All this noise around opencode go deepseek v4.1 flash vs direct API. I ran my own real world tests in Hermes - my everyday agent. Here's what it found - no difference between the two. Here is Hermes testing and results: Setup: - Model: deepseek-v4.1-flash - A: OpenCode Go · B: DeepSeek direct API - Temperature 0, identical effort, identical prompts, interleaved - logprobs blocked on A: HTTP 400 "not supported by the cell router yet" Test 1 — Differential code, 9 tasks × 200 random inputs - A 1,800/1,800 · B 1,800/1,800 - Tasks: CSV parser (quoted commas/newlines), LRU cache, LCS, interval merge, topo sort + cycle detect, token bucket, int→Roman, expression evaluator, bug hunt - Pass = all 200 cases match reference implementation Test 2 — Escalation, 3 tasks × 150 inputs - A 450/450 · B 450/450 - JSON parser from scratch vs json.loads · diff edit-script reconstruction · 40-line bug hunt Test 3 — Compounding reasoning, 10 problems, pre-verified answers - A 10/10 · B 10/10 - Byte-identical answers: 10/10 - One item: 2,906 vs 2,465 reasoning tokens, same final integer Test 4 — Mixed suite, 27 tasks - A 26/27 (96.3%, CI 82–99) · B 25/27 (92.6%, CI 77–98) — intervals overlap - Byte-identical: 22/27 - All 5 divergences cosmetic (valid variants, different syntax, same logic) - Verifiers: exact arithmetic, script-countable constraints, JSON schema, tool-call schema, unit tests Test 5 — Latency, 02:27 UTC (inside peak window) - B 1.29–1.9s median · A 2.15–2.6s median - Consistent 1.4–1.7x penalty Test 6 — Effort-mapping probe - Default (high): 36 completion tokens - reasoning_effort=low: 31 - reasoning_effort=none: 4, zero reasoning chars - medium/low/high/xhigh all map to high → medium is a no-op Totals 4,500 differential executions (2,250 cases × 2) 0 cases where one passed and the other failed 10/10 byte-identical on exact answers $0.07 total spend