> One-line conclusion: OpenClacky is genuinely much cheaper than OpenClaw and Hermes, but the "cheaper than Claude Code" claim needs an asterisk—it is more expensive on some tasks, and its measured cache hit rate (90.6%) falls short of the marketing claim ("approaching 100%"). Its real value lies not in "saving money" but in its BYOK model routing and Skill economy experiment.
Key points
- Headline benchmark (OpenClacky official, 2026-04-30, OpenRouter CSV reconciliation, all on Claude Opus 4.7):
- Counterintuitive finding #1: Claude Code's cache hit rate (95.2%) is actually *higher* than OpenClacky's (90.6%). OpenClacky wins on total cost because it uses fewer requests (51 vs 70)—each request carries baseline overhead (system prompt, tool descriptions). This is a request scheduling victory, not a caching breakthrough.
- Counterintuitive finding #2: On the marketing-page task, Claude Code was 43% cheaper ($1.20 vs $1.72). Task-level breakdown:
- Caches don't transfer across providers — splitting a session across Claude and DeepSeek means paying each provider's cache separately; total cost could rise.
- Operational complexity — BYOK means managing keys, quotas, and failover yourself.
- Cost transparency — showing dollar amounts per step is a UI improvement, but true transparency requires per-request token breakdowns and cache hit/miss details.
- Survivorship bias: only three tasks were benchmarked; conclusions may not generalize.
- Attribution error: cost differences are credited to "cache design" and "fewer tools," but the data points to request scheduling and subtask routing.
- Functional equivalence assumption: "identical capability" was subjectively judged, not objectively measured.
- OpenClacky: 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
- Test tasks: guizang-ppt-skill, marketing-psychology, social-content (all on Claude Opus 4.7)
| Agent | Total cost | Requests | Cache hit rate | Model | |---|---|---|---|---| | OpenClacky | $5.10 | 51 | 90.6% | Claude Opus 4.7 | | Claude Code | $5.49 | 70 | 95.2% | Claude Opus 4.7 | | OpenClaw | $15.70 | 81 | 88.7% | Claude Opus 4.7 | | Hermes | $30.14 | 218 | 60.3% | Claude Opus 4.7 |
| Task | OpenClacky | Claude Code | OpenClaw | Hermes | |---|---|---|---|---| | 01 PPT creation | $1.23 | $1.45 | $5.07 | $10.96 | | 02 Marketing page | $1.72 | $1.20 | $7.47 | $4.65 | | 03 Social content | $2.14 | $2.84 | $3.15 | $14.53 |
Caching: "Approaching 100%" — engineering or marketing?
OpenClacky's four claimed savings mechanisms: 1. No session restarts — the system prompt cache anchor stays valid 2. Dual cache-point marking — marks the last two messages, so advancing the conversation doesn't invalidate the anchor 3. Insert-then-Compress — context compression reuses existing cache 4. Idle-time auto-compression — don't pay while "waiting for you to come back"
The site claims "Measured cache hits approach 100%" while the benchmark shows 90.6%. That 9.4% gap compounds significantly in production. Still, 90.6% is already excellent among open-source agents (OpenClaw: 88.7%, Hermes: 60.3%)—the inflation was unnecessary.
The dual cache-point design is clever but depends on provider-specific cache APIs; whether it works equally well on OpenAI prompt caching vs Anthropic cache control needs independent verification.
16 tools: architecture choice or feature amputation?
| Agent | Core tools | Extension | |---|---|---| | Hermes | 52 | All built-in | | Claude Code | 40+ | Built-in + extensions | | OpenClaw | 23–50 | Built-in + plugins + MCP | | OpenClacky | 16 | invoke_skill meta-tool |
Back-of-envelope: at ~200 tokens per tool schema, 52 tools vs 16 tools means ~7,200 tokens saved per request—roughly $7.56 over a 70-request session. But this assumes full schema loading every request; real implementations may load subsets or compress descriptions.
The invoke_skill meta-tool is like OS system calls: a small stable kernel keeps cache anchors from drifting, while skills load on demand. Risks: multi-turn skill invocations can offset schema savings, and skill quality becomes the bottleneck.
BYOK: freedom or trap?
Supported providers: Claude, GPT, DeepSeek, Kimi, MiniMax, OpenRouter. Three risks:
Skill economy and investors
OpenClacky's creator program showcases vertical Skills (QingClaw for legal, JoyClaw for wealth management, MediClaw for medical) with encrypted distribution and license management—more a "skill app store" than a development framework. Success depends on skill quality, willingness to pay, and fair platform fees.
Investors include MiraclePlus, ZhenFund, Sequoia China, and Hillhouse Ventures—validation of the "cheaper agents" track, not of the product itself.
Cognitive biases to watch
Takeaway
OpenClacky's real contribution is proving that agent cost structures can be systematically optimized—via cache engineering, tool minimalism, idle compression, and model routing. Its honest positioning: meaningfully cheaper than OpenClaw (–67%) and Hermes (–83%), only marginally cheaper than Claude Code on average, and sometimes more expensive. "Approaching 100% cache hits" should be stated as the 90.6% it actually is.
> 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 more honesty.