AI Policy & Feature Tracker

Updated 2026-07-11 · verified, client-relevant items · every item confirmed against a vendor's own publication or reported press coverage with a retained source snapshot · see the Memo tab above for the print view
Vendors
Themes

AI Platform Policy & Feature Digest

Prepared 2026-07-10 · covering the 7 days ending 2026-07-10 (a story linked to an older event still shows that event for context) · internal work product

← back to partial (current) edition

Inclusion standard. Every item below is flagged client-relevant, carries confidence of vendor-confirmed or reported (never unverified), has no open verification flag, has every cited source page captured on file as it read on the retrieval date, and falls within this edition's 7-day window (or is part of a story where a more recent member does). 21 additional verified items fall outside this 7-day window and remain in the full record; 4 items are still held back pending verification or source capture. The full record, including everything outside this window, is in the interactive dashboard.
Summaries are AI-drafted from cited sources and human-reviewed before publication.
Themes this period:
model-lifecycle (9)enterprise-trust-and-vendor-risk (5)export-controls-and-regional-restrictions (2)distillation-and-model-extraction (2)pricing-and-licensing (2)security-vulnerabilities (2)undisclosed-behavior-and-transparency (1)government-access-and-review (1)

Security, Trust & Incidents

Hidden tracking code found in Claude Code, present since April 2026 — 2 developments

Hidden tracking code found in Claude Code, present since April 2026
2026-07-01 · Anthropic · Claude Code · undisclosed behavior
Reported (press/researcher; vendor position noted in summary)

Reverse-engineer LegitMichel777 found undisclosed code in Claude Code — present since v2.1.91 (April 2, 2026), never mentioned in release notes — that checked whether the system timezone matched Chinese regions and compared connection proxies against a list of Chinese domains and AI-lab addresses, encoding results as near-invisible Unicode substitutions sent back to Anthropic. Anthropic called it an experiment launched in March to prevent account abuse by unauthorized resellers and protect against distillation, and said a removal PR merged July 1 before the story broke. Core technical finding is consistent across outlets; the 'why now' narrative and framing still vary by outlet and Anthropic has published no detailed writeup we could locate.

effective 2026-07-01 · China-linked users · Claude Code users
distillation-and-model-extractionenterprise-trust-and-vendor-riskexport-controls-and-regional-restrictionsundisclosed-behavior-and-transparency
Cited: [1] [2]
Alibaba bans internal use of Claude Code
2026-07-03 · Anthropic · Claude Code · regulatory action
Reported (press/researcher; vendor position noted in summary)

Following the tracking-code disclosure, Alibaba instructed employees to stop using Claude Code by July 10. Official Alibaba statement quoted by The Next Web: 'As Claude Code was recently discovered to carry back-door risks, after comprehensive evaluation, Claude Code has now been added to a list of high-risk software with security vulnerabilities.' A customer-side policy action, not an Anthropic change, but directly relevant to enterprise clients evaluating vendor trust.

effective 2026-07-10 · China (Alibaba internal) · Enterprise
enterprise-trust-and-vendor-riskexport-controls-and-regional-restrictions
Cited: [3]
  1. Ars Technica (via secondary coverage) (press, retrieved 2026-07-08)
  2. Slashdot (aggregating Ars Technica) (aggregator, retrieved 2026-07-08, stance: neutral)
  3. The Next Web (press, retrieved 2026-07-08)

Policy, Pricing & Terms

2026-07-08 · Harvey · Harvey · pricing change

Harvey co-founder warns of 14x token-burn growth and looming $10M AI consumption bills

Reported (press/researcher; vendor position noted in summary)

Harvey co-founder Gabe Pereyra said token usage across Harvey's platform grew 14x over six months, warning that customers will start seeing AI consumption bills reaching $10 million and asking what drove the cost -- citing Uber's CTO's comment about burning a year's coding-token budget in four months as an industry-wide parallel. Pereyra frames this as the start of a broader "pricing reckoning" as agentic systems consume tokens at rates traditional SaaS pricing wasn't built for. (A single 100,000-document contract review can reportedly cost around $20,000 in tokens alone.)

pricing-and-licensing
Sources (2) — page snapshots on file as retrieved
2026-07-06 · OpenAI · ChatGPT · pricing change

Workspace Agent and Excel/Sheets usage move to token-based credit pricing

Confirmed by vendor's own publication

ChatGPT Business, Enterprise, and Edu workspaces moved Workspace Agent runs and ChatGPT for Excel/Sheets from a free preview to token-based credit pricing, based on input, cached input, and output tokens rather than a flat credit cost. ChatGPT for PowerPoint remains free for these tiers until August 6, 2026.

