fleet2.eduz.ai — agent & model fleet

Agent & model fleet — which model for which job, what's available, what it costs. Today: static catalog. v1: live session aggregator.

v0: static model catalog. Session-level live data (active runs, token spend, cost roll-up) needs a backend that reads ~/.claude/projects/-home-pk/*.jsonl — that's v1 work, and will be auth-gated (session content is private). For now: use this page as the routing reference when picking a model for a task.

Claude family (primary fleet)

Anthropic Claude 4.X — what's installed in this universe. Knowledge cutoff: January 2026.

model	tier	id	strengths	typical use
Opus 4.8	flagship	`claude-opus-4-8`	Deepest reasoning, longest-context judgment, most reliable on multi-step planning	Architecture, CTO-mode, design reviews, ambiguous problems
Opus 4.7	flagship (prev)	`claude-opus-4-7`	Same family as 4.8, slightly older; what this session is running on	Same as 4.8 when 4.8 unavailable
Sonnet 4.6	balanced	`claude-sonnet-4-6`	Good reasoning at ~5x cheaper than Opus; great for most coding work	Default for app-level coding, refactors, ports where judgment isn't critical
Haiku 4.5	fast/cheap	`claude-haiku-4-5-20251001`	Lowest cost, fast TTFB, still capable for mechanical work	Subagent dispatch (file copies, mechanical ports, batched lookups). Used heavily in this repo's build.
Fable 5	special	`claude-fable-5`	Specialized variant	Niche use cases

Non-Claude providers (registered in Doppler)

Other model providers we have keys for. Stored in Doppler workplaces below, fetched on demand.

provider	account / source	Doppler scope	tier note	typical use
DeepSeek	via API key	`eduz1-core/deepseek/prd`	Very cheap; strong coding	Mechanical-but-tricky code, when Haiku is too brittle
Gemini	prasoonky@gmail.com (AI Studio)	`eduz1b-core/gemini/prd`	Free tier (rate-limited) — verified via `-FreeTier` quota IDs	Free-tier experiments; long-context tasks within the rate limit
GLM	(zhipu)	`eduz1-core/glm/prd`	Z-family model series	Alternative when Western models are throttled
OpenRouter	aggregator	`eduz1-core/openrouter/prd`	One key, many models	Testing models we don't have direct accounts with
Mimo	(Xiaomi)	`eduz1-core/mimo/prd`	Open-source small-model family	Self-hosted experiments
RunPod	self-hosted compute	`eduz1-core/runpod/prd`	GPU pods for OSS model serving	When we want full control of the model + data

Routing rules of thumb

Hard judgment, ambiguous design, ARCH → Opus 4.7/4.8 (this universe's primary loop)
Implementation of a clear spec, multi-file edits → Sonnet 4.6
Mechanical: file copies, bulk renames, README ports, scaffold stamps → Haiku 4.5 (subagent dispatch)
Free experiments, rate-limit-tolerant exploration → Gemini (free tier)
Self-hosting / privacy-critical / batch → RunPod + open-source models

For this repo specifically: Opus drives the conversation; Haiku does the heavy file copies and README ports (cheap, parallel). DeepSeek + Gemini are available for future cost-sensitive batch work.

v1 ideas (auth-gated)

Live session list — read ~/.claude/projects/-home-pk/*.jsonl, render sessions (model, start time, message count, last activity).
Token roll-up — sum subagent_tokens across sessions to surface daily / weekly spend.
Fleet "now" — which agents are actively running (parallel forks, background tasks).
Cost projection — extrapolate spend at current burn rate.

All gated by auth-allowlist once that's wired across the 2-series.