AI Model Updates

DeepSeek moved the DeepSeek-V4-Flash API from preview into official public beta on 31 July 2026, with significantly upgraded agent, coding and tool-use capabilities and benchmark results exceeding the earlier V4-Pro-Preview; V4-Pro and the consumer app/website models are unchanged.

Google released Gemini Robotics ER 2, a significant upgrade to its embodied-reasoning model for robotics, improving video understanding, task planning, and coordination between multiple robots.

Perplexity has renamed and upgraded Spaces to Projects, adding a persistent file system so its Computer feature edits the same files across sessions instead of starting over, plus a self-improving memory feature called Brain for long-running team work.

OpenAI cut API pricing for GPT-5.6 Luna by 80% and GPT-5.6 Terra by 20%, and introduced a Fast mode for GPT-5.6 Sol offering up to 2.5x faster responses at double the price, effective 30 July 2026.

Google's Gemini Spark personal agent can now use a person's own Chrome browser — including logged-in accounts and saved passwords — to complete web errands like scheduling appointments or researching flights, pausing for user confirmation on sensitive actions like payments. Google AI Pro access for Spark also expanded to 160+ additional countries.

xAI released Grok Voice Think Fast 2.0, its latest speech-to-speech voice model, with improved intelligence, transcription accuracy, and conversational responsiveness.

Google added natural-language voice capabilities to the Gemini app for macOS — holding the Fn key lets users dictate, ask questions, or generate content naturally within any application window.

Grok launched Build Mode, letting SuperGrok Heavy subscribers describe an app, game, or website in chat and get a working, publishable product with a live preview - no coding required. It combines Grok's Build coding agent, Imagine generation tools, and browser-use tooling into one guided experience across web, iOS, and Android.

xAI's Grok 4.5 model is now available inside GitHub Copilot, selectable from the model picker across VS Code and other GitHub products.

Anthropic's Claude Opus 5 is now available on Amazon Bedrock, described as matching Claude Fable 5's top-tier intelligence at Opus-tier pricing, with zero data retention (ZDR) enabled by default to help meet data-governance requirements.

Anthropic released Claude Opus 5, a new flagship model described as thoughtful and proactive, approaching the frontier intelligence of Claude Fable 5 at roughly half the price. It is state-of-the-art on coding and knowledge-work benchmarks (Frontier-Bench, GDPval-AA), priced the same as Opus 4.8, and is now the default model on Claude Max and the strongest available on Claude Pro.

xAI released a free Google Workspace add-on that puts Grok alongside open files in Sheets, Slides and Docs to assist with data analysis, presentation building and document writing.

Grok Build added Workflows, letting users describe a large task in plain language which Grok then plans and fans out across hundreds of parallel agents, saveable as reusable team commands.

OpenAI launched a health feature letting users 18 and over securely connect health records and Apple Health data to ChatGPT; OpenAI states this data is not used to train its foundation models.

Voice support was extended into the ChatGPT Work and Codex modes of the desktop app, adding natural interruption handling during voice sessions.

Anthropic launched a no-install connector in claude.ai letting users query the Anthropic Economic Index directly, e.g. asking which occupations or tasks use AI most.

Manus launched Plan Mode, an optional review step that has the agent produce a structured plan (goals, steps, constraints) before executing a task, so users can review and edit the approach before any code is written. It can be triggered at the start of a task or mid-task, on both web and mobile.

Grok 4.5 rolled out broadly to grok.com, X, and the iOS/Android apps, delivering clearer everyday answers and handling comprehensive tasks like building a spreadsheet or slide deck from an outline.

OpenAI launched Presence, an enterprise platform for deploying and managing production-grade AI voice and chat agents with built-in policies, guardrails and human escalation paths, rolling out via limited general availability to select enterprises including BBVA, SoftBank and IAG.

Google expanded Gemini Intelligence's task-automation abilities to more than 40 popular apps for everyday tasks like shopping and booking travel, and made Gemini Notebook (formerly NotebookLM) a preinstalled app on Samsung's new foldable devices for synthesising research into slide decks and podcasts.

Google released Gemini 3.6 Flash, an efficiency-focused workhorse model for agent development that uses about 17% fewer output tokens than 3.5 Flash while improving coding, ML research, computer-use and document-analysis performance, priced at $1.50/M input and $7.50/M output tokens. Google also released 3.5 Flash-Lite for high-throughput use and a 3.5 Flash Cyber variant for a limited government/partner cybersecurity pilot.

xAI launched a Grok add-in for Microsoft Outlook that summarises email threads, drafts contextual replies, and helps organise the inbox.

xAI shipped a free Grok add-in for Microsoft Excel that answers questions in plain English, writes formulas from a description, and runs scenario analysis directly in a workbook.

