Correction — June 1, 2026:

This report's Pattern 1 section originally asserted that Microsoft had not published Copilot's per-Credit cost, per-feature consumption rates, or hard cap mechanism. All three were published — the per-Credit rate (1 AI Credit = $0.01 USD) and per-token model breakdowns have been in GitHub Docs since April 27, 2026. The hard cap is a user-settable $0 overflow budget option. The original framing was empirically wrong.

The substantive consumer-protection story — that GitHub's published rates predict structural monthly overage for typical agentic Copilot workflows — is preserved and strengthened in the corrected Pattern 1 below. The retraction is logged on the OneHuman Record page.

In May 2026, three of the eight AI tools OneHuman tracks restructured their fundamental revenue or billing model — and in each case, users were informed after the fact. The industry ran a live experiment on whether AI subscribers would accept ads, variable billing meters, and 10x pricing jumps without meaningful consultation. The month is over. The answers aren't in yet.

Pattern 1: The Billing Structural Break

The three changes happened independently but landed in the same month. OpenAI launched ChatGPT Ads Manager on May 21 with a stated $2.5B year-one revenue target and a $100B-by-2029 roadmap. The ads match conversation history — meaning the model you use for private research is the same surface that serves you ads based on that research. No opt-out was announced at launch. The 800M+ users on free and Plus tiers were not asked.

GitHub Copilot's June 1 transition moves every paid subscriber from a flat subscription to AI Credits — a published consumption meter. The base prices stay the same ($10 Pro, $39 Pro+, $19 Business, $39 Enterprise). The per-Credit rate is $0.01 USD, published in GitHub Docs since April 27, 2026 alongside per-token model breakdowns. A $0 overflow budget option prevents surprise bills for users who set it. The disclosure is technically complete. The math is the consumer-protection story. A Pro subscriber receives 1,000 AI Credits per month. A typical agentic session — Workspace, Agent Mode, large-codebase refactoring with premium models — consumes 3,000-4,000 Credits. That is 25-33% of monthly allowance in a single session. Two sessions exhausts the allowance entirely. Pro+ subscribers exhaust their 3,900 Credits in approximately one agentic session per month. The base prices are floors, not ceilings — and the published rates predict that every meaningful agentic Copilot user is structurally over allowance every month.

xAI launched SuperGrok Heavy at $300/month — a 10x gap above the next tier ($30 SuperGrok) with nothing between them. That pricing arrived inside SpaceX's corporate structure, post the February 2026 merger at a combined $1.25 trillion valuation. The data-flow boundary between Grok consumer accounts and SpaceX's classified government operations has not been published. I have looked. It is not there. The introductory $99/month price holds for six months, then resets to $300.

Tools affected: ChatGPT, Copilot, Grok Consumer verdict: Set GitHub Copilot's $0 overflow budget option before the first agentic session; locate the ChatGPT ad settings; do not subscribe to SuperGrok Heavy until the data-flow boundary is published.

Pattern 2: AI Tools Are Connecting to Your Accounts — And Obscuring What You're On

Five separate data-access expansions landed in May across three tools. Each was framed as a feature. In aggregate, they represent a structural shift: AI tools are moving from "you bring the data" to "we stay connected to your accounts."

ChatGPT quietly added Gmail integration to its default model between May 5–8 — a silent rollout affecting 500M+ users, buried inside a model naming change. Three weeks later, OpenAI launched personal finance integration, allowing ChatGPT to connect to bank accounts and view a unified spending dashboard. That is one tool with standing access to your email, your financial accounts, and your conversation history — and an ad platform operating on top of all three.

Google Gemini's I/O 2026 relaunch requires always-on Gmail access for Daily Brief and Gemini Spark, its two highest-profile new features. Claude's "dreaming" feature — launched alongside the Microsoft Office add-ins — scans past sessions in the background by default, with no documented opt-out mechanism at launch.

