The industry consensus is clear: models get bigger, smarter, and more capable. GPT-5 beats GPT-4. Opus 4.8 beats Opus 4.6. The frontier advances every quarter.
I think the real gains will come from somewhere else.
The Diminishing Returns of Scale
Each new model generation is more expensive to train and operate. GPT-5 required estimated $2B in compute. Claude Opus generations require clusters that most companies can’t afford.
Yet the performance gains per dollar are shrinking. GPT-5 is roughly 15% better than GPT-4 on complex reasoning. Opus 4.8 is around 10% better than 4.6. The raw intelligence gains are real but the cost-to-gain ratio is worsening.
The Orchestration Alternative
Instead of one $2B model that does everything, what if you had:
- One small, fast model for code generation ($0.15/M tokens)
- One medium model for planning and architecture ($2/M tokens)
- One specialized model for security review ($1/M tokens)
- Five tiny models for linting, formatting, and test generation ($0.08/M tokens)
Orchestrated together, this system outperforms a single large model on most tasks and costs a fraction as much.
Proof It Works
Hermes’s MOA (Mixture of Agents) system already does this. Multiple smaller models collaborate: one proposes, another critiques, a third synthesizes. The combined output is more accurate than any single model in the mix.
The cost savings are dramatic. A MOA round with 5 small models costs as much as 1-2 calls to a frontier model — but produces better results on complex reasoning tasks.
Why This Wins Long-Term
The orchestration approach has better economics:
- Smaller models improve faster — edge models catch up to frontier models every 6-8 months. The gap is closing.
- Orchestration is software — it improves with code, not compute. Better routing, better prompts, better task decomposition.
- Cost per task decreases — orchestration optimizes for total task cost, not per-token price.
What This Means
If you’re a team picking tools today, optimize for agents with good orchestration, not just access to the best model. Agents that route tasks intelligently across models will give you better results than agents that dump everything into one frontier model.
Hermes’s MOA and subagent systems are ahead here. Codex’s worktree isolation enables a form of orchestration. Claude Code’s subagents are a start.
The most productive coding setup in 2027 won’t be the one with the smartest model. It’ll be the one that uses models intelligently — matching task complexity to model capability, not using a sledgehammer for every nail.