💡 Intelligent Cost-Optimizing Model Router: Automatic Sonnet/Opus Task-Based Selection

[github-actions[bot]]

Summary

TalkTerm automatically routes each agent interaction to the optimal Claude model (Sonnet 5 vs Opus) based on task complexity, saving BYOK users 30-60% on API costs without any manual model selection. Simple Q&A and single-turn tasks route to Sonnet; multi-step workflows, complex analysis, and tool-heavy operations route to Opus. The avatar transparently indicates which model is active and users can override at any time.

Market Signal

Claude Sonnet 5 (June 2026) delivers near-Opus intelligence at Sonnet cost with a 1M context window. OpenAI's ChatGPT auto-selects models silently. Google Gemini Spark optimizes costs automatically. BYOK users bear API costs directly — automatic optimization is table stakes for cost-conscious adoption. Meanwhile, 56% of CEOs report zero measurable ROI from AI (Gallagher 2026 AI Adoption Benchmarking); cost transparency and optimization are emerging as competitive levers for tools that put users in control of their spend.

User Signal

Existing idea #55 (Effort Modes: Economy / Power / Turbo Session Tiers) requires users to manually select tiers — but TalkTerm's target users are non-technical knowledge workers. Goldman Sachs 2026 survey data shows 45% of small business AI users cite lack of technical expertise as their primary adoption barrier. Asking them to choose between model tiers is still a technical decision that creates friction. Automatic routing removes this cognitive burden entirely while delivering the same cost savings.

Technical Opportunity

The AgentBackend abstraction layer (architecture.md) already isolates the renderer from backend implementation details. Adding a routing layer between the use case and the Claude SDK backend is architecturally clean — a decorator or strategy pattern on AgentBackend. The TaskComplexityClassifier interface can be defined as a port in src/shared/types/ports/, with the initial implementation using heuristics (message length, expected tool count, workflow step count). Sonnet 5's benchmark parity with Opus on most tasks makes conservative routing (default Opus, offload only clear-cut simple tasks to Sonnet) low-risk.

Assessment

Dimension	Score	Rationale
Feasibility	high	Anthropic API already supports model selection per request; routing logic is a simple decorator on existing AgentBackend interface
Impact	high	30-60% BYOK cost reduction directly addresses cost sensitivity — the #1 concern for individual API key users
Urgency	high	Sonnet 5 released June 2026; competitors (ChatGPT, Gemini) already auto-optimize. Delay means launching without table-stakes cost optimization

Adversarial Review

Strongest objection: If Sonnet produces lower quality on tasks the router classifies as "simple," users lose trust in TalkTerm without understanding why the quality degraded.

Rebuttal: Start with only the most conservative routing (single factual questions, formatting requests). Always show which model is active via avatar UI indicator. Let users override per-session. Expand Sonnet routing gradually as confidence data accumulates from preference memory (FR51). The 80/20 rule: even routing just 20% of interactions to Sonnet saves meaningful cost with minimal quality risk.

Suggested Next Step

Define the TaskComplexityClassifier interface as a port in src/shared/types/ports/. Implement a heuristic-based classifier (message length, tool count, workflow step count) as the initial strategy. Add model routing to the AgentBackend decorator chain in main.ts composition root. Design the avatar UI model indicator (subtle badge showing "Sonnet" or "Opus" near the status indicator).

[github-actions[bot]]

Weekly Update

What Changed

Three major developments reshape the model routing landscape:

1. Claude Sonnet 5 released June 30, 2026 at introductory pricing of $2/$10 per M tokens (rising to $3/$15 after August 31). This creates a time-limited window where aggressive Sonnet routing saves 33% compared to post-introductory pricing.

2. Claude Fable 5 (Mythos-class) launched July 2026 at $10/$50 per M tokens, purpose-built for long-running autonomous agents. This is a new routing target: tasks requiring multi-hour autonomous operation should route to Fable 5 rather than Opus.

3. Compound savings now stack — Anthropic confirmed that prompt caching, batch processing, and long-context multipliers can compound for 95%+ total savings. A smart router that combines model selection with caching strategy could deliver dramatically lower costs for BYOK users.

Current model lineup (Aug 2026):

Model	Input/Output (per M tokens)	Best for
Haiku 4.5	$1/$5	Quick questions, low-stakes tasks
Sonnet 5	$2/$10 (introductory)	Standard work, balanced quality/cost
Opus 5	TBD (July 24 release)	Complex reasoning, high-stakes decisions
Fable 5	$10/$50	Multi-hour autonomous deep work

Sources: Claude Platform Release Notes, Anthropic Prompt Caching Update

Updated Assessment

Dimension	Previous	Current	Delta
Feasibility	high	high	→ Architecture supports model selection via AgentBackend
Impact	high	high	→ 4 model tiers + stacking discounts = significant cost optimization
Urgency	med	high	↑ Sonnet 5 introductory pricing ends Aug 31; Fable 5 is available now

Recommendation

Advance. The expanded model lineup and stacking discounts make cost-optimizing routing even more valuable than when originally proposed. Priority actions: (1) Add Fable 5 as a routing target for deep work sessions (see new proposal #434), (2) Implement Sonnet 5 as default for standard interactions, (3) Design cost estimation UI that accounts for compound savings. The Sonnet 5 pricing deadline (Aug 31) creates urgency for users to benefit from introductory rates.