The same ChatGPT model change that enabled Gmail access also introduced a naming problem worth flagging separately. OpenAI now runs three active models with near-identical names: GPT-5.5 (the premium API model launched April 23 at $5/M input tokens, with 82.7% agentic benchmark scores), GPT-5.5 Instant (the everyday default chatbot model used by everyone who opens chat.openai.com), and GPT-5.4 mini (the budget API option). They share naming prefixes. They are not the same product. Developers who used chat-latest to stay current were auto-migrated to GPT-5.5 Instant without notification. The naming structure makes it structurally difficult to compare what users on different tiers are actually receiving. That is not an accident — it is a transparency failure that belongs in the same category as silent data-access expansions.

I am not opposed to connected AI features. I am opposed to data access grants bundled into feature launches with no clear opt-out, no explicit consent step, and no published answer to "what happens to this data and who can access it" — and to naming conventions that obscure what model tier you are actually on. I checked each of these five expansions at the source. None of them published a clear opt-out path at launch. That is what I am documenting.

Tools affected: ChatGPT, Claude, Gemini Consumer verdict: Before enabling Gmail, bank account, background-session, or always-on features on any AI tool: check the data access comparison page, locate the opt-out, and read the data retention policy. Also verify which model you are actually on — the name on the dropdown may not match what you think you subscribed to.

Pattern 3: The Coding Tool Price Wedge Widened to 30x — And Split Into Two Categories

May 2026 created the widest price gap in AI coding assistance since the category launched. But the more important development is not the price range — it is the architectural divide underneath it.

OneHuman's Recommendation Engine was updated this month to split "coding" into two distinct categories: Code Assistance (debugging, autocomplete, code review; synchronous; runs in your IDE; human-in-the-loop at every step) and Agentic Coding (autonomous multi-file builds, long-running jobs, async cloud execution; human as reviewer, not operator). These were always different workflows. May 2026 made the infrastructure behind them diverge publicly.

OpenAI put Codex mobile on every ChatGPT plan, including free — competing squarely in the Code Assistance tier. Anthropic doubled Claude Code limits but kept the tool gated at $20/month minimum, anchoring in Agentic Coding. Mistral launched Medium 3.5 with 77.6% SWE-Bench Verified at $1.5/M input tokens — open weights, self-hostable on four GPUs, with async cloud execution via the Vibe CLI and session teleportation (move a live local coding session to cloud execution mid-task, preserving history and state). Grok Build launched at $300/month as Agentic Coding infrastructure only. Copilot's IDE-lock model, which broke under agentic workloads in April, transitions to metered billing June 1.

The AI coding market now spans $0 (Codex free, Code Assistance) through $10–$39 (Copilot, metered) and $20 (Claude Pro) to $300 (SuperGrok Heavy/Grok Build, Agentic Coding). The tools that bet on IDE lock-in are repricing under pressure. The tools betting on cloud-native agent infrastructure are shipping. Mistral's structural position remains unique: open weights, self-hostable, subscription dependency eliminable at the source — the only entry in the tracked set where you can exit the vendor billing relationship entirely. I tracked this divergence across seven tools this month. The infrastructure split is real and it is widening.

Tools affected: ChatGPT/Codex, Claude, Copilot, Grok, Mistral Consumer verdict: Identify which category your actual workflow falls into — Code Assistance or Agentic Coding — before choosing a tool. Model your June Copilot bill before it arrives. If self-hosting is viable, Mistral Medium 3.5 is the strongest open-weights coding option currently tracked.

