Augment Code Ships Prism: Per-Turn Model Routing Cuts Coding Agent Costs 20-30% at Frontier Quality
Augment Code shipped Prism on May 2. It is a model router that sits inside the IDE and makes routing decisions turn-by-turn rather than at the start of a session.
What It Does
Prism offers two configurations:
- Prism (GPT + Kimi): targets GPT-5.5 quality
- Prism (Claude + Gemini): targets Opus 4.7 quality
The router switches between models only when the expected quality gain from a different model exceeds the cost of evicting the cache. This is the key distinction from session-level routing: most existing tools fix the model at session start. Prism re-evaluates every turn.
On Augment’s internal multi-turn coding benchmark, Prism matches the best individual model on quality at 20-30% lower cost per task. Teams sending 10,000 user messages per month can expect to save $20,000 compared to running a single frontier model throughout.
The Architecture Argument
No single model wins every turn in a multi-step coding session. File reads, grep operations, documentation lookups, and docstring generation do not need the same model as the turn that writes a complex algorithm or debugs a race condition. Session-level model selection forces users to pick a model that is appropriate for the hardest task in the session, paying frontier rates for every turn.
Prism’s routing logic applies this distinction automatically, routing cache-eviction-free turns to cheaper models and escalating only when the difficulty warrants it.
The billing rolls up under a single Prism line item. The underlying model that handled any given turn is not surfaced to the user or the billing system — a deliberate choice that avoids the cognitive overhead of second-guessing the router.
Why This Matters Now
Frontier model costs for heavy coding agent users run $500-2,000 per engineer per month. GitHub Copilot moved to token-based metering in June. Claude Code Pro added a $100/month 5X tier after standard usage caps proved inadequate for production agent workloads. The cost pressure on engineering teams is real and accelerating.
Prism is currently available in VS Code, JetBrains, the Augment CLI (/model), and the web interface. It is the first commercially deployed per-turn coding router targeting both the GPT and Claude model families.
Limitations
Augment’s internal benchmark is not independently verified. The 20-30% cost figure will vary significantly depending on the task mix — sessions heavy in complex multi-step reasoning will see lower savings than those mixing complex turns with simpler tooling calls. The router’s quality equivalence claim has not been validated on standard external benchmarks.
That said, the underlying economics are sound. Frontier model pricing at $75-150/M output tokens is hard to justify for file reads. Routing to cheaper tiers for lower-complexity turns is where the 2026 cost curve for production coding agents gets controlled.