AI changes

Authorized administrators can review changes made through Codex Policies & Configurations in the UI or retrieve them through the Compliance API.

Administrators can cap newly created keys' lifetimes at organization or project level, making rotation a concrete production task.

The new MultiVectorEncoder brings ColBERT-style late interaction into the main library, widening support for token-level text and visual search while increasing index and migration complexity.

The new experience automatically routes identified 13-to-17-year-olds into teen-specific learning tools and protections, while independent evidence about outcomes remains absent.

The August 18 cutoff removes Deep Research from the consumer Copilot app while preserving saved reports and steering Microsoft 365 Premium users to Researcher.

Customers can choose European or US processing, request a preview service tier and run GLM-5.2 on Mistral infrastructure, while residency and capacity claims remain vendor-defined.

The agent-focused model keeps its existing API name while adding native Responses API support, new reasoning controls and a revised time-based pricing schedule.

Unity AI Gateway can choose a model for each coding task, but the beta relies on first-turn classification, restricted integrations and vendor-reported savings.

Eligible Claude Enterprise reviewers can now retrieve Claude Code and Cowork conversations, but coverage excludes ZDR, HIPAA-ready tenants, some cloud surfaces and device activity that never reached the API.

Consumer and Edu users can answer quiz questions inside a chat, but OpenAI has not published assessment-quality or learning-outcome evidence.

Eligible unshared projects can now move between default and project-only memory, while shared projects stay isolated and ChatGPT Work remains unavailable in project-only mode.

Eligible web users can browse Drive content, add files or folders to Chat and Work, and keep Google documents beside a conversation, while Shared Drives and mobile remain outside the first rollout.

The timeline records selected interaction events rather than screenshots or audio, while access, administrator approval and regional exclusions limit the initial rollout.

The new desktop build supports five current Ubuntu, Debian and Fedora releases worldwide, with browser actions enabled but control of other Linux applications still unavailable.

The compact open-weight model keeps document layouts and fine detail at native resolution, but its capability evidence and deployment readiness remain first-party and incomplete.

The 3.1-billion-parameter model adds screen understanding, multi-image input, grounding and function calling, while its speed and benchmark claims remain first-party results.

Earth-observation teams can generate geospatial vectors for a chosen place and time, then take the resulting raster into their own analysis tools.

The August 14 cutover leaves existing runs working but moves new custom-model projects toward a public-preview serverless GPU environment.

The August expansion adds partner actions across meetings, travel, entertainment, music and services, with access split by market, account and Gemini mode.

Developers get more control over long agent runs through subagent tracking, queued commands, headless plan-to-implementation flow and context-preserving app handoff.

The new reasoning-model option reaches five paid plan tiers and eight coding surfaces, with gradual availability, an administrator gate and usage-based billing.

Developers can choose review depth per pull request, while organizations set inherited defaults and account for different resource use.

SL2T 1.0 reaches Pixel 11 with explicit limits against high-stakes use or replacing qualified interpreters.

The builder guide adds deterministic cache breakpoints and same-engine routing hints for repeated agent context, while leaving real-world savings to workload testing.

The new model option spans eight coding environments, but access is gradual, organization administrators must enable a preview policy, and numeric 3.7 pricing is not yet shown in the linked reference.

LangChain says enterprise customers can keep sensitive agent data in their own AWS accounts while it manages the LangSmith platform lifecycle.

Python and TypeScript teams can move a code-first agent into a hosted LangSmith runtime that supplies durable execution, sandboxes, memory, identity and evaluation plumbing.

One package can now carry skills and MCP server configuration across VS Code, Copilot CLI, the SDK and Copilot app, with enterprise controls applying across supported clients.

All ChatGPT plans can surface available tables and refine choices in conversation, with partner coverage varying by country and Work excluded.

Approved security teams can now use Daybreak Red and Blue inside AWS environments through Bedrock, adding a governed cloud route for defensive and authorized testing workflows.

Copilot customers can now trace input, output and cached tokens behind AI-credit consumption instead of seeing only the final credit charge.

Microsoft's smaller coding model is rolling out across Copilot with a lower stated list price, while teams have until September 10 to migrate from its predecessor.

The IDE plugin can carry selected context across agent chats, connect to Ollama as a bring-your-own-key provider, expose Codex sessions in debug logs and accept tighter enterprise controls.

The advertising pilot is now live in the United Kingdom, Mexico, Brazil, Japan and South Korea, extending a model that keeps paid plans ad-free and gives entry-tier users personalization controls.

Developers can route supported open-weight language models to Baseten from Hugging Face model pages and client libraries, using either provider credentials or Hugging Face billing.

