How to work with GPT
- 6 Sol together with Kimi K3? This is my suggestion after testing. In a word: Sol is a better "brain", K3 is a cheaper "stamina + long memory". The matching principle is: leave judgment to Sol, and leave throughput to K3. Relative advantages and disadvantages of the two models (actual measurement data) • Sol is stronger: difficult agentic coding (DeepSWE 73.0 vs K3
- 5), slightly higher comprehensive intelligence, 12 times faster output, and the most complete ecological tool chain • K3 is stronger or equal: terminal tasks are almost tied (Terminal-Bench
- 3 vs
- 8), 1M context + cache hit input $0.30/M (Sol 1/17), the output price is only half of Sol, open source can be self-hosted, and the security boundary evaluation is good • K3’s shortcomings: slow output (62 t/s), the hardest reasoning questions are slightly inferior to mine My personal experience is that just send the following plan to codex-orchestration. codex-orchestration is a Codex plug-in: it turns "who plans, who consults, who designs, who executes" into configurable seat routing, allowing the root agent to assign different roles to different models according to tasks. Specific matching plan (sorted by income)
- Sol planning + K3 execution + Sol final review (main assembly line) This is the link with the greatest income. Sol splits the tasks into packets with clear boundaries and clear acceptance criteria → K3 implements them in batches as executors → Sol conducts adversarial reviews. The gap between K3 and Sol is very small in terms of "explicit execution of instructions", but the cost is only half. The planner/executor/reviewer seats in AGENTS.md in your workspace correspond exactly to: planner=Sol, executor=K3, reviewer=Sol.
- All the work of "swallowing" long contexts is given to K3 (millions of contexts). This is K3's underrated trump card. Reading the entire warehouse, hundreds of pages of logs, long documents, and collections of papers - 90% of the cost of such tasks is the input token, and inputting after the K3 cache hits is almost free (the hit rate in encoding scenarios is >90%). Let K3 read the same large context repeatedly for summary, positioning, and extraction, while Sol only looks at the refined conclusion to make judgments.
- For parallel sharding exploration, 3-5 K3 workers are responsible for sharding tasks such as code base auditing, batch reconstruction, and multi-file translation, and Sol is responsible for merging and adjudicating conflicts. The slowness of 62 t/s is evened out in parallel, and the cost is saved by half.
- Overnight batch processing is all provided to K3 for test generation, document completion, type annotation, and mechanical migration—work that does not require human attention, speed is not important, K3 runs all night, and Sol accepts it in the morning.
- Cross-review (cheap quality insurance) Error patterns differ between the two model families. The key code written by Sol allows K3 to find faults (or vice versa), and can catch blind spots that single model self-examination cannot catch. The cost of a K3review is approximately equal to free. In turn, do not give these to K3 • Architectural decisions, clarification of fuzzy requirements, and debugging of difficult problems - the
- 5-point gap in DeepSWE is reflected here • The final check of security-sensitive code • Scenarios that require rapid interactive iteration and human output (62 t/s will affect your flow)