OpenClacky Deep Dive: Does the 'Money-Saving' Agent Really Save Money, or Is It a Numbers Game?
> One-line verdict: OpenClacky is indeed far cheaper than OpenClaw and Hermes, but the claim of being 'cheaper than Claude Code' deserves a discount—it is more expensive on some tasks, and the measured cache hit rate (90.6%) falls short of the marketing line ('approaching 100%'). Its real value lies not in 'saving money' but in its BYOK model routing and the Skill economy experiment.
Key Points
- Benchmark results (same prompts, model, skill versions, OpenRouter per-request CSV reconciliation): OpenClacky $5.10 / 51 requests / 90.6% cache hit; Claude Code $5.49 / 70 / 95.2%; OpenClaw $15.70 / 81 / 88.7%; Hermes $30.14 / 218 / 60.3% — all on Claude Opus 4.7.
- Counterintuitive finding #1: Claude Code has a *higher* cache hit rate. OpenClacky wins on total cost because it uses fewer requests (51 vs 70); each request carries fixed overhead. This is a request-scheduling victory, not cache-technology superiority.
- Counterintuitive finding #2: On the marketing-page task, Claude Code is 43% cheaper ($1.20 vs $1.72). Task-level breakdown: PPT $1.23 vs $1.45; marketing page $1.72 vs $1.20; social content $2.14 vs $2.84 (OpenClacky vs Claude Code).
- Marketing vs engineering: The site claims cache hits 'approach 100%' while its own benchmark shows 90.6% — still best-in-class among open-source agents, but the wording overstates the data.
- Cache design: OpenClacky uses dual cache-point markers (marking the last two messages), session persistence, insert-then-compress, and idle-time compression. Effectiveness across providers (OpenAI prompt caching vs Anthropic cache control) needs independent verification.
- Tool economy: 16 core tools with an
invoke_skillmeta-tool vs Hermes' 52. A back-of-envelope estimate (200 tokens per tool schema) suggests ~$7.56 savings per 70-request session, though real implementations may load tool subsets conditionally. - BYOK risks: cache is not portable across providers (Anthropic/OpenAI/DeepSeek/Kimi/MiniMax/OpenRouter); routing subtasks to cheaper models can backfire on cache costs; BYOK shifts operational complexity to users.
- Skill economy: Encrypted, license-managed skill distribution (creator-set pricing) — closer to a 'skill app store' than a framework. Legal/medical/wealth-management skills face compliance and quality-standardization challenges.
- Investors: MiraclePlus, ZhenFund, Sequoia China, Hillhouse Ventures — capital backs the 'cheaper agent infrastructure' track, not necessarily this product.
- OpenClacky site: https://www.openclacky.com
- Benchmark: https://www.openclacky.com/benchmark (2026-04-30, OpenRouter CSV reconciliation)
- GitHub: https://github.com/clacky-ai/openclacky
- Creator program: https://www.openclacky.com/creators
- Comparison baselines: Claude Code ($5.49), OpenClaw ($15.70), Hermes ($30.14); tasks: guizang-ppt-skill, marketing-psychology, social-content
Cognitive Biases to Watch
1. Survivorship bias: Only three tasks tested; conclusions don't generalize to refactoring, data analysis, or multi-step reasoning. 2. Attribution error: Cost differences are attributed to 'cache design' but likely come from request scheduling, subtask routing, and context compression. 3. Capability-equivalence assumption: Output quality equivalence is asserted subjectively, not measured.
Honest Framing
> On PPT and social-content tasks, OpenClacky is ~15–25% cheaper than Claude Code; on the marketing-page task, Claude Code is cheaper. On average, OpenClacky is slightly cheaper under identical models and skills, but the gap is small. Its decisive advantages are being 67% cheaper than OpenClaw and 83% cheaper than Hermes, plus BYOK model-choice freedom.
Conclusion
OpenClacky's real contribution is proving that an agent's cost structure can be systematically optimized (cache engineering, tool minimalism, idle compression, model routing) while staying open source. Its marketing ('approaching 100% cache', 'beats Claude Code across the board') needs scrutiny—90.6% is a good number that doesn't need exaggeration.
> The second half of the agent era isn't 'who is smarter' but 'who dares to leave it running every day.' OpenClacky's direction is right; the numbers need to be more honest.