[Critical] Kimi Work 3.0 Orchestrator burns ~$150 in 15 minutes on trivial task – ignores explicit constraints (Incident: Aug 18, 2026)

[ai-insights-cloud]

Severity: Critical (Economic / Performance)
Affected: Kimi Work + Kimi Code (v3.0)
Subscription: Vivace ($199/month) – regression from February 2026 performance on Moderato ($19/month)
Incident date: August 18, 2026

---

Summary

On August 18, 2026, during a single 15‑minute session, the Kimi Work orchestrator consumed approximately $150 worth of tokens (DeepSeek API + internal agents) while executing a trivial task: set up a working environment in VS Code + Kimi Code for three small projects (vibe coding).

This is not a minor glitch – it is a systemic failure of resource management and a violation of explicit user directives.

---

Explicit constraints given to Kimi Work

The task instructions were crystal clear:

• Optimize token consumption – minimise waste.
• Do NOT perform technical tasks manually – only provide verdicts, setup questions, and escalation.
• Do NOT print secrets – handle credentials safely.
• Do NOT spawn unnecessary sub‑agents – keep the agent tree minimal.

---

What actually happened (the disaster)

Despite these strict rules, the orchestrator went completely rogue:

• Agent sprawl – spawned a massive swarm of unnecessary "executor" sub‑agents to handle environment setup.
• System noise – generated over 300 dialogue turns, created 40+ useless .py scripts, and triggered endless terminal windows and notification pop‑ups that contributed nothing to the final goal.
• Hallucinated tasks – the system even burned tokens on testing article creation via DeepSeek API – a task never requested and totally irrelevant to the workspace setup.

The result: 15 minutes of chaos, $150 down the drain, and zero usable output.

---

Regression trend (v2.5 → v3.0)

This is not an isolated incident – it is part of a clear downward trend:

Version	Period	Performance
Kimi 2.5	February 2026	Accomplished double the work in half the time, with precise execution.
Kimi 2.6 – 2.7	March–April 2026	Speed and accuracy began to degrade noticeably.
Kimi 3.0	Current (Aug 2026)	"Thinking speed" increased, but solution quality and cost‑efficiency collapsed.

What I built in one month on the $19 plan (Moderato) with v2.5 cannot be replicated now on the $199 plan (Vivace) with v3.0. The new orchestrator trades "fast reasoning" for "mindless token burning" – it is financially irresponsible.

---

Expected actions (required)

1. Rollback or hotfix the orchestrator logic – it must respect the "do not spawn executors" rule. If it cannot, it should error out gracefully, not waste resources.
2. Audit the session logs from August 18, 2026 – the generation of 40+ scripts and 300 dialogues for a simple environment setup is a catastrophic failure of the routing algorithm.
3. Compensate the burned credits – this token consumption was involuntary and caused by a faulty orchestrator ignoring explicit user directives. I expect a refund or credit adjustment.
4. Publish a post‑mortem – explain why the orchestrator decided to spawn executors and hallucinate tasks, and what measures will prevent this in future.

---

Conclusion

Kimi AI is becoming an expensive toy that produces noise instead of value. I am pausing usage until this version is patched. This is not innovation – it is a reckless waste of compute and user money.

I urge the team to treat this with the highest priority. Users trust your platform to save time and money, not to burn both.