Total deductions for 39 runs in OpenCode totaled $0.67 for the DeepSeek Flash model compared to $8.76 for Claude Opus. The experiment was based on solving 13 applied Python tasks with three attempts for each, yielding a total of 156 runs.
The economics of autonomous coding are rewriting development budgeting rules: the token cost gap between models differs by a factor of multiple times despite identical test quality. In non-interactive mode using OpenCode 1.19.6 without modifying default parameters, all four test participants successfully passed 39 out of 39 checks. The question of whether it is worth overpaying for Anthropic's solutions comes down to generation architecture and hidden overhead costs.
Expense architecture and price gap
The testing involved four models: DeepSeek V4.1 Flash (1017), DeepSeek V4 Pro, Claude Sonnet 5.5, and Claude Opus 5.5. Requests to DeepSeek solutions were routed through an OpenAI-compatible endpoint, while the Claude family was accessed via an Anthropic-compatible interface. Processing went through the TeamoRouter gateway, which recorded total expenses for 156 runs at $15.98. The Sonnet model cost $3.65, and V4 Pro required a deduction of $2.89.
Based on unrounded deduction amounts, Opus proved to be 13 times more expensive than Flash, whereas Sonnet exceeded the base DeepSeek model in expenses by 5.4 times. On the DeepSWE benchmark, the V4.1-Flash-1017 model demonstrated a result of 77.0 percent, as noted in previous measurements.
Recalculation of recorded tokens using official vendor tariffs yielded a total of $47.64: for Flash, the calculated figure was $0.79, while for Opus it reached $30.67. The discrepancy between real deductions through the gateway and the nominal price list is explained by context caching rules.
Clean code versus redundant tests
The test suite included ten separate functions with strict edge conditions and three small repositories ranging from 87 to 299 lines of code. All four models successfully passed the checks without reaching their own complexity limits. However, identical green tests masked fundamentally different agent behavioral patterns in the project working directory.
In the transaction task, Claude left dozens of new tests in the repository, whereas DeepSeek verified the code and removed temporary scripts.
An objection exists stating that the thirteen-fold gap in the receipt is fully justified by the depth of analysis and the reliability of code coverage through automated checks. Such bloating of the test base did not alter the final validation result, but it did increase context overhead costs.
What came before
While in mid-October DeepSeek V4.1-Flash-1017 was benchmarked on DeepSWE at an average cost just below $0.20 per task, thirty-nine tasks inside non-interactive OpenCode ran for under a dollar in total fees. On tiny codebases where every participant closed all edge cases, deploying heavy Claude options simply drained routing balances on redundant automated validation. Will a thirteenfold billing gap hold any practical merit once repository sizes exceed three hundred lines and push agents against their actual complexity limits?
Our stories: Claude Code Costs 2.7x More Than Rivals for AI Agent Tasks (August 6); Deepseek-V4.1-Flash Cuts Enterprise AI Inference Costs to $0.20 (October 17).