The unified endpoint now leads Google's model-and-agent developer stack, adding managed sandboxes, background jobs and optional server-side state while the older API remains supported.

The August 7 change opens more everyday health, educational and clinical queries while continuing to reroute dual-use research requests to Opus 5.

The cloud-hosted browser runs in V8 isolates and is designed around AI-agent automation, offering developers a lighter alternative for some browser-based tasks.

The open-weight speech model adds Arabic, Korean and Brazilian Portuguese voices, plus a production NIM for teams deploying voice agents on their own infrastructure.

The Apache-2.0 model targets always-on agents on consumer hardware, combining tool use, long-horizon reasoning and interleaved text-and-image input.

The 30-billion-parameter open model targets specialised agent tasks, while an open routing library directs requests across mixed-model systems.

Repository teams can launch documentation, error-investigation and follow-up workflows from issue or pull-request comments.

Paid-plan users can choose how much supported models reason for each delegated task, trading potential quality gains against token and credit use.

The new packages give students and educators guided workflows that can use approved course materials, documents, calendars and connected apps.

The integration lets customers connect bank accounts, make payments and verify finances without leaving a Sierra agent conversation.

Supported voice output can now carry an embedded watermark that OpenAI's public tool and API are designed to detect.

The 276-billion-parameter mixture-of-experts model activates 12 billion parameters per token and ships for self-hosting, fine-tuning, and multimodal use.

The in-chat builder turns prompts into interactive projects that can be revised in conversation and published to a link.

The re-post-trained API model keeps the preview architecture and endpoint while adding native Responses API support and a sharper agent focus.

The July update gathers recaps from the previous 30 days and adds agent workflows for follow-ups, inspections and in-person meeting notes.

The public beta centralizes spend caps, traffic limits, model fallbacks and sensitive-data handling across production agents.

The security product traces sensitive information across employees, agents, applications, databases and endpoints, with policy alerts and investigation context.

Forge developers can now call hosted Claude models through one SDK while keeping application data inside Atlassian's cloud boundary.

The open-weight 230M and 350M models target classification, routing and extraction across inputs as long as 8,192 tokens.

The new controls can stop inference requests when a shared budget is exhausted, extending cost governance beyond individual API keys.

The speech-to-speech model is priced at $0.08 per audio minute, with existing latest-alias integrations scheduled to move automatically.

The update adds text-only generation, native 1080p output and reference controls across consumer surfaces and the xAI API.

The U.S. rollout lets Spark use signed-in browser sessions while returning payments and other sensitive steps to the user.

The new policy mode lets administrators keep a company-wide baseline while granting additional models to selected enterprise teams.

The retirement reaches every Copilot surface, while enterprise administrators may need to enable the suggested replacement models.

The July Gemini Drop also adds macOS voice controls, Dropbox, Zillow and Viator links, avatar reuse and personalized images.

Terra falls to $2/$12 per million tokens, Luna to $0.20/$1.20, while Sol gains a premium low-latency API path.

Administrators can require organization SSO, disable remote control or limit hosting through managed configuration channels.

The July update adds an SDK-based agent, opt-in .NET and Azure expertise, selected-code review and organization-wide instructions.

The July editor bundle makes concurrent Copilot, Claude and Codex sessions easier to isolate, inspect and review.

The music model update targets melody, lyrics, vocals, tempo and duration, but Google has not published pricing, API access or independent quality tests.

The account-linking beta begins with Airtable, GitLab, HubSpot, Notion, Supabase and Vercel while keeping plugin permissions as a separate approval step.

Copilot reviews can now use repository skills and read-only MCP context across paid individual and enterprise plans.

A new global policy will make eligible generally available models available automatically from August 26 unless administrators opt out.

The enterprise release combines an AI tutor, course-authoring tools, learning administration, and Workday workforce context in one product.

The managed platform can route certain high-risk workflows back to Opus 4.8, making model selection conditional rather than absolute.

The agentic security system divides discovery, investigation and remediation across red-, blue- and green-team agents while keeping human control in the workflow.

Developers can start a Copilot coding-agent investigation from a failed mobile check, then review the proposed repair in a separate pull request.

The server drops initialization sessions and Redis traffic while adding a lighter request path, revised elicitation and conformance testing.

The latest plugin release brings custom tools into Claude agent flows while adding observability, token-limit and model-management settings.

Administrators can now govern app access separately, extend managed settings into app and cloud-agent sessions, and attribute usage to individual users.

Google is positioning two new Gemini Flash models around lower-cost, lower-latency agent workloads, while keeping its cyber-focused variant in a limited-access pilot.

The Associated Press now permits more assistive uses of AI but keeps reporting, sourcing, editorial judgment and verification with its journalists.

The final retirement removes the service from existing customers as well as new ones, leaving active projects with a short window to move model access or absorb failures.