effective 2026-07-06 · Business/Enterprise/Edu
pricing-and-licensing
Sources (1) — page snapshots on file as retrieved

Deprecations & Removals

2026-07-05 · Anthropic · Claude · model deprecation

Claude 3.5 Haiku deprecated on Google Cloud Vertex AI Model Garden

Confirmed by vendor's own publication

Google's Vertex AI partner-model deprecation notes list Claude 3.5 Haiku as deprecated effective January 5, 2026 with shutdown July 5, 2026 -- this is Google's own hosting schedule for the model on Vertex. Separately, on Anthropic's own API (and other Anthropic-operated platforms), the same model (claude-3-5-haiku-20241022) was deprecated December 19, 2025 and fully RETIRED February 19, 2026 -- already resolved, on an earlier and independent schedule from the Vertex one. Partner-hosted platforms set their own retirement dates per Anthropic's own documentation.

effective 2026-07-05 · Vertex AI Model Garden
model-lifecycle
Sources (2) — page snapshots on file as retrieved

New Features & Releases

Anthropic ships new features (4)

2026-07-09 · Anthropic · Claude · feature launch

Anthropic launches "Reflect" usage-reflection dashboard

Confirmed by vendor's own publication

Anthropic introduced Reflect in beta, a Claude usage-reflection dashboard in the style of "Spotify Wrapped." It summarizes a user's Claude activity over the past 1, 3, 6, or 12 months -- key topics, usage patterns, and task types -- using the 4D AI Fluency Framework, and periodically surfaces reflective questions such as "What's one thing you want to keep doing yourself, even if Claude could do it faster?" Users can set quiet hours or schedule a break nudge from within the dashboard. Available in beta for Free, Pro, and Max users with memory turned on; it does not draw from incognito chats or pull in underlying files from connected tools (only summaries Claude already produced).

model-lifecycle
Sources (2) — page snapshots on file as retrieved
2026-07-09 · Anthropic · Claude Code · feature launch

Claude Code v2.1.206 released

Confirmed by vendor's own publication

Anthropic shipped Claude Code v2.1.206, adding directory path suggestions to /cd (matching /add-dir behavior), a /doctor check that proposes trimming checked-in CLAUDE.md files of content Claude could derive from the codebase, confirmation prompts before entering a git worktree outside the project's .claude/worktrees/ directory, expanded git push remote auto-allowlisting, gateway login support for Anthropic-operated public gateway endpoints, and background agents upgrading in the background right after an update instead of on next attach. Also included fixes for expired logins, keyboard input responsiveness, MCP server configuration, and model picker display.

model-lifecycle
Sources (1) — page snapshots on file as retrieved
2026-07-08 · Anthropic · Claude Platform / API · feature launch

API and Admin key expiration dates added to Claude Console

Confirmed by vendor's own publication

Anthropic added the ability to set an expiration when creating an API key or Admin API key in the Claude Console -- a preset, a custom duration, or Never. For keys with a lifetime of at least 7 days, Anthropic emails the creator before expiration; existing keys are unaffected. The Admin API reports each key's expiration in its expires_at field.

security-vulnerabilities
Sources (1) — page snapshots on file as retrieved
2026-07-07 · Anthropic · Claude · feature launch

Claude Cowork expands from desktop to web and mobile

Reported (press/researcher; vendor position noted in summary)

Anthropic expanded Claude Cowork, its remote/background task-execution feature, from desktop-only to web and mobile (iOS and Android), starting as a beta rollout for Max subscribers with more plans to follow. Users can start a task at their desk, get status updates on their phone, and receive finished output later, including tasks that run in the background with no device online, scheduled for a specific time. Chat and Cowork now share one home tab on web and desktop. Anthropic extended doubled Cowork usage limits through August 5 to mark the rollout; usage data shared alongside the launch showed over 90% of Cowork sessions are non-coding work, led by business process/operations tasks.

model-lifecycle
Sources (2) — page snapshots on file as retrieved

Google ships new features (3)

2026-07-07 · Google · Gemini API / Vertex AI · feature launch

Gemini Distillation Service launched in Early Access

Reported (press/researcher; vendor position noted in summary)

Google Cloud's Gemini Enterprise Agent Platform release notes list a Gemini Distillation Service in Early Access (access by request), alongside Semantic Governance Policies (SGP), a runtime layer for evaluating an agent's proposed tool calls against organizational rules — SGP also appears throughout the platform's API reference. Both are notable for enterprise governance conversations. Exact dated entry not yet pinned: the release-notes page renders dynamically and could not be read directly from this environment; existence confirmed via search-indexed content of the official page.

