The Price of a Second Opinion: Lower-Priced Model Routes, Frontier Baselines, and What Actually Moved the Score

Can lower-priced API model routes match a frontier-tier solo model at lower estimated cost? This preprint reports locally recorded experiments by task, keeping registered raw scores separate from secondary and post hoc analyses. In the original TAGTEAM harness, the optional helper was not called in 112 Haiku runs where it was configured. In 64 Sonnet runs it was also not called, but every session reported a helper-server connection failure, so those zeros do not establish a model choice. A forced handoff later loaded and answered in 62 of 62 calls. A promising 24-task broad-team result did not replicate on 96 cases. On a frozen 80-puzzle set, the registered prompt gave Astra 71/80 and Terra to Sonnet 68/80. A clearer answer contract, written after inspecting failures on the same questions, raised Astra to 80/80 and three ordered pairs to 79/80; those rerun results are post hoc. On 20 selected LongMemEval-S chats, eight routes were tested. Full-history format extraction brought Astra, Sonnet, and Terra to Sonnet to 20/20; Gemini Pro reached 19/20 and Grok 4.3 high 14/20. With retrieved sessions, five routes tied at 18/20 and Gemini Pro was cheapest at an estimated USD 1.15. Gemini Pro cost about 30 times Terra on the puzzles but about 2 times on the long chats, showing that value on one task did not carry over. Costs are estimates from saved token usage, not invoices. No general model ranking or tested adaptive router is claimed. The companion archive contains the run manifest, redacted result notes, registrations, case tables, runner scripts, and an audit addendum. AI contributions are credited by role in the manuscript.