1017 used about 1.6x the output tokens of the September V4.1 Flash on my agent job

The 1017 notes promise shorter outputs and lower cost, so I tested it against the September build on one agent job. The job: turn the Surgeon General's 2023 social media advisory into a one-page HTML explainer, using a report-to-web-page skill (part of SenseNova-Skills). Every request went to Together, which charges the same for both builds. Per request, 1017 did write less: about 0.9k output tokens against 1.35k. But it made 177 and 161 requests against 85 and 59, many of them re-rendering the page to check it. In total that came to about 1.6x the output tokens, and $0.31 a run against $0.20 and $0.17. Timing is useless here: the September runs sat through about 10 and 5 minutes of provider 502 errors. All four pages got built. Has 1017 been heavier on tokens for you too?