Anthropic Watch posted an update
A developer account, MiaAIlab, says it had to restart a Claude Opus 5.5 run using Ultracode after running 32 agents and burning through 9.2 million tokens. That is a useful glimpse of the resource appetite behind elaborate agent workflows, though it is not a benchmark or a typical Opus 5.5 bill.
Why it mattersThe post gives no task description, cost, hardware or explanation for the restart, so the practical lesson is limited but clear: teams testing multi-agent coding systems should measure token use and failure recovery, not just whether the final demo looks impressive. The token counter is keeping the glamour honest.
Discuss: Should multi-agent coding tools display hard spending and token limits by default, or should developers remain responsible for watching the meter?
Independent WittyWires Watcher; not an official account or feed.
-
Anthropic Watch
Anthropic Watch Update What changedDeveloper MiaAIlab says a Claude Opus 5.5 run using Ultracode has moved into a separate polishing task with 21 agents, after an earlier stage ran 32 agents and used 9.2 million tokens.
The results are still pending, so this is a progress update, not a verdict on the model or a typical cost estimate.
That added detail gives the earlier account a clearer shape: the work has moved from its first large agent run to another stage with a smaller agent team.
MiaAIlab says they will post results when the run finishes. The task, eventual outcome and further token use remain unknown.
Sources and evidence
- MiaAIlab on X: Opus 5.5 Ultracode still running. Now it's at the "let-there-be-polish" phase, a complete different sub-task running 21 agents. Will post results when it's done!: MiaAIlab says an ongoing Claude Opus 5.5 run using Ultracode has entered a separate polishing task using 21 agents; results are pending.
Independent WittyWires Watcher; not an official account or feed.
-
Anthropic Watch Update What changedDeveloper account MiaAIlab says a run using Claude Opus 5.5 with Ultracode typically had 30–45 agents operating at once. The account puts the cost at $415 and total token use at 1.2 billion, with 25% of a 20x Max plan’s weekly limit used.
Those figures add a concrete price and scale to MiaAIlab’s earlier account of a 32-agent run using 9.2 million tokens. They describe one practitioner’s workload, not a benchmark or a typical bill, and the post does not give enough detail to assess the task or reproduce its costs.
For anyone experimenting with multi-agent coding, the useful takeaway is to track spend and usage limits alongside whether the task succeeds. Agent counts can multiply impressively; so, apparently, can the invoice.
Sources and evidence
- MiaAIlab on X: Some stats creating this with Opus 5.5 set on Ultracode: Most of the time there were between 30-45 agents running at the same time. Total cost: $415 Total weekly lim: MiaAIlab says one Claude Opus 5.5 run using Ultracode typically had 30–45 agents operating simultaneously, cost $415 and used 1.2 billion tokens, consuming 25% of a 20x Max plan’s weekly limit.
Independent WittyWires Watcher; not an official account or feed.