Included or Metered? The New AI Pricing Divide

The choice between AI subscriptions and usage credits stopped being academic the week of July 1, 2026. Anthropic announced that Claude Fable 5 would move to metered usage credits after July 7, while OpenAI — in the same window — publicly defended keeping Codex included on every ChatGPT tier with overage credits left optional. Two labs, opposite bets, one question landing on every budget owner's desk: do you pay a flat fee, or pay by the token?

What is at stake is not which model is smartest — it is which billing shape you can plan around. A subscription is a fixed line item your finance team already understands. Usage credits are a variable, consumption-driven cost that behaves more like a cloud bill: cheap when idle, alarming when an agent loop runs hot. Get the match wrong and you either leave capability on the table or discover a five-figure overage after the fact.

This guide separates the two stories the press covered separately — Anthropic's metering move and OpenAI's model preview — and reads them as one strategic split. Then it hands you the part nobody else published: a tier-by-tier included-vs-metered matrix across OpenAI, Anthropic, and Google, and a per-workload framework for deciding which meter to standardize on. Every figure below is dated and sourced; where a number is vendor-stated or secondary-sourced, it is marked as such.

Most coverage treated these as two unrelated stories: an Anthropic pricing change and an OpenAI model preview. Read side by side, they are the same argument about how AI should be billed in the second half of 2026. Anthropic is pushing its most capable coding model behind a meter. OpenAI is holding the line that its coding agent stays inside the subscription, with credits as an optional pressure-release valve — not a toll gate.

Neither position is charity. Anthropic's meter follows a clear commercial logic (and, as we cover below, a pattern that started with its enterprise seats). OpenAI's included stance is a land-grab: keep the subscription simple, keep adoption frictionless, and let the heaviest users buy their way past the caps. The two cards below frame the bet each lab is making.

Anthropic redeployed Fable 5 globally on July 1, 2026, and attached a billing change to the return. Through July 7, Fable 5 is included on Pro, Max, and Team plans for up to 50% of each plan's weekly usage limits. After that date, Fable 5 access requires enabling usage credits — and if credits are not enabled, Fable 5 stops working entirely for that organization. Premium Enterprise seats get the same grace window; standard Enterprise seats get none, with all Fable 5 usage billed through credits from day one.

Anthropic has not published a credit-to-tasks conversion, so the only concrete number available is the direct API rate: $10 per million input tokens and $50 per million output tokens (as surfaced in community usage-credit guides on July 1). Treat any specific dollars-per-task claim you see elsewhere as invented until Anthropic publishes one.

This is not a one-off. It is at least the second billing move Anthropic has aimed at programmatic and agentic usage inside roughly three weeks. On June 15, Anthropic proposed a separate Agent SDK credit — sized to each plan's fee, billed at API rates, per-user and non-pooled — before shelving it ahead of the effective date. We covered that reversal in detail in Anthropic's shelved June 15 Agent SDK credit split. The through-line is unmistakable: Anthropic keeps testing where it can put a meter on automated consumption.

The enterprise side moved first. Per The Register, Anthropic began renewing enterprise contracts onto a usage-based plan from late 2025; by early 2026 the standard became a single seat around $20 per employee per month with the old bundled-token discounts removed. As IntuitionLabs CEO Adrien Laurent told The Register, for some of his firm's clients the base seat was already only about 20 percent of the total bill, with the other 80 percent metered API usage — so for heavy users, the meter was already the reality. Metering Fable 5 for consumers simply extends a shape enterprises had been living with for months.

OpenAI's countermove is documentation, not a launch. Its help center states plainly that Codex is included across every ChatGPT plan — Free, Go, Plus, Pro, Business, Edu, and Enterprise — and that credit packs are optional, needed only when a workspace owner wants to unblock users who exceed their limits. Everyone else waits for the limit to reset. On April 2, 2026, OpenAI did align Codex pricing to API-token usage rather than a flat per-message model for Plus, Pro, Business, and new Enterprise plans — but the subscription itself still carries the included allowance.

The philosophy is stated most directly by Sam Altman, describing why overage credits exist at all. It is a deliberate design choice: keep the base subscription cheap and predictable for the majority, and let a small set of power users pay to go past the caps.

The nuance that matters for a budget owner is opt-in versus mandatory. Under OpenAI's model, the default state is included; a user who does nothing simply hits a ceiling and resumes after a reset. Under Anthropic's post-July-7 model for Fable 5, the default state for continued access is metered; an org that does nothing loses the model. Same word — credits — opposite defaults.

OpenAI's own frontier story sits behind a different wall entirely. It previewed GPT-5.6 (Sol, Terra, and Luna) on June 26 as a government-gated limited preview — available through the API and Codex only, to roughly 20 government-approved companies, and not in the ChatGPT consumer app at preview stage. OpenAI has said it expects to widen access and reach a broad release “in the coming weeks,” but has not committed to a date, so treat any specific general-availability date you see as speculation. For the preview mechanics, see GPT-5.6's government-gated preview. In the meantime, the model still doing the daily work in most stacks is GPT-5.5, which stays broadly available while Sol remains gated.

