Fact-checking the 200x/400x claim for a Jev + Opus 5.5 trading agent
The author traces the source of the claim that Jev with Opus 5.5 makes a trading agent 200x faster and 400x cheaper, noting there is no public trading test and neither the paper nor the code is released.
OPUS 5.5 + JEV JUST RAN 200X FASTER AND 400X CHEAPER THAN CLAUDE ALONE, ITS PITCH CLAIMS Opus 5.5 is Claude, so I priced both halves from public numbers. A 1.1M-view post pitches the pair for trading agents and claims a 6-page paper plus a Rust codebase. The author's posts never assign roles; public guides to this unofficial pairing do. → OPUS 5.5: plans and writes code at $4 in, $20 out per million tokens, thinking always on → JEV: TypeSafe AI's fast "System One" model that returns typed decisions instead of text → QUICK CALLS: per those guides, Jev picks the next action and flags risky commands → LOOP: agents hunt strategies 24/7; the author claims about 3 days of running so far → UNSAID: data feed, exchange, risk gates, sizing, and who places the trade Tracing the two multiples > claimed: 200x faster, 400x cheaper than Claude alone, per the author > claimed: 193.6x and 444.6x on TypeSafe's own homepage, the likely source > checked: an independent eval found Jev at 1/295th of Opus 5.5's cost on one task but gave no speed ratio against Opus > checked: 3,105 self-reported user posts give medians of 7x faster and 30x cheaper > derived: input prices give 95x ($4 vs a vendor-listed $0.042); 400x needs 0.64+ Opus output tokens per input, thinking included > checked: no public trading test of Jev; it placed last of twelve on a small legal-benchmark slice > not public: 3 days of "incredible" results, the paper, the code; no figures, statement or repo found Even at 400x, the pair still pays for every Opus call. Jev decides but doesn't enforce permissions, so trading limits live in your code. Before trusting "incredible", ask for a statement, a repo and each multiple's baseline. ↓ Terminal below: speed and cost bars, pair vs one big model, each tagged claimed, checked or derived.
This post quotes another post: x.com/0xkarmi/status/2108207473379180586