GitHub is rolling Anthropic's newest Opus model across its editors, command line, cloud agent, mobile apps and web experience—with usage billed at the provider's API list price.

Powered by Muse Spark 1.1, Meta's assistant is moving beyond one-off answers into scheduled briefings, app-connected planning and work that continues without repeated prompting.

The email client shuts down on September 22, but Notion's email agents will remain. Users have until September 21 to export drafts, scheduled messages and workflow settings.

OpenAI is turning voice from a conversational feature into a control surface for starting, steering and checking delegated agent work.

OpenAI’s new health experience can ground answers in personal records and wearable data, making its privacy controls as important as its medical reasoning.

New approvals, confidence ratings and rationales give repository teams a clearer way to supervise automated issue changes without mistaking review for security.

With Gemini 3.5 Pro still missing, Google competed on price and efficiency — cutting 3.6 Flash output pricing to $7.50 per million tokens and restricting a new security model to governments and certified partners.

From July 20, frontier access becomes a plan differentiator: Max and Team Premium include Fable 5 at half of weekly limits, while Pro and Team Standard pay per use with a one-time $100 transition credit.

Anthropic's introductory $2/$10 API rate ends August 31. The same text can also consume up to 1.35 times as many tokens as it did with Sonnet 4.6, so buyers need to measure real workloads.

Demand for Kimi K3 pushed Moonshot's infrastructure to its limits within two days of launch; new signups are halted while existing subscribers are prioritized, with reopening planned in batches.

The rename keeps every notebook, link and Audio Overview intact — but as of publication it rests on a single detailed source, with confirmation from Google's own channels still pending.

SpaceXAI started European availability on July 16 after EU AI Act evaluations — but channel-by-channel checks show Cursor live while the developer API console still flags EU access as pending.

The strongest open-weight-class model yet shipped July 16 as a hosted service, with full weights committed for July 27 and no license published — buyers should treat 'open' as a dated promise, not an artifact.

Mira Murati's lab shipped a 975B MoE with full open weights at launch — no gating, no wait — while candidly conceding it is 'not the strongest overall model available today.'

Enterprise admins can now configure HIPAA-ready workspaces and execute a business associate agreement directly from the admin console — removing a procurement bottleneck for healthcare pilots.

The July 14–21 rollout moves Claude from read-only grounding into M365's center of gravity — and forces enterprises to re-review connector permission scopes.

The July 9–13 wave takes Anthropic's knowledge-work agent beyond desktop — still research-preview tier, pitched against Microsoft Copilot Cowork and OpenAI's agent mode.

Genkit Agents unifies conversations, tool execution, deployment and app state — with human-approval checkpoints for sensitive actions — as the July agent-framework race intensifies.

Mistral confirmed a new sparse-MoE open-weight family with partner-only July early access and a broader release expected 'later in the summer' — though no specs, benchmarks or firm date have been published.

Claude's memory now stores individual categorized entries instead of a daily summary, and new Time-and-focus settings make Anthropic the first frontier lab to ship explicit wellbeing controls.

Microsoft's July 2026 Copilot rollout pairs new GA capabilities with a governance-relevant change: OpenAI models, including GPT-5.6, become a named subprocessor admins can disable.

With GPT-5.6's July 9 general availability, the flagship model becomes a paid-tier privilege across ChatGPT's seven plans — while ads arrive on the free and Go tiers.

OpenAI's new flagship became Copilot's default-preferred model the same day it went GA — even as Microsoft quietly routes some Excel and Outlook prompts to its own MAI models.

AWS's newsroom confirms all three GPT-5.6 tiers went GA on Bedrock July 9 — day one, not weeks later — via the Responses API on Bedrock's next-generation inference engine.

OpenAI shipped its flagship with High cyber/bio classifications and 700,000 GPU-hours of red-teaming — while its independent evaluator couldn't produce a trustworthy capability number for Sol.

The Breeze Prospecting Agent finds persona-matching net-new contacts and drafts outreach; documentation indicates Professional/Enterprise gating, qualifying 'GA for all paid customers' claims.

Advertisers using Meta's generative tools or C2PA-tagged third-party AI content now get automatic disclosure labels — landing two days after Meta's own AI-likeness controversy and weeks before EU AI Act transparency duties apply.

Meta Superintelligence Labs' closed-weight agent model and $1.25/$4.25 API mark the company's definitive pivot from open-weight goodwill to metered AI revenue.

Message Center notice MC1319216 gives organizers a live toggle for Copilot, Facilitator and recap — but the July 9 update pushes Targeted Release to mid-August and GA to late August 2026.

