Why We're Not Pre-Buying Claude 5 "Mythos" Credits — And How We're Preparing for Model Churn Instead

TL;DR: Rumored 3x Opus pricing ($16M/$80M per 1M tokens) makes lock-in the real risk. We’re building model-agnostic tooling and eval harnesses now — so any model swap takes hours, not sprints — full comparison.

The Context

Team of 4, currently routing architectural decisions to Claude 3.5 Sonnet (capped $50/mo). X leaks suggest “Mythos” (Claude 5) launches June with “Oceanus” enterprise tier at ~3x Opus pricing. Opus already costs $15/1M in, $75/1M out. If Oceanus is $45/1M in, $225/1M out, our $50/mo cap covers 1M tokens — ~20 complex architectural queries. That’s not a budget; that’s a constraint.

What We Tested

Strategy	Use Case	Verdict	Why
Pre-commit to Claude 5 / Oceanus	Architecture, security, unknown libs	❌	Price unknown; vendor lock-in; no eval baseline for new model
Model-agnostic router + eval harness	All cloud-tier queries	✅	Swap models in config; weekly eval catches regressions; provider-agnostic
Multi-provider fallback (Claude → GPT-4o → DeepSeek)	Cost control + availability	✅	If one provider spikes price or degrades, router fails over automatically
Local-first with cloud only for “unknown”	80% local, 20% cloud	✅	Reduces cloud dependency; only routes to cloud when local confidence < threshold

The Pivot Point

Claude 3 Opus launched at $15/1M in. 6 months later, 3.5 Sonnet launched at $3/1M in — 5x cheaper, better coding. Then token-based billing hit Copilot/Cursor. Pattern: model quality improves, pricing shifts unpredictably, tooling layers capture value. Teams locked to one provider (via proprietary SDKs, custom prompts, no eval) pay the “loyalty tax.” We decided: never again. Every prompt, every tool call, every eval must be portable.

What We Use Now

Three-layer portability stack:

1. Prompt/schema registry (JSON, versioned): Every agent prompt stored as {version, model_hints, expected_schema, golden_examples}. Not in code — in config. Swappable per model.

2. Provider-agnostic router (TypeScript, 400 LOC): Interface ModelProvider { complete(prompt, schema), stream(...), estimateCost(...) }. Implementations: AnthropicProvider, OpenAIProvider, DeepSeekProvider, OllamaProvider. Router picks by complexity tier + cost ceiling.

3. Weekly eval harness (50 golden prompts × 4 models): Runs every Monday. Tracks: pass rate, latency, cost, schema compliance. Dashboard in Grafana. If new model (e.g., Claude 5) beats current on ≥3 metrics → promote in router config. No code changes.

Current routing: Tier 4-5 → provider: anthropic, model: claude-3-5-sonnet, max_cost: $50/mo. When Mythos drops: add AnthropicProvider(model: claude-5-mythos), run eval, promote if it wins.

When You’d Choose Differently

• Enterprise with negotiated Anthropic contract: Volume discounts + SLA may justify lock-in. Still build eval harness.
• Single-model shop by design: If you’ve standardized on Anthropic stack (Tools API, MCP, SDK), migration cost > portability ROI.
• No eval capacity: Portability requires eval. If you can’t run weekly golden prompts, stick to one provider and accept the tax.

Tool Crucible Rating

Overall	Ease	Value	Support
4.0/5	2.5/5	4.5/5	3.0/5

---

This is part of our AI model strategy series. See full comparison: Claude 5 Mythos Preparation 2026