Run postmortem · Published 2026-10-13
Gemini 4 Argon at high effort completed a valid 29/100 run in 56m 33s, at an estimated API cost of $4.43.
One-hour track · high reasoning · provisional
Gemini 4 Argon explicitly finalized a valid package and answered 29 of 100 independent hidden questions correctly. Its final public-validation score was 25/25. The controller finished cleanly in 56 minutes 33 seconds, with no provider errors or broker infrastructure errors. One training call hit OpenCode's default two-minute command timeout; the worker was released immediately and Gemini recovered with a longer timeout. This is one valid seed in the normal one-hour benchmark and remains provisional. ARI Bench grades the compact system the agent builds, rather than asking the controller to answer hidden questions directly.
One valid autonomous run produced a provisional 29% hidden exact-match score.
1 of 3 required seeds · Hidden exact-match accuracy, not public calibration accuracy.
Gemini built a hybrid byte predictor that combines a trained transformer, a prefix n-gram copier and a symbolic solver. It generated synthetic examples, trained several models and repeatedly tested packaged candidates.
The inference code handles local mappings, records, arithmetic, sequences and passages with explicit rules. Otherwise it uses the trained byte transformer, lightly blended with a prefix n-gram copying distribution.
Eight training attempts included seven successful jobs totaling 36,700 steps, including the 200-step starter warmup. The substantive models used six layers, width 384, six attention heads, RoPE, RMSNorm and SwiGLU. Training mixed the public corpus with generated examples and weighted answer bytes more heavily. The selected weights came from the successful 8,000-step model_v6 pass, with a 512-byte neural context.
Fourteen candidate submissions included thirteen valid packages and one oversized package. Gemini explicitly finalized its selected candidate. The 10,098,911-byte gzip artifact fits the 16 MB cap. The graded code, configuration and weights match the saved selected candidate; large weight matrices use int8 quantization with row scales.
After a 200-step starter warmup, Gemini generated synthetic exam examples and implemented a symbolic solver. It corrected two generator IndexErrors and changed a rejected inline Python call into an allowed script call. Its first substantive hybrid package passed 25/25 public validation.
A 3,500-step training attempt hit the default 120,000 ms tool timeout. The broker cancelled it and immediately released the worker. Gemini retried successfully with a longer timeout. It also corrected an oversized package before continuing.
Successful passes trained for 3,000, 3,500, 6,000, 8,000, 6,000 and 10,000 steps after the starter warmup. Later packaged candidates continued to pass 25/25 public validation. Gemini tested inference variants and selected its earlier 8,000-step weights.
Gemini finalized hybrid_v10_best and exited cleanly. All 124 broker requests received responses. The independent hidden test scored the saved selected artifact at 29/100.
Gemini corrected development errors, recovered from a training timeout, completed seven training jobs and finalized a package under the size cap. Artifact hashes confirmed that the saved candidate was the one graded.
The final package passed all 25 public cases but answered only 29 of 100 hidden cases correctly. Its perfect public score did not carry over to the separate hidden test.
This single seed does not establish a stable ranking. High was the highest supported reasoning setting in the model's retained launch catalog. The run used the standard Google AI Studio endpoint through OpenRouter, OpenCode 1.18.32 and a 128,000-token response allowance.
Every archived run selected for this exact model-and-effort row is shown below. The narrative uses aggregate run evidence and final artifact structure; it does not expose hidden questions, answers, or raw private transcripts.
Swipe horizontally to see all run columns.
| Run | Score | Time | Cost at run | Artifact | Candidates | Grade |
|---|---|---|---|---|---|---|
| run-0137 | 29% | 56:33 | $4.43 | 10.10 MB | 14 | Valid |
197 reasoning tokens are included in the estimated cost. Response allowance: 128,000 tokens.
Autonomous runs are stochastic. Two more valid seeds are required before this can be treated as an official estimate.
The agent made 14 candidate submission actions across the published run set. Selected artifacts averaged 10.10 MB compressed.
The runs averaged 56:33 of wall time. 1 run explicitly finalized before the one-hour limit.
Retained totals: 140 agent messages · 14 candidate actions · 1 published run.
$4.43 estimated for this run, including reasoning.
The estimated API cost is US$4.43, calculated from complete metadata for all 139 model responses: 471,157 uncached input tokens, 137,375 visible output tokens, 197 reasoning tokens and 21,122,121 cached input tokens. Published endpoint rates were $2 per million uncached input tokens, $10 per million output tokens including reasoning, $0.10 per million cached input tokens and $0.1875 per million cache-write tokens; no cache writes were recorded. These are the introductory rates verified on October 13, 2026. No separate long-context tier was published in the retained endpoint snapshot. The provider-billed amount has not been independently reconciled. The charts use this complete estimate, including reasoning.
At current configured API-equivalent rates, the displayed run cost is $4.43 and value is 6.55 score points per dollar.
Estimated API cost, including reasoning / run · pricing snapshot 2026-10-13.
Adjacent rows provide score context without treating small one-seed differences as settled model rankings.
Compare Gemini 4 Argon with the complete ARI Bench field, including score, current cost, value per dollar, effort, and seed status.
View the leaderboard