AI Product Updates, July 21: Gemini 3.6 Flash, Mistral in Foundry, and New AI Tools (2026)

The short version

Google released three Flash models on July 21: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber. Microsoft and Mistral expanded their enterprise partnership, putting Mistral Medium 3.5 and OCR 4 in Microsoft Foundry and Medium 3.5 in Copilot Studio. GitHub began rolling Gemini 3.6 Flash into Copilot, while Google Cloud started a three-month deprecation window for 16 open-model MaaS endpoints. 1 2 3 4

GrowthBook, Algolia, Google Shopping, and Handoff also shipped product-level changes for experimentation, commerce discovery, consumer shopping, and construction estimating. The practical split today is access: some releases are live now, some are gradual rollouts, and the Google Cloud endpoints now need a migration plan.

Product	Releasing entity	Release date	Core change	Availability
Gemini 3.6 Flash, 3.5 Flash-Lite, 3.5 Flash Cyber	Google	July 21, 2026	New Flash family: general agentic and coding work, lower-cost lightweight work, and cybersecurity. 1	Gemini 3.6 Flash is listed as stable in the Gemini API and available to try in AI Studio. The source does not give a plan or endpoint matrix for the two 3.5 variants. 5
Mistral Medium 3.5 and OCR 4 in Microsoft Foundry	Microsoft and Mistral	July 21, 2026	Expanded partnership adds both models to Foundry; Medium 3.5 also reaches Copilot Studio, alongside expanded European GPU capacity. 2	Available globally in Microsoft Foundry and Medium 3.5 is available in Copilot Studio. Azure and Azure Local support cloud, cloud-connected, and fully disconnected deployments. Pricing is not disclosed in the announcement.
Gemini 3.6 Flash in GitHub Copilot	GitHub	July 21, 2026	Adds Google's model to Copilot's model picker for coding and longer-horizon agentic work, with configurable reasoning and parallel tool use. 3	Gradual rollout to Pro, Pro+, Max, Business, and Enterprise across eight Copilot surfaces. Business and Enterprise admins must enable the Gemini 3.6 Flash Preview policy. Usage-based billing uses provider list pricing.
Open-model MaaS endpoints	Google Cloud	July 21, 2026	Deprecates 16 managed endpoints spanning DeepSeek, GLM, GPT OSS, Kimi, Llama, MiniMax, E5, and Qwen families. 4	Existing endpoints remain functional during the deprecation period, but new use may be restricted. Retirement is October 21, 2026; Google recommends self-deploying replacements through Model Garden.
AI Visual Editor	GrowthBook	July 21, 2026	Plain-language or WYSIWYG edits can change copy, styles, layouts, images, and Figma imports, then launch experiments or multi-arm bandits. 6	Available now as a Chrome extension connected to a GrowthBook account. The announcement does not state a new price or plan requirement.
Dynamic Facets	Algolia	July 21, 2026	Automatically prioritizes filters from search context and shopper behavior instead of showing the same fixed order for every query. 7	Available now to Algolia customers. Algolia reports 10% to 15% filter clickthrough improvement in its strongest-performing markets; that is vendor-reported early usage, not an independent benchmark.
Try On and Lens shopping features in New Zealand	Google New Zealand	July 21, 2026	Google Shopping Try On uses one uploaded photo to preview clothing; Lens helps identify products such as handbags. 8	The announcement is for New Zealand shoppers. It does not establish a global rollout or give a paid-plan requirement.
H1 blueprint takeoff model	Handoff AI	July 21, 2026	Upload a blueprint PDF and receive an end-to-end material takeoff. Handoff reports an 81.6% score on 15 residential blueprint sets, versus about 55% for eight major AI models. 9	Available now to Scale Plan subscribers and API customers. The benchmark is Handoff's own evaluation; pricing is not stated in the release.

Google releases three Flash models

Google's July 21 release adds three different bets under the Flash name. Gemini 3.6 Flash is the broad model for code generation, agentic execution, and spatial reasoning. Gemini 3.5 Flash-Lite targets cheaper, faster work such as managing AI agents. Gemini 3.5 Flash Cyber is tuned to find and patch software vulnerabilities at a lower cost than larger models, according to Google. The company still has not broadly released the higher-end Gemini 3.5 Pro described in the same report. 1

Gemini 3.6 Flash has the clearest access path. Google's developer page lists the stable model ID gemini-3.6-flash, a 1,048,576-token input limit, a 65,536-token output limit, and support for caching, code execution, function calling, search grounding, structured outputs, batch inference, and priority inference. Computer use is available in preview. 5

For developers, that means the 3.6 model is ready to test in AI Studio and through the API. For the two 3.5 variants, the release report confirms the product roles but does not give a comparable endpoint or plan table. Do not assume that the three models share the same access route.

