Fable 5 Before July 7: The Six-Day Window Playbook

Claude Fable 5 returned to Pro, Max, Team, and select Enterprise plans on July 1, 2026 — with a deadline attached. Through July 7, the model is included for up to 50% of each plan’s weekly usage limits. From July 8, every Fable 5 token is billed through metered usage credits at standard API rates. That leaves six days, counted from today, in which the most capable coding model Anthropic has shipped is effectively bundled into a subscription you already pay for.

The stakes are concrete. At Fable 5’s listed API rates of $10 per million input tokens and $50 per million output tokens, the kind of heavy agentic session the model was built for — multi-hour autonomous runs, whole-repo migrations, deep architecture reviews — moves from “covered by my plan” to “a line item I approve.” Teams that sequence the window deliberately will bank weeks of high-leverage work on subscription tokens. Teams that treat this like a normal week will spend their included allowance on tasks a cheaper model could have done.

This guide covers what actually changed at the relaunch, how the 50% inclusion mechanic works, what the July 8 flip to usage credits means operationally, and — the part nobody else has published — a task-by-task sequencing matrix for deciding what deserves window tokens before Tuesday and what can wait for metered access.

On Wednesday, July 1, 2026, Anthropic announced it was redeploying Claude Fable 5 and Claude Mythos 5 across the Claude Platform, Claude.ai, Claude Code, and Claude Cowork, with cloud-marketplace access on AWS, Google Cloud, and Microsoft Foundry to follow. The restore ended a 19-day pause — June 12 through June 30 — triggered by a reported jailbreak-bypass technique and resolved once U.S. export controls were lifted on June 30. We covered why Fable 5 disappeared for 19 days in the first place in depth when the pause hit.

The relaunch is not a simple rewind. Anthropic shipped a new safety classifier alongside the restore, which it says blocks the reported jailbreak technique in over 99% of cases — a vendor-stated figure, not an independently audited one. Requests the classifier flags are redirected to Opus 4.8 rather than answered by Fable 5, at what Anthropic estimated pre-incident would be under 5% of sessions. That fallback behavior has real implications for coding sessions, which we unpack in the coding tradeoffs of the new safety classifier.

The second change is the one this playbook exists for: the access terms. Fable 5 did not come back as a permanent subscription feature. It came back with an expiry date on plan inclusion — and a metering model waiting on the other side.

The core mechanic fits in one sentence of Anthropic’s announcement, and it rewards careful reading.

Three details matter. First, the 50% is relative — half of your plan’s weekly usage limit, whatever that is. Anthropic has not published exact per-plan weekly token counts for Pro, Max, or Team, so any absolute number you see quoted is invented. Second, the wording is “select Enterprise plans” — standard Enterprise seats get no inclusion window at all and are on usage credits from day one. Third, this cap is new. Hacker News commenters comparing notes against the original June 9 launch messaging concluded the 50% restriction wasn’t part of the first rollout.

The window is also shorter than what subscribers were originally promised. PCWorld reports the promotional access works out to half the originally promised two weeks: subscribers were told 14 days of included access when Fable 5 first launched on June 9, the export-control pause consumed most of that, and the makeup window on relaunch is seven days at up to half of weekly limits — of which six remain as of this writing, July 2 through July 7 inclusive.

One caveat before you plan the week: at least one user report on Hacker News suggests weekly usage counters were not reset at the relaunch, meaning usage consumed before the June 12 pause may still count against the current week. Anthropic has not confirmed or denied this — treat it as anecdotal, and check your own usage meter before committing to a heavy schedule.

From July 8, Fable 5 is not part of any subscription tier. It becomes a usage-based add-on: you enable usage credits under Settings > Usage in your Claude account, attach a payment method, set spending preferences, and every Fable 5 token is billed at standard API rates — $10 per million input tokens, $50 per million output tokens — as charges separate from your subscription. Credits come with guardrails: a $2,000 daily redemption limit, a configurable monthly spending cap (or unlimited, if you choose), usage alerts as you approach the cap, and an optional auto-reload that tops up the balance when it crosses a threshold you set.

The full mechanics — credit packs, plan-by-plan differences, and the billing edge cases — are in the full usage-credits pricing breakdown. The operational summary is the table below.

The under-covered detail in the relaunch coverage is a single line that changes the optimal strategy. TestingCatalog’s breaking-news summary of the new terms — corroborated verbatim by Hacker News commenters — ends with: “Fable 5 draws down usage faster than Opus 4.8.” The 50% inclusion shares a pool with everything else you run on the plan, and Fable 5 consumes that pool at a higher rate than Opus 4.8 does for equivalent work.