Enterprise / Agent Platform
distillation-and-model-extractionenterprise-trust-and-vendor-risk
Sources (1) — page snapshots on file as retrieved
2026-07-07 · Google · feature launch

Gemini Enterprise adds egress FQDN and allowed-data-source restrictions for connectors

Confirmed by vendor's own publication

Google Cloud added two managed organization policy constraints for Gemini Enterprise data connectors, both GA: "Restrict allowed data sources" lets admins control which external data sources (such as Jira, Box, or Confluence) are permitted when adding a data store, and "Restrict allowed egress FQDNs" lets admins control which fully-qualified domain names a data store can connect to.

enterprise-trust-and-vendor-risksecurity-vulnerabilities
Sources (1) — page snapshots on file as retrieved
2026-07-06 · Google · feature launch

Developer logs support added for Interactions API in AI Studio

Confirmed by vendor's own publication

Google added developer logs support for the Interactions API: logs for supported Interactions API calls are now viewable directly from the Logs page in Google AI Studio, for real-time payload debugging.

model-lifecycle
Sources (1) — page snapshots on file as retrieved

Harvey ships new features (2)

2026-07-10 · Harvey · Harvey · feature launch

Harvey partners with Lumio on legal-AI economic impact framework

Confirmed by vendor's own publication

Harvey and Lumio (a firm specializing in legal market strategy, pricing, and commercial growth) partnered to build an economic framework helping law firms understand AI's financial impact beyond simple efficiency/productivity metrics -- covering how AI reshapes work division, talent development, client demand, pricing, growth, and profitability. Per Harvey's Tara Waters: "Together, we can help firms understand AI's economic impact and the strategic choices that follow."

model-lifecycle
Sources (1) — page snapshots on file as retrieved
2026-07-08 · Harvey · Harvey · feature launch

Cross-user Vault sharing on iOS

Confirmed by vendor's own publication

iOS users can share Vaults with other Harvey users directly from the mobile app, with granular permission levels (view/edit/full access) and controls over downloads/duplication — a new access-control and data-sharing surface.

enterprise-trust-and-vendor-risk
Sources (1) — page snapshots on file as retrieved

OpenAI ships new features (2)

2026-07-07 · OpenAI · ChatGPT · model release

GPT-5.6 lineup (Sol, Terra, Luna) expands to global preview access

Confirmed by vendor's own publication

OpenAI expanded preview access to the GPT-5.6 lineup — Sol (flagship, positioned for advanced coding and cybersecurity), Terra (balanced/lower-cost), Luna (fastest/cheapest) — with public availability announced for July 9. Pricing per 1M tokens: Sol $5 in/$30 out, Terra $2.50/$15, Luna $1/$6. The models initially went to ~20 organizations after OpenAI shared them with the US government under a pre-release review; the Commerce Department approved broad launch after additional testing. Press framing of the hold ('Trump administration pause') varies; OpenAI's own pages describe a staged preview.

effective 2026-07-09 · Global · API, Codex; ChatGPT general availability not yet dated
government-access-and-reviewmodel-lifecycle
Sources (5) — page snapshots on file as retrieved
2026-07-06 · OpenAI · ChatGPT · feature launch

ChatGPT for PowerPoint reaches General Availability

Reported (press/researcher; vendor position noted in summary)

OpenAI made ChatGPT for PowerPoint generally available for Business workspaces, and GA for Enterprise and Edu workspaces as well. The add-in runs inside Microsoft PowerPoint, letting teams create and revise editable presentations, ask questions about deck structure and narrative, and use Skills and connected apps to build slides from repeatable workflows. Workspace admins control app access. Per already-tracked event 6, PowerPoint remains free for Business and Enterprise customers through August 6, 2026, after which it follows the same token-based credit pricing as Excel/Sheets.

model-lifecycle
Sources (2) — page snapshots on file as retrieved
2026-07-07 · Legora · Legora aOS · feature launch

Ally Law network collaborates with Legora on member-firm AI adoption

Reported (press/researcher; vendor position noted in summary)

Ally Law, a global legal network of nearly 80 law firms across 50 countries (a Band 1 Chambers and Partners Leading Law Firm Network), announced a collaboration with Legora to help member firms evaluate and adopt AI tools for efficiency, client service, and long-term competitiveness. Since late 2025, roughly 25% of Ally Law's member firms have signed agreements with Legora and begun implementation.

model-lifecycle
Sources (1) — page snapshots on file as retrieved
Generated mechanically from the tracker database on 2026-07-10. Summaries are AI-drafted from the cited sources and human-reviewed before publication; where outlet framing conflicts, the conflict is stated rather than resolved. Source-page snapshots (as-retrieved copies) are retained in the tracker archive. This digest describes vendor conduct and published terms; it is not legal advice.