The July 9 consolidation merges Chat, Work and Codex into a single ChatGPT desktop app — renaming the old client 'ChatGPT Classic' and scheduling the Atlas browser's retirement for August 9.

OpenAI's July 9 launch resets its model taxonomy and mid-tier pricing, with Sol at $5/$30 per million tokens and Terra undercutting GPT-5.5 at half the price.

Seedream 5.0 Pro generates editable multi-layer images and infographics with native text in 14 languages, targeting enterprise design workflows — though third-party tests note 2–3 minute generation times.

A US agent company built its flagship coding model on a Chinese open-weight base — disclosing it up front — and matched near-frontier performance at a fraction of per-task cost.

OpenAI's full-duplex voice models make interaction decisions many times per second — and ship with an unusual candor note: small disclosed safety regressions versus the mode they replace.

The multi-agent software-development system adds orchestration across the SDLC and dashboards for measuring agent productivity — IBM's bid for the highest-ROI agent use case in consulting pipelines.

The 8B vision-language-action model steers robots from one RGB camera and plain-language commands, reporting 76.6% on simulated R2R-CE — but ships neither weights nor API at launch.

Hand-written Metal kernels for sparse attention, native safetensors loading and a memory-saving fused cross-entropy loss headline a release aimed partly at local model development on Macs.

Microsoft's official lifecycle doc confirms the July 7 realtime-voice GA pair — and a hard July 23 retirement for the October 2025 mini build that forces an in-window migration.

The open-source robotics toolkit now standardizes the full predict-evaluate-correct-retrain loop, landing the same week a wave of open robot foundation models hit Hugging Face.

Final Token Preference Optimization retrains only the token that starts a repetition loop — cutting doom-loop rates from 22.9% to 1% on one test model with a one-GPU training run.

Meta's July 7 image model shipped with default-on AI remixing of public Instagram accounts; after backlash from SAG-AFTRA and CAA, Meta pulled the feature July 10 saying it 'missed the mark.'

Bloomberg reports tens of thousands of weekly prompts in Microsoft's flagship apps now run on in-house models — a small but strategic shift away from paying OpenAI and Anthropic per prompt.

The 3B/8B/14B open model family can change decoding modes at inference time — and its biggest training lift came from combining the two objectives everyone treats as rivals.

A 30B/3B-active hybrid Mamba-Transformer MoE handles ASR, translation, TTS, audio generation and speech-to-speech while preserving its text backbone's reasoning — but under a non-commercial license.

The 295B MoE topped Hugging Face trending in its launch week and was free on OpenRouter through July 21 — though hands-on reviews find its coding below the frontier and its headline claims vendor-reported.

Weights and inference code shipped under MIT on July 5 for what Meituan calls the first trillion-parameter model trained and served end-to-end on non-NVIDIA silicon — with export-control implications in both directions.

Anthropic's restoration statement argues the capability behind the export ban was already freely available in unrestricted models worldwide — undercutting the rationale for model-specific controls.

The July 2 release is the first major vendor cost-governance response to enterprise AI billing blowouts — though audit-grade logging still needs the separate Compliance API.

The consumer Sora app actually closed April 26; the July-accurate story is the Videos API countdown, replacement-model migration guides, and an economics autopsy built on analyst estimates.

Agent registry, governance and Defender-integrated runtime protection for AI agents convert from preview freebie to budgeted line item on July 1 — an explicit 'guardrails tax.'

The Commerce Department's June 30 reversal ended an 18-day suspension and set the template for government gating of frontier AI — with Anthropic adding a new cyber classifier as a condition of return.

GitHub made Moonshot's open-weight coding model generally available in the Copilot picker July 1, hosted on Azure with no traffic to Moonshot — a normalization milestone with enterprise admin gates attached.

List prices rise roughly 5–10% across E3/E5 and Business tiers effective July 1, justified by bundled Copilot and security value — the first suite-wide reset of the Copilot era.

Agents consuming Copilot Credits are now metered with monthly true-up and pay-as-you-go requirements once included capacity is exhausted — per admin notices, not a single launch announcement.

Effective July 1, the standalone Copilot Pro individual plan effectively ceases to exist, simplifying Microsoft's consumer AI tiers around Microsoft 365 bundles.

Agent definitions move from point-and-click configuration to versionable, testable code — a governance-relevant shift for enterprise SDLCs.

The consumption-currency model replaces flat AI add-on SKUs at cloud renewal — a consequential shift given SAP's installed-base leverage.

Announced June 17 with rollout continuing through July, the new Slackbot turns Slack into a surface where third-party agents are discoverable and invocable.

Cognition's hard July 1 deadline retired the Cascade agent, forcing remaining Windsurf automations onto Devin Local — a rare hard-date deprecation in the AI IDE market.