Skip to Content

Release Notes: Adaptive Model Control

Plan with the strongest models, execute with cheaper capable models, and switch back anytime in Neo.

You now have real model-layer flexibility.

Use premium models like Opus 4.7 and GPT 5.5 for planning and architecture decisions, then switch to lower-cost models like Qwen 3.6 27B (OpenRouter) or GPT 5.4 mini (OpenAI) for execution-heavy loops. Neo keeps the workflow stable while you change model strategy as requirements change.

Up to 7x faster across core workflow stagesUp to 67% lower cost in execution-heavy phasesSwitch anytime by task, phase, or rerun

Why this release matters

Phase 1: Plan with high-capability models

Use models like Opus 4.7 or GPT 5.5 when you need better decomposition, stronger reasoning, and cleaner execution plans before any heavy run starts.

  • Complex planning and architecture choices
  • Risky refactors and deep debugging strategy
  • Better first-pass plans, fewer wrong turns

Phase 2: Execute with cost-efficient models

Once the plan is clear, switch to models like Qwen 3.6 27B via OpenRouter or GPT 5.4 mini via OpenAI to run iterative implementation cycles at lower cost.

  • Code edits, retries, and repetitive transforms
  • Validation loops and test-driven iterations
  • Lower unit cost while preserving useful quality

Reswitch whenever needed: If execution quality drops or scope changes, move back to a stronger model for a difficult step, then return to a cheaper model for bulk implementation.


Provider flexibility in practice

Neo does not lock you into a bundled model layer. You keep provider control, and choose model/provider per workflow need.

Workflow phaseRecommended model postureExample options
Planning and scopingHigher-reasoning, premium modelsOpus 4.7, GPT 5.5
Execution and iterationCost-efficient capable modelsQwen 3.6 27B (OpenRouter), GPT 5.4 mini (OpenAI)
Hard blockers or regressionsTemporarily move back up-modelReturn to Opus 4.7 or GPT 5.5 for critical steps

What shipped

Execution speed improvements

  • Faster preprocessing, feature engineering, and evaluation loops
  • Up to 7x acceleration across common AI engineering workflows
  • Higher throughput for iterative coding cycles

Execution clarity in chat

  • Upfront execution plans before long runs
  • Clearer step-level progress and state visibility
  • More context on reasoning and approach decisions

Runtime reliability upgrades

  • Improved stability under long, multi-step sessions
  • Fewer runtime interruptions and retries
  • Stronger recovery behavior when failures occur

Benchmarks snapshot

Data preprocessing

Before: 5-8 minutes for 1M rows

Now: 1-2 minutes (about 5-7x faster)

Model training

Before: 15-20 minutes for ensemble runs

Now: 3-5 minutes (about 4-6x faster)

Feature engineering

Before: 10-15 minutes for complex transforms

Now: 2-3 minutes (about 5-7x faster)

End-to-end workflow

Before: 45-60 minutes

Now: 10-15 minutes (about 4-5x faster)


Migration notes


Need help choosing model strategy? Start with Bring Your Own LLM Keys, then review FAQ and Getting Started.