep raising the capability ceiling.
Your move today: if you run agent workflows, price out a Flash-tier swap for your highest-volume, lowest-stakes steps. Keep your reasoning-heavy steps on your current model. Cost per token drops on the parts that don't need frontier smarts. That's a same-day win you can test without rebuilding anything.
Before you wire anything into a billable workflow, confirm the model ID in your console. Google lists the production code for Gemini 3.6 Flash as gemini-3.6-flash.
The Stack
The Signal is about squeezing cost out of agent steps. The same discipline applies to how you assemble the rest of your workflow: match the tool to the job instead of running everything through one chat window.
A tight operator stack for this week:
Google Gemini for anything living inside Docs, Sheets, and Gmail, with live Search grounding when you need current answers.
Claude for long-document reasoning — contracts, research packets, and executive-ready rewrites where nuance survives.
Perplexity AI for cited first-pass briefs, so you can fact-check a claim before it goes in a deck.
Cursor or GitHub Copilot for multi-file edits and inline code, depending on whether you want a new editor or a plugin.
Midjourney for campaign visuals and moodboards, now on v7 with a full web app.
Pick two you already pay for and route real work through them this week. A five-tool stack you actually use beats a fifteen-tool stack you browse.
Prompt of the Day
Image work is where most operators waste tokens and time. The fix is a repeatable structure, not luck. Some of the best AI image prompts follow the same skeleton every time: subject, setting, lighting, camera or medium, and mood.
Use this prompt in Midjourney or your image tool of choice, and swap the bracketed nouns for your own:
Photorealistic product shot of a matte-black water bottle on a concrete surface, soft morning window light from the left, shallow depth of field, 50mm lens look, muted neutral palette, editorial minimalist mood, 3:2 aspect ratio.
Change one variable at a time — light, then lens, then palette — and save the versions that land. That variable-isolation habit is exactly what operators trade in AI Freedom Lab, where the current thread is on which Midjourney v7 prompt structures hold up across a full campaign.
Tool of the Day
ChatGPT (rating 4.7, freemium) is still the default general assistant, now powered by GPT-5 with an o3 reasoning model for multi-step problems.
Why it earns the slot today: it collapses a scattered workflow into one window. Drafting client emails, turning messy notes into an SOP, analyzing an uploaded spreadsheet, and generating image concepts via DALL-E all happen without tab-switching. For a solo operator, that consolidation is the value.
Test it on one real task today: upload your last messy meeting notes and ask it to return a structured action plan with owners and dates. If the output is deck-ready, you've replaced a 30-minute cleanup job.
It won't beat a specialist on every axis — Midjourney still wins on image craft, Claude on long-document nuance. But as the hub of your stack, ChatGPT is worth keeping open. Grab it via the ChatGPT Tool of the Day listing.
Quick Hits
Cisco open-sourced two small AI models called Antares (350M and 1B parameters) on Hugging Face, built to scan codebases for known vulnerabilities without shipping source code to an outside provider, per Axios. Access is gated to verified cyber defenders. Relevant if you run SecOps or engineering.
Moonshot AI's new Kimi K3 is priced at $15 per million output tokens, according to Tom's Hardware, while AP reports that demand forced a temporary pause on new subscriptions. Existing subscribers were prioritized while Moonshot added capacity.
OpenAI Presence is an enterprise product for running agents against internal data and workflows. It is available only to eligible enterprise customers through a limited general-availability program, so it is not a broad-market launch.
Closing the loop
The through-line this week: cost and fit beat frontier bragging rights. Google optimized its cheap tier, Cisco went small on purpose, and Moonshot competed on price. The operator lesson is the same — match the model to the step, then measure.
Run one test today. Price a Flash-tier swap on a high-volume agent step, or save three image variations from the prompt above. Small, measurable, done before lunch.
Try today's Tool of the Day: ChatGPT on AI Tools Index
Browse past issues in the Indexed archive