Fable 5 returned July 1 — use it well before July 7

Quick summary

• Fable 5 is included on paid plans only through July 7, capped at 50% of your weekly limit.
• After July 7 it's metered credits at $10 / $50 per million tokens — 2x Opus 4.8.
• Front-load genuinely valuable heavy jobs now, while access is cheapest — not junk to burn quota.
• After the window, drop to Sonnet 5 or open-weight GLM-5.2 by task, not by brand.

The clock: your plan covers Fable 5 only until July 7

Right now, on Pro, Max, Team, and select Enterprise plans, Fable 5 is included — but it can eat up to 50% of your weekly usage limit, so it's included, not unlimited. After July 7, access shifts to metered usage credits, and Fable 5 is the most expensive model Anthropic sells. In plain terms: the price of using it is about to change, and this is the cheapest it will be for a while. Plan the next few days around that, not around panic.

What the credits actually cost

The usage-credit and API rate is $10 per million input tokens and $50 per million output tokens — twice Opus 4.8's $5 / $25, and the top of Anthropic's price list. A heavy coding run of roughly 50k input plus 15k output tokens works out to about $1.25; do 30 of those a day and you're near $37.50/day, or $1,100+ a month. Two levers cut that: the Batch API halves the rate to $5 / $25 for work that isn't time-sensitive, and cache hits drop input to $1 per million. Honest caveat: Anthropic hasn't published what '50% of weekly limits' equals in tokens or dollars, nor the credit-to-dollar conversion after July 7 — so treat the plan-window value as unknown and the API math above as your firm reference point.

Front-load your real heavy work now

This window is the moment to run the genuinely valuable, token-hungry jobs you were going to do anyway — big refactors, bulk test generation, framework migrations, documentation passes, large research syntheses. These are exactly the tasks that get expensive under metered credits, so doing them now is where the real savings live. This isn't about wasting quota on junk; it's about pulling forward work that has to happen and cashing in the lowest price on it. Make a short list of your actual backlog and start with the biggest, clearest wins.

Loop it or drive it

Some work runs fine unattended — well-specified, repetitive, verifiable tasks like refactors, test-writing, migrations, and bulk transforms can be batched or scheduled to grind through a backlog while you do other things. Other work needs you at the keyboard: ambiguous debugging, architecture calls, and design judgment where the next step depends on what you just saw. A good rule is to orchestrate the well-defined jobs as a multi-step workflow you can leave running, and reserve live sessions for the messy, judgment-heavy ones. Sort your backlog into those two piles before the window closes.

What to use after the window

When Fable 5 turns into premium credits, the cheaper routes stay open. Claude Sonnet 5 launched alongside it as a cheaper agentic model — $2 / $10 per million tokens through Aug 31 (then $3 / $15), roughly 40-60% less than the flagship, scoring 63.2% on agentic coding versus Opus 4.8's 69.2%. Open-weight GLM-5.2 (MIT license) runs about one-sixth the cost of Opus and lands close to it on coding. Pick by the task in front of you, not by the logo — many jobs never needed the most expensive model to begin with.

What's next — and why this isn't your last shot

This isn't your only window. GPT-5.6 (Sol, Terra, Luna) launched June 26 but only to about 20 US-government-approved organizations via API and Codex; general availability is 'in coming weeks,' mid-July at best per Axios, gated by a June 2 executive order requiring federal benchmarking — and OpenAI has reportedly been slow this week. Meanwhile the cheap options keep improving: GLM-5.2 and Nous Research's Hermes MoA (a mixture-of-agents approach) stay inexpensive and keep closing the gap. Separately, the team behind this site is building Coily, a layer that routes each task to a model you can actually afford and runs it for you — it's in development, not launched, so stay tuned.