Gemini 4 Argon reaches 29/36 basic verdicts with proof

Basic investigations: 29 verdicts with accepted proof

Gemini’s four-case proof lead over Sol and Astra is an observed count from one pass. High effort does not imply equal internal computation across providers.

Intrusions: the proof lead does not carry over

Gemini matched Sonnet’s four accepted-proof verdicts, one fewer than Sol, Astra and Opus. It completed 47/48 workflows overall. One basic response hit the output limit and ended without a valid final submission; that case stays in the denominator. The other four completed 48/48.

Cost:

Gemini recorded $21.69 in settled charges, with a conservative upper bound of $23.59 after three temporary rate-limit responses with unresolved charges. An abandoned Vertex attempt adds $0.08–$0.62 in setup overhead, excluded from the scored workload. Combined accounting, including every retained reservation, is $24.21 against the approved $40 cap.

Gemini used Standard Google AI Studio pricing: $2/M input and $10/M output through November 30. Astra was measured on Flex; its Standard estimate reprices the recorded workload. Estimates preserve caching and retries; price-component ratios were checked against archived rates. Ranges retain unknown charges, rather than express statistical uncertainty.

Gemini took longer on both tiers

Gemini averaged about 6m 20s per basic case and 9m 44s per intrusion. Time sums provider requests, including retries, and excludes cooldowns and restart gaps. Astra’s measured Flex time remains Flex time; a Standard price estimate does not predict Standard speed.

Explore scores, cost, speed and tokens

Switch tiers and choose correct verdicts or verdicts with proof. The charts use a frozen 22-model snapshot, initially showing current measured flagships.

Token coverage: Gemini has usage for 942/945 attempts: 583/586 basic requests and all 359 intrusion requests. Its combined totals are lower bounds (≥). The basic token mean stays unknown and is omitted from that plot; the intrusion mean is available. Missing usage is never zero. Older Astra and Opus request journals are unavailable here. Provider tokenizers differ, so token totals do not measure internal compute.

What this comparison supports

Gemini produced more evidence-backed basic conclusions in this pass, at a higher recorded bill and longer request time than Sol. The basic proof lead did not carry over to intrusions. Judge the task, proof, completion and operating cost together.

The limits: One pass per model, with 36 basic cases in 12 related families and 12 intrusions in four. Related variants are not independent incidents. These preliminary results need replication and do not establish statistical superiority or performance on an unseen investigation. We measured model endpoints in the same harness; consumer ChatGPT, Gemini and Claude apps are outside this comparison.