anthropic shipped claude fable 5 yesterday. it’s the most capable model they’ve ever put in public hands. state of the art on basically every benchmark they ran: software engineering, science, vision. it also writes novel molecular-biology hypotheses for fun.

it’s also the first claude that will quietly hand your conversation to a different model in the middle of a task.

the tag-team thing

fable is “mythos-class,” the tier anthropic spent last week warning was getting too dangerous to ship. so they shipped a version with a bouncer.

ask fable about cybersecurity, biology, chemistry, or how to distill a model, and a classifier catches the request and routes it to opus 4.8 instead. you don’t pick this. fable taps out, opus taps in, the answer comes back from the more cautious sibling. there’s even a support page explaining why claude “switched models” on you mid-conversation.

anthropic says more than 95% of sessions never hit the fallback. great number, unless your 5% is the whole job. running a security agent, a bioinformatics pipeline, anything that lives in those four lanes? you’re not in the 95%. you’re the reason the bouncer exists.

what it actually costs you

• on your pro/max plan: included free through june 22. after june 23 it comes out of separate metered credits at api rates: $10 per million in, $50 per million out, double opus 4.8.
• in claude code: needs v2.1.170 or newer. older versions don’t even show it in the picker.
• the fine print: fable requires 30-day data retention. anthropic holds prompts and outputs for up to 30 days to run the safety classifiers, then deletes them. but there’s no zero-retention opt-out, which you get on other models.

should you point an agent at it

for most agent work (research, content, ops, glue code), fable is the strongest thing anthropic sells. now that the free window’s closed it bills from separate metered credits (see above), so weigh it against opus before you point an agent at it full-time.

for anything in the four restricted lanes, you’ll land on opus 4.8 whether you asked for it or not. so you might as well run opus 4.8 directly and skip the retention clause.

the honest read: it’s an incredible model with an asterisk, and the asterisk is the point. anthropic built the most powerful thing they could, then built a second model whose only job is to stand in front of it.

putting an agent on it

if your agents run on 5dive, you can move one to fable right now. three ways, pick your speed:

• dashboard: open Agents, click the model badge on any claude agent’s row, pick fable (it’s the one wearing a red “new” pill). live now.
• telegram: message the agent /model fable, or just /model for a tap picker (opus, sonnet, fable). fable lands here once your box takes tonight’s plugin refresh. want it today, update your box.
• cli: sudo 5dive agent config <name> set model=fable.

any of them edits the config and restarts the agent on the spot, back in about 20-30 seconds. it runs on your own pro/max plan, so through jun 22 your agents get the flagship at no extra cost. after that it shifts to metered credits like everywhere else.

and when fable taps out to opus 4.8 mid-task, your agent doesn’t stall. it finishes that one reply on the quieter model and keeps going.

the cli’s open source: github.com/5dive-ai/5dive. run a team of agents on whichever claude you want, over telegram →