Google integrated Parallel Web Systems as a native web-grounding provider on the Gemini Enterprise Agent Platform, letting enterprise agents pull real-time web results with precise source citations.

xAI released Grok 4.5, its strongest model yet, purpose-built for coding, agentic tasks and knowledge work, trained partly alongside Cursor on real developer session data.

Grok added Automations, letting users set up recurring or trigger-based jobs that Grok runs independently using attached files, connectors and skills, saving results to a run history.

OpenAI updated the ChatGPT desktop app (macOS and Windows) with a global switcher between ChatGPT and Codex, a unified Recents view combining Chat and Work conversations with sort/filter/pin, and Projects support inside the desktop app, alongside syncing Work across devices.

Organisations can now submit agents built in Agent Builder to a "Built by your org" section of the Agent Store after admin review, creating a governed publishing workflow for internally built agents.

Agents built using the Model Context Protocol are now accessible directly inside Word, Excel, PowerPoint, Outlook, and Catalyst, extending MCP-based agent access across the core Office apps.

Microsoft shipped a batch of Copilot updates including submitting custom agents to the Agent Store from Agent Builder with admin review, a company-wide Prompt Gallery for publishing tenant-wide prompt collections, MCP-based agent access inside Word/Excel/PowerPoint/Outlook/Catalyst, brand kit generation from uploaded brand guidelines, and nested permissions support for Confluence and ServiceNow connectors.

Perplexity detailed SPACE, a new sandbox architecture built with Firecracker microVMs that isolates AI agents while they handle sensitive credentials and data, with rolling snapshots for time-travel recovery.

xAI released Grok Build's coding-agent/terminal codebase as open source on GitHub, letting developers inspect the implementation, build custom extensions, and self-host it.

ChatGPT's search was enhanced across web, iOS, and Android to search across chats, projects, and files with improved filtering.

Anthropic launched free premium Claude access for verified US K-12 educators, including standards-mapped lesson planning and class-data analysis, under FERPA-compliant terms.

Manus can now generate native, fully editable .pptx files with data-backed charts and real tables directly from spreadsheets, documents or meeting notes, currently in beta for 1.6/Max users.

ChatGPT access was restored on WhatsApp for users in the European Economic Area via a verified WhatsApp contact number.

OpenAI's new GPT-5.6 model family — Sol for advanced reasoning, Terra for balanced production work, and Luna for fast/affordable high-volume tasks — is now generally available on Amazon Bedrock. The models run on Bedrock's inference engine with prompt caching and hardware-enforced security, at pricing matching OpenAI's direct rates.

Perplexity launched Brain, a self-improving memory system for Computer that builds a private context graph from past sessions to improve answer correctness, alongside faster orchestrator model options and site publishing.

A new toggle automatically deploys a website's live URL every time a build succeeds, removing the need to manually confirm each publish while iterating.

OpenAI launched ChatGPT Work, a new agent for longer-running tasks like research, analysis, app integration, and document creation, and simultaneously retired group chats and the ChatGPT Atlas browser.

As part of the ChatGPT Work launch, OpenAI discontinued its standalone Atlas browser, folding its agentic browsing capabilities into the new unified ChatGPT desktop app instead.

Group chats began winding down from July 9, 2026: users could no longer create or join new group chats, existing ones became read-only, and users retained access to their conversation history.

ChatGPT Skills, reusable custom capabilities accessible under Plugins in the sidebar, became generally available for Enterprise and Edu users.

Mistral Studio added version control for prompts and skills, giving enterprises immutable version history, ownership tracking, audit logs, and controlled CI/CD-based deployment.

OpenAI shipped a redesigned desktop app for macOS and Windows that combines Chat, Work, and Codex modes in one application, available globally.

ChatGPT Sites entered public beta, letting users turn their work into interactive websites or apps, with Business and Enterprise customers able to publish them publicly.

Meta released an upgraded version of its Muse Spark reasoning model with a 1-million-token context window, stronger agentic tool-use and coding performance, and made it available through a new public-preview Meta Model API as well as a "Thinking" mode in the Meta AI app.

Users can now branch any point in a Manus conversation into a new parallel session that inherits all prior context, letting them explore multiple directions from the same starting point without losing the original thread.

OpenAI launched a new voice model, GPT-Live-1, that can listen and speak simultaneously, enabling more natural turn-taking and interruption in voice conversations.

