LatentEval
LatentEval
Everyday AI

The model you never picked is the one about to bill you.

Anthropic switched Claude Fable 5 back on July 1, included on Pro, Max, and Team through July 7, then metered at twice Opus 4.8's price. A routed verdict on whether you need it, every number sourced.

By LatentEval Published

Open Claude and Fable 5 sits at the top of the model list. Your plan shows no extra charge for it. Leave it selected and it will draft your emails, tidy your paragraphs, and patch your one-line bugs. The picker shows today’s price, and nothing about July 8th’s. After July 71, that same habit bills at twice Opus 4.8’s rate2, on work Opus was already doing faster and for half the money. You would never notice. The bill would.

Route by task, not by default: switch Fable 5 on for long, multi-step research and agent runs, and turn it off for everyday chat, drafting, and short coding, where Anthropic itself points you to Opus 4.82 at half the price. Backwards, and it costs you both ways: the free window spent on work Opus already nails, then double the rate after July 7 for a gain you never feel.

Through July 7: Fable 5 is included on Pro, Max, and Team, up to half your weekly usage1, at no extra charge. After that: it runs on usage credits (the metered balance your plan bills against) or the standard metered rate, at twice what Opus 4.8 costs2.

Free right now, but only through July 7, and not on every plan

Anthropic pulled Fable 5 on June 12 under a US export-control order, then switched it back on worldwide on July 1. The version that returned came with a fuse. On Pro, Max, Team, and select Enterprise plans, it runs free for at most half your weekly usage and only through July 71; cross either line and it starts drawing usage credits.

The word “included” does not mean the same thing on every plan, so check your row before you plan around it. Standard Enterprise gets no included allowance1: there, Fable 5 runs on usage credits from the first message. Everyone else (Pro, Max, Team, and select Enterprise) holds the capped free window until the 7th.

Is Fable 5 actually better for your work?

Anthropic calls Fable 5 its most capable widely released model, built for the most demanding reasoning and long-horizon agentic work: jobs that chain many steps toward one result.3 Whether the bill is worth paying comes down to a single question: is your work demanding in that specific way?

Two-panel diagram contrasting how an error behaves in chained versus standalone work. Left panel, Chained work, each step feeds the next: four steps, Step 1 to Step 4, stacked and connected by downward arrows. Step 1 is clean. A small red error enters at Step 2, is inherited by each step below it, and grows larger at Step 3 and larger again at Step 4 until it dominates the final step. The arrows below Step 2 are red, showing the error propagating down the chain. Caption: each step feeds the next; Fable 5 can move the outcome. Right panel, Standalone tasks, each stands alone: four separate tiles, Task 1 to Task 4, in a two-by-two grid with no arrows between them. Three tiles are clean; one tile, Task 3, carries a small red error of the same small size that stays inside its own tile and never spreads or grows. Caption: each task stands alone; a stronger model changes nothing here. This is an explanatory, non-data visual with no numbers.

Fable 5 is a senior specialist you fly in for the day. On a weeks-long project, that judgment pays for itself many times over. On a fifteen-minute question, you buy the whole day and get the answer the colleague at your desk already had.

Picture Maya, nine hours into a literature review across dozens of papers, feeding Claude each new source to reconcile against everything it has already summarized. The latest source never gets read fresh; the model reasons over the running summary that every earlier source built.

Each step builds on the one before it. A sharper reconciliation at paper six becomes the ground paper seven stands on. A small error there is inherited as settled fact by every source that follows. That compounding is where a stronger model earns its price. A single email gives it nothing to build on.

Marcus, opening Claude the same morning, is fixing the tone of a cover letter and a function that throws every time the input list arrives empty. Two quick, self-contained jobs, with nothing downstream waiting on either. For exactly that work, Opus 4.8 is the model Anthropic routes you to2, and it answers faster at half the metered price. Maya switches it on before July 7; Marcus leaves it off.

We put that to our own test. On short tasks with a checkable right answer (structured extraction, tool calls, code you can unit-test), we could not measure Fable beating Opus at all: the two land inside the same margin at every effort setting we tried.4 The work Fable is built to win is the other kind, the long, open-ended, many-step runs, and that is the work this sort of test cannot reach. So the edge you pay for is real in theory and, on everyday checkable work, unmeasured in practice. It is insurance you buy sight unseen.

Anthropic draws the line the same way: start with Opus 4.8 for most work, and reach for Fable 5 only for workloads that need the highest available capability.2 If each step of your task feeds the next, Fable 5 can move the outcome; if your prompts stand alone, it cannot, and you would pay double for nothing.

