Compare concrete examples without mistaking a token price or a benchmark score for the outcome of your deliverable.
WHAT THE TOKENS COST
An illustrative API basket: 1 million uncached input tokens + 100,000 output tokens. Concrete models make the price comparison auditable; Junior’s value proposition is the delegation workflow.
| Example model | Input / 1M | Output / 1M | Example basket |
|---|---|---|---|
| CommodityDeepSeek V4.1 Flash · off-peak | $0.15 | $0.60 | $0.21 |
| CommodityDeepSeek V4.1 Flash · peak | $0.30 | $1.20 | $0.42 |
| CommodityDeepSeek V4 Pro · off-peak | $0.66 | $1.98 | $0.858 |
| Mid-tierClaude Sonnet 5.5 | $2.00 | $10.00 | $3.00 |
| FrontierClaude Opus 5.5 | $4.00 | $20.00 | $6.00 |
Checked October 6, 2026. Sources: commodity example pricing and frontier and mid-tier pricing. These are direct API rates, not Claude Code or Codex subscription allowances. Cache discounts, reasoning/output volume, retries, routing-provider fees, and manager review change the bill. The Pro example also has peak rates: $1.32 input / $3.96 output.
At the listed rates, this basket costs $0.21 for the Flash off-peak example and $6.00 for the premium example—about 29× apart. That is a token-price comparison, not a promised reduction in the cost of completing your task.
Total delivery cost = planning + worker execution + retries + checks + review. Delegation earns its place when the work is substantial enough and the finish line is clear enough to outweigh that overhead.
CAPABILITY IS TASK-SPECIFIC
A low-cost model can be competitive on repository fixes while trailing on a harder terminal test. Choose around the deliverable, then verify the actual work.
| Evaluation | Commodity example | Frontier snapshot | Gap |
|---|---|---|---|
| DeepSWE v1.1Repository issue resolution | 74.2% | 74.0% | Commodity +0.2 points |
| Terminal-Bench 2.1Agentic terminal tasks | 90.6% | 89.1% | Commodity +1.5 points |
| Terminal-Bench 4.0A different, harder terminal evaluation | 31.2% | 51.8% | Frontier +20.6 points |
Source: publisher model card, checked October 6, 2026. Publisher-reported results, not a Junior evaluation or an independently controlled head-to-head test. The commodity results use maximum reasoning effort and benchmark-specific harnesses. Its DeepSWE score with Pi is 66.2%, versus 74.2% with mini-SWE: the harness matters. Scores from different benchmark versions must not be compared.
In Anthropic’s newer FrontierCode 1.1 results, proprietary Sonnet 5.5 scores 52.1% at Xhigh and Opus 5.5 scores 54.4%—a 2.3-point gap. This is a separate evaluation, not an open-weight comparison. A narrow benchmark gap does not establish equal judgment across open-ended work.