Mistral released Robostral Navigate, an 8B-parameter vision-language model that lets robots navigate using a single RGB camera and natural-language instructions with no depth sensors or LiDAR.

Meta Superintelligence Labs launched its first media-generation models: Muse Image, an agentic image generator that edits with precision and composes from multiple references, and Muse Video, a video model with native audio support.

A new fallback model replaced GPT-5.3 Instant Mini, with improved tracking of evolving user intent during conversations.

ChatGPT for PowerPoint, a sidebar tool for drafting, revising, and analysing presentations directly inside Microsoft PowerPoint, became generally available across all tiers, with credit-based usage pricing.

xAI added 21 new multilingual voices to Grok Voice alongside the original five, each cast for specific use cases, accessible via the Voice Agent API, TTS API and Voice Agent Builder.

Mistral released Leanstral 1.5, an open-source formal-verification model with 119B total/6B active parameters, saturating the miniF2F benchmark and hitting state-of-the-art scores on FATE-H and FATE-X.

Copilot Chat in Outlook moved beyond single-email-thread help to reasoning across a user's entire inbox, calendar, meetings, and enterprise data to surface action items and manage schedules — and this expanded capability doesn't require a full Microsoft 365 Copilot licence.

Organisations can now add watermarks to AI-generated video and audio content produced by Copilot, controlled via a new admin policy setting.

Files Copilot generates now automatically inherit the highest sensitivity label detected in the source data used to create them, helping generated content stay compliant with an organisation's data-governance policies.

xAI launched a no-code Voice Agent Builder that lets businesses deploy production voice agents on Grok Voice, priced at $0.05 per minute of audio plus optional telephony costs.

Anthropic released Sonnet 5, described as its most agentic Sonnet model yet, with performance approaching Opus 4.8 at lower cost and introductory pricing rising later in the year.

After the US export-control order was lifted, Anthropic restored global access to Fable 5 with an improved safety classifier and announced a new jailbreak-severity framework built with Amazon, Microsoft and Google.

Anthropic launched Claude Science, a research workbench integrating lab tools, databases, and notebooks so scientists can run multi-step literature and lab analyses with auditable, reproducible outputs.

Google shipped Nano Banana 2 Lite, its fastest and cheapest image-generation model, and Gemini Omni Flash, a cost-efficient video generation and conversational-editing model, both in Google AI Studio and the Gemini API.

OpenAI began a limited preview of a new GPT-5.6 model family: Sol (flagship), Terra (balanced, half the cost of GPT-5.5), and Luna (fastest, lowest cost), initially restricted to trusted partners.

OpenAI's personal finance experience in ChatGPT expanded from a limited rollout to all Plus users and Android, alongside improved dictation across more languages.

GPT-4.5 was removed from ChatGPT; existing conversations that had been using it automatically continued on GPT-5.5.

Codex Remote, which lets users direct and approve long-running coding-agent tasks from their phone, became generally available on all ChatGPT plans, alongside a new DigitalOcean Droplet Workspace plugin.

Google launched study notebooks in the Gemini app, an adaptive learning space where students upload class materials to get a diagnostic quiz, personalised lessons, and a progress dashboard.

GPT-5.5 Instant received a conversational quality improvement aimed at decision-making, advice, planning, researching options, and shopping tasks.

Perplexity launched a legal-vertical version of Computer that connects to lawyers' research databases, document repositories and contract tools, using 20+ frontier models plus premium legal sources.

Manus Website Builder now offers a free-when-idle Autoscale hosting mode and an always-on Reserved mode, switchable with one click and no code changes.

Google made "computer use" a built-in capability of Gemini 3.5 Flash, letting developers build agents that see, reason, and act across browser, mobile, and desktop apps, with adversarial-training safeguards.

Mistral shipped connector governance features for enterprise customers, including workspace/org-level admin controls, scoped API keys, multi-account connectors, and GA support for connectors inside Vibe Code and Workflows.

Anthropic launched Claude Tag, a Slack-integrated always-on Claude teammate that can be tagged for tasks, learns from channel context, and proactively surfaces updates.

Mistral released OCR 4, a document-parsing model supporting 170 languages with bounding boxes, block classification and confidence scores, topping OlmOCRBench and available self-hosted or via API.

Grok Build added a /goal autonomous mode that runs long, unattended tasks to completion and verification without needing constant user back-and-forth.

Perplexity brought Deep Research into Computer so users can turn multi-step research into reports, spreadsheets, dashboards or websites, and added a command panel, thread forking, and per-member enterprise credit controls.

xAI shipped a free Microsoft Word add-in that turns prompts into structured documents, pulling in web/X research and connector data from email and cloud storage.