The naive strategy — route everything through Fable 5 this week because it’s included — is therefore exactly wrong. Every routine task you give Fable 5 burns allowance that a heavy migration or architecture review needed, and burns it faster than the same task would have on Opus. The window rewards discrimination, not enthusiasm. Our standing guidance on which model to reach for once the window closes applies inside the window too — just with the thresholds shifted.

What does “bank the heavy work” mean concretely? The first Fable 5 window, June 9–12, produced a public record of what the model is uniquely good at. Anthropic’s launch announcement cited Stripe reporting that Fable 5 completed a code change across a 50-million-line Ruby codebase in one day — work estimated at roughly two months for a team. Community build roundups from those 72 hours catalogued the same pattern at smaller scale: one developer ran nine concurrent overnight workstreams without losing context; another ran a persistent 10–12 hour benchmark-harness session that reportedly consumed over 100 million tokens; a full design-validation loop for a robotics actuator compressed into about 30 minutes and roughly 400,000 tokens.

Those examples share a profile: high context, high autonomy, high token volume, and output that stays valuable after the session ends. That profile is what the window is for. The matrix below is our triage tool — no other coverage we’ve seen organizes the window this way. When we run this triage for clients, it happens inside our AI transformation engagements — the same matrix, applied to a real backlog.

Sequenced across the six days, the matrix becomes a simple three-phase plan:

The single most reusable idea from the relaunch discussion isn’t Anthropic’s — it’s the community’s. Multiple Hacker News commenters independently described the same workflow: use the strongest available model to author a durable plan document, then hand execution to cheaper models running as subagents. One called the Opus-plans-Sonnet-implements split “great”; another described the fanout version directly:

The window supercharges this pattern. For six days, the planner seat is occupied by a strictly stronger model at zero marginal cost inside your allowance. A plan document authored by Fable 5 this week is an asset that keeps paying out after July 7 — every Sonnet subagent that executes against it inherits the quality of the plan without inheriting the price of the planner. The failure mode to avoid is the inverse: burning window allowance on execution-grade work while the planning-grade opportunities expire unused.

We’ve broken the full workflow — what makes a plan “durable,” how to structure handoff packets, and the prompts that produce executable plans rather than essays — into a dedicated companion piece on banking durable plans this week for cheaper models to execute later.

Two standard discounts reshape the post-window economics, and both stack with usage-credit billing. Prompt caching cuts repeated context to roughly 90% off — about $1 per million tokens on cache reads against Fable 5’s $10 input rate. The Batch API prices asynchronous work at 50% off standard rates: $5 per million input tokens and $25 per million output tokens. Inside the window these savings are invisible, because plan tokens are covering the bill. From July 8 they are the difference between a metered workload you can justify and one you can’t.

One community observation worth stealing: an HN commenter asked whether coding harnesses could route planning-type, non-interactive work through the Batch API for the extra discount. That instinct is right — anything tolerant of async turnaround belongs on the cheapest rail, and structuring workloads to maximize cache hits (stable system prompts, reused repo context) compounds the saving. The worked numbers, harness configuration, and spend-cap setup are in the cache-and-batch math for after July 8.

The subscriber reaction is genuinely split. The relief is real — one Hacker News commenter described canceling a Max plan and requesting a refund during the pause, then re-subscribing on July 1 the day Fable 5 returned. So is the frustration: subscribers were promised 14 days of included access at the June 9 launch, got about three before the pause, and returned to a shorter window at half of weekly limits.

On the post-window economics, one commenter’s back-of-envelope estimate — an individual guess, not a vendor or analyst figure — put heavy Fable 5 usage via credits at 10–20 times the price of a top-tier subscription for someone trying to sustain their in-window pace. Whatever the true multiple, the direction is right: from July 8, always-on Fable 5 usage is an enterprise-grade line item, not a hobbyist upgrade.

Our read of the trend: the window is less a promotion than a pricing experiment. Anthropic has a model that, by its own positioning and the community’s usage reports, is more expensive to serve and drains plan allowances faster than anything else in the lineup. A hard flip to pure metering — with a 50%-of-limits transition week as the offramp — lets Anthropic observe genuine willingness-to-pay before deciding whether Fable 5 ever fits inside a subscription tier again. One HN commenter argued market pressure will force exactly that outcome, speculating Anthropic may need “another pricing tier” for the model. That is plausible but unannounced; plan as if metering is permanent.

Looking forward, the sensible default for teams is a split portfolio: durable Fable 5 artifacts banked this week, Opus 4.8 as the everyday senior model, Sonnet for execution volume, and metered Fable 5 reserved for the workloads where its edge is provable. That calculus also depends on the competitive picture — how Fable 5 compares to the strongest alternative is the other half of the routing decision, and it’s why we treat model choice as a per-workload call rather than a vendor loyalty question.