Tool-by-Tool: What Changed This Month

Tool	Key Move	Impact	User Action
ChatGPT	Ads Manager launched ($2.5B year-one target, no opt-out); Gmail + bank account integrations; GPT-5.5/GPT-5.5 Instant naming confusion; Codex mobile free for all plans	High	Audit connected accounts; check ad settings; verify which model you're actually on
Claude	Limits doubled via SpaceX compute deal; Word/Excel/PowerPoint add-ins GA; "dreaming" session scan launched with no documented opt-out	High	Check dreaming opt-out before enabling; review Office add-in data terms
Gemini	I/O 2026 relaunch: Daily Brief, Spark, Omni Video; always-on Gmail required for key features; Spark standalone product positioning unclear	High	Read Gmail access scope before enabling Spark or Daily Brief; confirm which paywall tier covers your needed features
Copilot	June 1 transition to AI Credits at $0.01/Credit; per-token model rates published April 27; $0 overflow budget option available; typical agentic sessions consume 3,000-4,000 Credits (3-4x Pro monthly allowance)	High	Set $0 overflow budget in billing settings before first agentic session; evaluate Cursor or Claude Code if agentic use is frequent
Grok	SuperGrok Heavy $300/mo (10x gap above next tier); data-flow boundary with SpaceX undisclosed; $6.4B 2025 operating loss revealed in IPO filing	High	Do not subscribe to Heavy tier until data-flow boundary is published; $99 intro expires in 6 months
Perplexity	Personal Computer agent opened to all Mac Pro subscribers — local + cloud hybrid, auditable actions	Medium	Mac Pro users: test local agent for workflow automation; Pro plan required
Mistral	Medium 3.5: 77.6% SWE-Bench Verified, $1.5/M input, open weights, self-hostable on 4 GPUs, Vibe CLI async execution	Medium	Developers: evaluate as Copilot alternative; only tracked tool where subscription dependency is eliminable
DeepSeek	No significant May signals	—	No action required

The Watchdog Call

May 2026 was a live experiment: how much change can AI subscribers absorb in 31 days before they push back? Three billing restructures (Copilot disclosed but with structural overage math, ChatGPT ads with no opt-out, Grok at $300 with undisclosed data-flow boundary), five data-access expansions, a naming convention that obscures what model tier you are on, and a 30x price range in coding tools — all shipped as product announcements, none framed as what they actually were: fundamental changes to the consumer relationship. Three open consumer-protection questions remain unanswered at month-end: Grok's data-flow boundary with SpaceX, ChatGPT's ad opt-out mechanism, Gemini's always-on Gmail data terms. The fourth — Claude's dreaming opt-out path — is documentation-pending. The fifth original question — Copilot's per-credit rate — was already answered in GitHub Docs; OneHuman missed it, and the corrected analysis is now published with the math-driven story that should have run on May 27. What June brings: Copilot's first metered billing cycle closes mid-month, and the real-world overage bills will tell us whether the structural overage math holds in practice. If those bills are larger than expected, May's quiet developer backlash becomes a loud consumer-protection story. I will be watching.

Sources

OneHuman Breaking News: ChatGPT Ads Manager, May 25 2026
OneHuman Breaking News: GitHub Copilot Usage-Based Billing, May 27 2026
OneHuman Breaking News: Grok $300/Month Under SpaceX, May 30 2026
OneHuman: ChatGPT Five Changes in Four Days, May 8 2026
OneHuman: Claude Office Add-ins, Limits, and Dreaming, May 12 2026
OneHuman: OpenAI Codex Mobile Free Tier, May 16 2026
OneHuman: Google Gemini I/O 2026 Relaunch, May 20 2026
OneHuman: Mistral Medium 3.5 + Vibe Agentic CLI, May 4 2026
OneHuman Pricing Sweep Persistence Layer: entries 2026-05-01 through 2026-05-30
Cross-reference: AI Tool Pricing Comparison
Cross-reference: AI Tool Data Access Comparison
GitHub Docs — Copilot Billing Models and Pricing (docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing) — added in correction June 1, 2026
GitHub Blog — "Copilot is moving to usage-based billing" (April 27, 2026) — added in correction June 1, 2026

Verified by OneHuman · May 31, 2026 — corrected June 1, 2026

Independent. AI-assisted. Human-verified. No ads. No affiliates. No investors.

OneHuman Watchdog: May 2026 — Three Tools Changed How They Bill You