A new dedicated Scheduled page let users manage automated recurring tasks in ChatGPT, with improved reliability and notification handling.

PowerPoint's Copilot image generation gained a new FLUX.2 Flex model option, producing higher-quality visuals with improved text and layout rendering.

Users can now select Anthropic's Claude as a model option inside Copilot Chat for complex analysis, document understanding, and structured content generation, extending Copilot's model choice beyond OpenAI.

Google's June Pixel software update added Gemini Omni-powered conversational video creation and editing plus a Gemini music-generation tool, restricted to a Google AI subscription and users 18 and older.

xAI launched a free Microsoft 365 add-in that lets users build full slide decks from an outline, complete with web/X research, diagrams and images.

Grok Imagine Video 1.5 became generally available with better motion, physics and audio generation than prior versions, plus new productivity features like Projects and multiple agents.

Grok Build added an Agent Dashboard that consolidates multiple coding-agent sessions on one screen, letting users monitor progress and take over any session while others keep running.

OpenAI removed GPT-5.2 Instant, Thinking, and Pro from ChatGPT; existing conversations on those models automatically continued on the corresponding GPT-5.5 model.

The US government issued an export-control directive citing a possible jailbreak method, forcing Anthropic to abruptly disable Fable 5 and Mythos 5 worldwide for all customers.

xAI launched a plugin marketplace inside Grok Build so developers can browse, install and update plugins without leaving the terminal, launching with MongoDB, Vercel, Sentry, Cloudflare and Chrome DevTools.

Anthropic launched its first "Mythos-class" models: Fable 5, generally available with strong safety guardrails, and Mythos 5, restricted to trusted cybersecurity partners with fewer safeguards.

ChatGPT's memory system was upgraded to update automatically as conversations evolve, and a stricter "Lockdown Mode" privacy setting was made available to all logged-in users.

OpenAI extended its ads-in-free-tier model to UK Free and Go plan users, with paid plans remaining ad-free.

xAI released an image-to-video preview model that turns a single still image plus a motion description into a fluid, cinematic video clip up to 720p.

AWS added a new serverless way to fine-tune models on complex, multi-step agent tasks in SageMaker AI, training against a full sequence of decisions rather than single responses. Amazon Nova Lite 2.0 is one of the supported models, alongside Qwen and Gemma.

Copilot can now securely connect to supported third-party data sources in real time using the Model Context Protocol (MCP), with admins able to discover, review, and manage these connectors centrally from the Microsoft 365 admin center.

Word's default chat experience now lets Copilot directly edit the open document, with all changes fully reviewable and reversible; users can turn the behaviour off if they prefer.

xAI released Composer 2.5, a fast model built to excel at long-running tasks and complex instruction-following, available for free in Grok Build.

xAI released grok-build-0.1, a coding-focused model in public beta on the xAI API, trained for agentic coding tasks including web development, debugging and MCP support.

Mistral unified its coding and work agents under one product, Vibe, adding a Work Mode for cross-tool tasks, a dedicated Code Mode web interface tied to GitHub, and a VS Code extension, with new pricing tiers.

Anthropic released Claude Opus 4.8, an upgraded flagship model at the same price as Opus 4.7, with better coding honesty, adjustable effort levels, and a faster/cheaper fast mode.

Computer became directly usable inside Word, Excel, PowerPoint, Outlook and Teams for Pro/Max/Enterprise subscribers, and Perplexity added a live context panel plus usage-analytics dashboards.

Mistral released Search Toolkit in public preview, an open-source framework that bundles document ingestion, hybrid retrieval and built-in evaluation metrics into one pipeline for building production search and RAG systems.

Mistral introduced a "physics AI" product line that predicts physical behaviour from geometry and boundary conditions in seconds instead of hours, targeting product design, tooling and digital-twin use cases.

xAI launched Grok Build, a new coding-agent CLI for professional software engineering, available to SuperGrok and X Premium Plus subscribers, with approval workflows and parallel subagents.

Manus added a direct Shopify integration letting users build, manage inventory/collections, and run marketing campaigns for a live Shopify store entirely through chat.

Mistral released Medium 3.5, its first flagship "merged" model combining instruction-following, reasoning and coding, scoring 77.6% on SWE-Bench Verified, with open weights under a modified MIT licence.

Mistral added programmatic connector management to Studio, letting developers register MCP-based integrations once and reuse them across Le Chat and Studio, including human-in-the-loop approval for tool calls.

Perplexity open-sourced Bumblebee, an internal read-only security scanner that checks developer machines for vulnerable packages, risky AI agent (MCP) configurations, and browser/editor extensions without executing any code.