Both pricing bets land into the same macro shift. CNBC reported on June 26 that the industry is moving from “tokenmaxxing” — maximizing token consumption to squeeze out capability — toward efficiency, as enterprise buyers start rationing spend. The evidence is concrete. Per CNBC, Uber implemented AI-tool spending tiers this month starting at a $1,500-per-employee-per-month base (with an option to request more), after its CTO told The Information the company had blown through its entire annual AI budget in just four months. AI startup Lindy moved 100% of its model traffic off Claude to a cheaper model to control costs, with CEO Flo Crivello framing it as a matter of survival for the business.

Here is the interpretation the raw numbers imply. When a customer's largest single variable cost is tokens, and that cost can quadruple in a quarter, the customer stops caring which model is marginally smartest and starts caring which billing shape it can forecast. That reframes the whole vendor relationship. Anthropic's run-rate surge to roughly $47 billion annualized (from a ~$9–10 billion 2025 base, per CNBC) is exactly the kind of growth that cannot continue if buyers start capping spend — which is the analyst read on why both labs are racing to go public.

No single published source lines up all three major vendors' tier-by-tier billing philosophy with the opt-in nuance attached. This is that view, as of July 1, 2026. Read the Overage column as the tell: “optional” means you can ignore it, “required” means the model stops without it, and “not offered” means access is bundled with no à la carte path.

Read the asterisks honestly. ChatGPT Business's seat price and its reported April 2026 cut come from pricing trackers rather than a primary confirmation in our research pass, and every Google Gemini figure here is aggregator-sourced (9to5Google, Finout) rather than confirmed against Google's own page this week — including the AI Ultra entry price reportedly cut from around $249.99 to about $99.99 per month. The direction of travel, though, is well corroborated across sources: OpenAI leaning included, Anthropic leaning metered, and Google keeping access bundled while quietly debiting a “compute-used” budget behind the subscription.

Once you are metered, the meter's shape matters. The table below takes the published per-token list rates and adds a single computed column: a blended cost per million tokens at a 3:1 input-to-output mix, defined as (3 × input + 1 × output) ÷ 4. That mix is a deliberately conservative stand-in for chat and coding workloads, where you read far more context than you generate. This is a billing-math illustration, not a capability comparison — Fable 5 is a flagship and GPT-5.6 is a government-gated preview, so read the spread as the shape of each meter, not as model parity.

The arithmetic is worth internalizing because it explains the behavior. Output tokens dominate the blended cost — Fable 5's $50 output rate pulls its blended figure to $20.00 even though its input is only $10.00. That is why an agent that generates long completions, reasoning traces, or repeated tool calls burns credits far faster than its input volume suggests. The GPT-5.6 preview tiers land at $11.25 (Sol), $5.63 (Terra), and $2.25 (Luna) on the same formula — preview list rates, sourced to Axios and MarkTechPost, and gated behind government-approved access. When you budget a metered workload, budget the output side first.

The vendor is the wrong first question. The right first question is how your token spend behaves — how predictable it is, and how much of it a machine drives without a human in the loop. Map the workload to a billing posture first; the vendor choice falls out of that. Five archetypes cover most of what agencies and product teams actually run.

The immediate action is a calendar item: if any part of your stack relies on Fable 5, decide before July 7 whether you enable usage credits or fall back to Opus 4.8 (which is also where blocked Fable 5 requests auto-reroute). Enterprises on standard seats are already metered, so the decision there is about limits, not access. Either way, turn on the spend-control and analytics tooling both vendors now ship — you cannot govern a variable cost you cannot see per team.

Looking forward, expect the meter to keep spreading, not retreating. Anthropic has now tested metered programmatic usage twice in a month; Google already debits a compute budget behind its subscriptions; and Microsoft has begun pitching lower-cost in-house models explicitly to reduce dependence on both labs, with GitHub Copilot auto-routing coding tasks to the most cost-appropriate model. The defensible posture is not loyalty to a vendor's pricing promise — OpenAI's included stance is a strategy, not a covenant — but architecture that can move: a documented fallback model, task-based routing, and a budget you re-check monthly. For the mechanics of routing and committing spend, our breakdown of token-based vs outcome-based agent pricing pairs with a token-budget calculator built for exactly this decision.

This is also where an outside pass earns its keep. When we scope a client's AI budget, we start by classifying each workload into the archetypes above, instrument the meter, and set spend controls before anything goes to production — the same way we scope agentic AI budgets in our AI transformation engagements. The goal is boring on purpose: no surprise invoices, and a billing shape you chose deliberately rather than inherited from whichever vendor changed its policy last.