After July 7, the bill doubles for the same prompt

All the date does is remove the cover. Underneath, the price was always double. Claude bills per token, the chunks of text a model reads and writes, and Fable 5 lists at $10 per million in and $50 out3, against Opus 4.8’s $5 and $25. For every dollar Opus charges, Fable 5 charges two.

Claude Fable 5Claude Opus 4.8
API price per 1M tokens$10 in / $50 out$5 in / $25 out
LatencySlower, you wait while it spins upModerate
Best fitlong, multi-step reasoning and agent runseveryday chat, drafting, short coding

Prices and the latency row come from Anthropic’s model overview2; the best-fit split from that page’s routing guidance and Fable 5’s own description3.

On an everyday request, Fable 5 makes you wait longer and costs more for the answer Opus already hands you. That is waste: the same answer for a higher bill. But does Fable 5 ever make everyday work actively worse?

On everyday coding, most declines route straight to Opus anyway

On one slice of quick coding, yes. Fable 5 ships a stricter safety filter, one that blocks more requests before answering, and Anthropic is candid about the trade: it comes at the cost of flagging benign requests more often during routine coding and debugging tasks1. Flagged inside Claude’s own apps, you still get an answer: you see a notice, and the request is handed to Opus 4.81.

Pictograph of 28 dots, one per benign, checkable everyday coding task, from LatentEval's routing-eval, low effort, fallback-on arm, arranged 7 columns by 4 rows. The 2 left columns, 8 orange dots, are tasks Fable 5 answered itself. The 5 right columns, 20 green dots, are tasks Fable 5 declined that Claude silently served from Opus 4.8 (fallback fired on 20 of 28, 71%). Zero tasks were left with no answer; 27 of 28 were correct overall (96.4%, 95% Wilson CI 82 to 99). Left to itself, Fable 5 declined 75 to 86% of these tasks across effort settings (21 of 28 at low effort, 24 of 28 at maximum; more thinking did not reduce the refusal level). Fable 5 lists at twice Opus 4.8's per-token price. Sample is small, n=28, and covers short-horizon checkable coding only. Source: LatentEval routing-eval, Anthropic API, 2026-07-02.

How often is “more”? In the same test, on benign, checkable coding tasks (parsers, state machines, string transforms, small simulations), Fable declined most of them: 21 of 28 at low effort and 24 of 28 at maximum, roughly 75 to 86% either way. More thinking held the rate right where it was, at both settings. Of the 45 declines, 44 were the filter reading ordinary code as a cyber risk. The answer a user actually received landed on about one try in six, 5 of 28 at low effort and 4 of 28 at maximum, against roughly nine in ten for Opus (25 to 27 of 28).

Set that beside the price. On a slice of ordinary coding help, the pricier model declines, then reroutes you to the cheaper one, and you read the Opus answer regardless. You land in the same place. You just paid Fable 5’s premium for the detour to an Opus answer you could have picked yourself.

Route the work, then watch the date

  • Switch it on for long, multi-step reasoning or agent tasks, and test them before July 7 while the free window still covers them. That is where a stronger model moves the outcome.
  • Leave it off for short chats, quick drafts, and one-off code fixes. Anthropic routes that work to Opus 4.82, at half the API price and lower latency.
  • Check your plan. Pro, Max, and Team carry the capped free window; standard Enterprise starts on usage credits1.
  • Decide before the 7th. If you do keep Fable 5 on, settle whether the work is worth usage credits or the metered $10 in, $50 out3, because paying becomes the default the moment the date passes.

The deadline passes on July 7. The routing choice underneath it does not: every time you open Claude, you are still picking a model. “Most capable” measures peak benchmark scores. Whether a model gets your specific job right, run after run, is a separate measurement, and our research desk covers that gap in how to measure agent reliability past a single pass rate. Pick the model your task needs, no matter what sits at the top of the leaderboard.

Footnotes

  1. Anthropic, Redeploying Claude Fable 5, https://www.anthropic.com/news/redeploying-fable-5 ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7

  2. Anthropic, Claude models overview, https://platform.claude.com/docs/en/about-claude/models/overview ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7

  3. Anthropic, Introducing Claude Fable 5 and Claude Mythos 5, https://platform.claude.com/docs/en/about-claude/models/introducing-claude-fable-5 ↩ ↩2 ↩3 ↩4

  4. Our own routing-eval harness: deterministic exact-match / unit-test scoring (no LLM judge), Anthropic API, pricing re-verified live 2026-07-02; an effort sweep of 14 pruned tasks x 2 runs = 28 trials per arm (prior round n=40), 95% Wilson intervals, a pre-registered stop rule. It measures short-horizon checkable correctness only, not long-horizon agentic or open-ended quality. Method and definitions: agent reliability testing. ↩