Google launched the Gemini 3.5 model family, led by Gemini 3.5 Flash, built for AI agents that can execute complex, multi-step tasks and run roughly four times faster than earlier frontier models.

Copilot Chat gained access to GPT-5.5 Thinking, a model built for multistep reasoning and structured analysis, enabling deeper reasoning on complex problems.

Microsoft added admin controls letting organisations selectively grant access to Anthropic models, with xAI's Grok models also becoming available in Copilot Studio for US customers via admin opt-in.

Google rolled out "Gemini Spark," a background personal agent that handles recurring tasks without being asked each time, alongside a redesigned interface, Gemini Omni video generation, and a personalised "Daily Brief."

Recurring automations can now run inside the same ongoing task conversation instead of spawning fresh tasks each time, with new calendar views and finer control over timing and connectors.

Grok introduced Skills, letting users teach the assistant preferences and workflows once and have them retained across future conversations, with pre-built and custom skills.

Google introduced Gemini Intelligence, a system-level layer that lets Android automate multi-step tasks, summarise and act on web content in Chrome, and build custom widgets from plain-language descriptions.

Perplexity's Personal Computer feature — which securely uses local files, apps and the Comet browser as context — rolled out to all Mac users, alongside a Personal CFO upgrade.

Manus Projects can now review past task conversations and propose updates to project instructions, files and reusable skills, with every change requiring explicit user approval before it takes effect.

Anthropic doubled Claude Code's five-hour rate limits for Pro, Max, Team and seat-based Enterprise plans and raised Opus API rate limits, backed by a new compute partnership with SpaceX.

Perplexity launched a Finance Search capability in its Agent API that combines licensed financial data, real-time pricing and cited web sources into a single API call.

xAI (via SpaceX) agreed to supply Anthropic with compute from its Colossus 1 supercomputer, with the two companies also exploring orbital compute infrastructure.

Grok launched Connectors, deep integrations that let the assistant read and act on data in SharePoint, Outlook, Google Workspace, Notion and GitHub directly, without manual copy-pasting.

xAI added a Quality Mode to the Grok Imagine image API for enterprise developers, producing higher-realism images with stronger text rendering and better prompt adherence.

Anthropic released ten ready-to-run agent templates for financial-services tasks that integrate with Cowork, Claude Code and Managed Agents, plus new Microsoft 365 add-ins.

Perplexity gave finance teams direct access to licensed data from providers like Morningstar, PitchBook, FactSet, Daloopa and Carbon Arc inside Computer.

Perplexity started surfacing clinical-grade medical information from sources like the New England Journal of Medicine, BMJ Journals and BMJ Best Practice directly in relevant health answers, with full citations.

GPT-5.5 became the default orchestration model for Computer on Pro and Max plans, OpenAI's GPT Image 2 replaced the prior image generator, and Perplexity added website publishing plus Microsoft Teams, Snowflake and Databricks integrations.

Manus launched an always-on persistent virtual machine so users can run bots, scripts and scheduled automations continuously without managing servers or writing code.

OpenAI added an optional security setting letting users require stronger sign-in methods, such as passkeys or compatible hardware security keys, for their accounts.

Perplexity expanded Computer's workplace integrations, giving business users a broader agentic assistant that plugs into existing work tools and files for research, drafting, and task execution.

xAI launched Custom Voices, letting users clone their own voice from about a minute of audio for use in Grok's Text-to-Speech and Voice Agent APIs, with two-stage verification to prevent unauthorized cloning.

Manus integrated OpenAI's GPT Image 2 model into its slide generator, giving presentations 4K photorealistic images and far more legible in-image text.

Mistral launched Workflows in public preview, an orchestration layer for running AI agents in production with durable state, full audit trails, and built-in human-approval steps.

DeepSeek released a preview of its V4 model ahead of general availability, alongside a deprecation notice for the older deepseek-chat and deepseek-reasoner model aliases.

OpenAI released GPT-5.5, its new flagship model, with major gains in agentic coding, computer use, and research tasks while matching the prior model's speed.

A new image generation model launched across all plans, adding a "thinking" capability for paid users to improve prompt understanding and output quality.

ChatGPT began showing ads to Free and Go plan users in these markets, while paid tiers remain ad-free, marking a new monetisation model for the free product.

A new lightweight fallback model replaced GPT-5 Instant Mini, described as more natural in conversation with stronger writing and contextual awareness.

OpenAI introduced a new $100/month Pro tier aimed at intensive Codex users, offering up to 10x more Codex usage than Plus, alongside the existing $200/month Pro plan.