Microsoft turns Mistral into a deployment choice

Microsoft and Mistral expanded their partnership around both models and infrastructure. Mistral Medium 3.5 and OCR 4 are now in Microsoft Foundry, and Medium 3.5 is also in Copilot Studio. OCR 4 is positioned for structured document processing and agent workflows; Medium 3.5 is the open-weight model Microsoft says customers can build, customize, and deploy through Foundry. 2

The deployment detail matters more than the partnership headline. Microsoft says the same tools, APIs, and workflows can span Microsoft Foundry in the cloud and Foundry Local on Azure Local. The supported operating choices are cloud, cloud-connected, and fully disconnected environments. Mistral is also adding Europe-based GPU capacity under a multibillion-dollar agreement; the announcement says the expansion uses thousands of NVIDIA Vera Rubin GPUs. Neither the release nor its linked material gives a new customer price.

GitHub adds a second route to Gemini 3.6 Flash

GitHub's Copilot rollout gives the same Google model a separate developer-facing access path. Gemini 3.6 Flash is being added gradually to the model picker in Visual Studio Code, Visual Studio, Copilot CLI, the Copilot cloud agent, the Copilot app, JetBrains, Xcode, and Eclipse. GitHub lists Pro, Pro+, Max, Business, and Enterprise as eligible plans. 3

Business and Enterprise administrators must enable the Gemini 3.6 Flash Preview policy before members can select it. GitHub says the model is billed at provider list pricing under usage-based billing, but the announcement does not give a numeric rate or a previous-versus-current comparison. Teams should verify both the policy setting and the billing model before treating the rollout as included plan capacity.

Google Cloud starts the retirement clock on 16 open endpoints

Google Cloud marked 16 open-model endpoints for deprecation on July 21, with retirement scheduled for October 21. The list covers managed DeepSeek, GLM, GPT OSS, Kimi, Llama, MiniMax, multilingual E5, and Qwen endpoints. During the deprecation period, existing workloads can continue to function, but Google says new use may be restricted and the endpoints will receive no new features or updates. After retirement, requests to those model IDs will fail. 4

The migration path in the documentation is self-deployment through Model Garden. That changes the operational burden: teams keeping one of these model IDs in production now need to choose a self-deployed alternative, a different managed endpoint, or a code and evaluation migration before the October deadline. This is a deprecation notice, not a model-quality announcement, so existing clients should treat the date as an engineering deadline.

GrowthBook removes the engineering ticket from visual experiments

GrowthBook's AI Visual Editor turns a plain-language prompt or WYSIWYG edit into a live website variation. Users can change copy, styles, spacing, layouts, and images, import a Figma frame, and then launch an experiment or multi-arm bandit from the same workflow. The editor is available as a Chrome extension connected to a GrowthBook account. 6

The boundary is also clear. Visual changes can validate a landing-page idea, but application logic, backend behavior, authentication, and pricing calculations still require engineering. The release does not publish a new price, so teams need to check their existing GrowthBook plan before planning a wider rollout.

Algolia makes filters respond to the query

Algolia's Dynamic Facets changes the order of product filters from a fixed catalog configuration to one based on search context and shopper behavior. A television query can prioritize screen size, while a skincare query can surface ingredients or skin type. Algolia says the capability is available now to customers and reports 10% to 15% higher filter clickthrough in its strongest-performing markets. That performance number comes from Algolia's early usage, not an independent study. 7

The product change is operational as much as it is user-facing: merchandising teams can inspect which filter values receive interaction and use that data for quick-filter chips. Pricing and plan packaging are not stated in the announcement.

Google adds AI shopping help for New Zealand users

Google's New Zealand team introduced a shopping flow built around Search, Lens, and Try On. The virtual try-on feature uses one uploaded photo to show how an item may drape across different body types, while Lens can identify an item such as a handbag and help find it online. The post frames this as a New Zealand shopping feature and does not promise a global launch. 8

For product teams, the update is a useful reminder that AI shopping features are moving into the search and catalog layer rather than living only in a chatbot. For users outside New Zealand, market availability remains unconfirmed by this announcement.

Handoff launches a domain-specific estimating model

Handoff H1 reads construction blueprint PDFs and produces a complete material takeoff. Handoff says H1 scored 81.6% on its Construction Blueprint Takeoff Benchmark, which used 15 residential blueprint sets scored against expert-validated quantities. The same release says eight major AI models clustered around 55%, while an experienced human estimator took a week compared with H1's two hours. These are vendor-reported benchmark results, and the blueprint sets and ground truth are available by research request. 9

H1 is available now to Scale Plan subscribers and API customers. The API path matters for contractors and software vendors that want takeoff generation inside an existing workflow; the release does not state an API price.

What teams should check today

Developers can test gemini-3.6-flash in AI Studio or the Gemini API, then compare that access path with the separate Copilot rollout before changing model routing.
Copilot Business and Enterprise administrators should enable the Gemini 3.6 Flash Preview policy and confirm usage-based billing before opening it to a team.
Google Cloud users should inventory the 16 deprecated MaaS model IDs and set a migration owner ahead of October 21, 2026.
Enterprise buyers evaluating Mistral should decide whether Foundry, Copilot Studio, Azure Local, or a disconnected deployment is the right operating surface; the announcement leaves pricing to the existing commercial path.
Growth, commerce, and construction teams can test the specialized products now where their account or regional eligibility permits, but should treat the published performance numbers as vendor claims until independent evaluations are available.

AI Product Updates, July 21: Gemini 3.6 Flash, Mistral in Foundry, and New AI Tools