HelloLucia LLC · Internal Exploration
Which of Google, Anthropic, or OpenAI is best for the public-facing Guest Agent? The Lucia-shaped answer: the guest-facing eval battery should decide, not vendor reputation. The canon already requires that battery before launch readiness — it is also the instrument that can compare providers.
The exploration still surfaces three concrete findings. First, the current default — GPT-5.6 Sol at $5/$30 per million tokens — is likely over-provisioned for the Guest Agent's deliberately bounded role, and OpenAI's cheaper tiers (Terra, Luna) are a zero-migration cost win worth trialing immediately. Second, Claude Sonnet 5 ($2/$10, pricing now permanent) is the strongest external candidate on the two priorities Lucia's doctrine weighs most — hospitality voice and refusal to invent — and deserves a head-to-head eval. Third, Gemini 3.7 Flash leads on cost and latency ($0.75/$3.75 promotional through December 2026), but its economics double in January 2027 and should not be anchored on the promo. The recommended path is a thin provider abstraction in the Guest Agent's model layer, then an evidence-based bake-off.
The canon docs constrain this decision helpfully. The Guest Agent is a warm front-of-house concierge, not an exposed database: it orients guests (booked, joining, planning, exploring), assists verification, and collects structured signals — while identity, verification status, calendar availability, booking data, and routing remain deterministic, system-owned truth. Models "improve natural language and reasoning expression only." The runtime already answers deterministically without model invocation on several paths, and the provenance contract (meta.model_runtime) records whether a model ran, which one, and whether fallback fired.
This bounds the requirement. The Guest Agent does not need any vendor's frontier reasoning tier. It needs a model that can hold Lucia's voice — warm, first-person, privacy-conscious, never bureaucratic — while following strict instructions about what it may not say or invent, returning reliable structured output, at guest-chat latency and public-facing scale. The tone doctrine and the four eval dimensions (identity orientation, verification safety, routing integrity, hospitality tone) are effectively the model requirements document.
One doctrine point governs any provider change: proof does not transfer. A new provider means new execution proof, new provenance evidence, and a re-run battery. That is a real cost of going multi-vendor, and it recurs with every model added.
OpenAI (incumbent). The Guest Agent already runs GPT-5.6 Sol via the Responses API through its own Netlify function. Current flagship-family pricing: Sol $5/$30, Terra $2/$12, Luna $0.20/$1.20 per million input/output tokens, all ~1M context, cached input at 10% of base. Migration effort is zero, and the provenance contract is already built around this API. The open question is whether Sol's "ambitious agentic work" positioning buys anything for a bounded concierge role that Terra — or even Luna — would not.
Anthropic. Current lineup: Claude Sonnet 5 at $2/$10 (introductory pricing made permanent), Haiku 4.5 at $1/$5, Opus 5 at $5/$25, Fable 5 at $10/$50. Sonnet 5 is the relevant candidate: 1M context, mature tool use and structured outputs, and a reputation — which the battery should test rather than trust — for warm, consistent persona and for honesty discipline (saying "I don't know" instead of inventing), which maps directly onto Lucia's truth-state doctrine. Two caveats: Claude models from the 4.7 generation onward use a tokenizer producing roughly 30% more tokens for the same text, so effective cost runs somewhat above sticker price; and integration means abstracting the current Responses-API-specific layer, since Anthropic's Messages API is not a drop-in for it.
Google. The relevant lineup: Gemini 3.7 Flash and 3.6 Flash at $0.75/$3.75 promotional through December 31, 2026, rising to $1.50/$7.50 in January; Gemini 3.5 Flash-Lite at $0.30/$2.50; older Flash-Lite tiers cheaper still. Flash models are fast, 1M-context, with schema-constrained output. Google's OpenAI-compatibility layer mimics Chat Completions, not the Responses API Lucia migrated to, so Gemini requires the same abstraction work as Anthropic. On Antigravity: it is Google's agent development platform — IDE, CLI, and an SDK in preview — potential build tooling, not a runtime the Guest Agent would depend on. And Gemini's no-training-on-your-data posture applies to the paid tier only; the free tier trains on inputs and is not appropriate for guest traffic.
Assume a 10-turn guest conversation accumulating ~30K input and ~5K output tokens, before caching (which cuts repeated-prefix input cost to ~10% on all three platforms):
| Model | Per conversation | Note |
|---|---|---|
| GPT-5.6 Sol | ~$0.30 | Incumbent default |
| GPT-5.6 Terra | ~$0.12 | Zero migration |
| Claude Sonnet 5 | ~$0.11–0.14 | Upper bound is tokenizer-adjusted |
| Gemini 3.7 Flash | ~$0.04 / ~$0.08 | Promo / from Jan 2027 |
| GPT-5.6 Luna | ~$0.012 | Zero migration |
| Haiku 4.5 · Flash-Lite 3.5 | ~$0.07 · ~$0.02 | Routing-tier candidates |
Two observations. Per-conversation costs are small everywhere — at controlled-preview volume, cost should not drive the decision; voice and truth discipline should. And the largest single saving available today requires no vendor change at all: Sol to Terra is a ~60% cut, Sol to Luna a ~96% cut, testable this week against the existing eval scenarios.
Conversation quality and voice. Unresolvable from benchmarks — hospitality tone is one of Lucia's four eval dimensions precisely because it must be judged on Lucia's own scenarios. Working hypothesis worth testing: Sonnet 5 and Sol strongest, Terra close, Gemini Flash competent but historically more utilitarian in register.
Disclosure: this report was researched with a Claude model. Treat the voice hypothesis as a claim to falsify, not a finding.
Reliability and truth discipline. All three support schema-enforced structured output and tool calling. The differentiator is behavioral — whether the model respects "missing evidence stays unknown" under pressure from a chatty guest. That is exactly what the verification-safety and routing-integrity eval dimensions measure. Lucia's deterministic spine bounds the blast radius of any model failure, which argues that the cheapest model that passes the battery is the right model.
Cost and latency. Gemini Flash leads on both; Luna is the dark horse at near-Flash-Lite pricing with zero migration. Latency for all candidates is adequate for chat; measure it in the bake-off rather than trusting published numbers.
Migration effort. OpenAI: none. Anthropic and Google: comparable — both need the model layer abstracted behind Lucia's own contract (which the "provider swaps must preserve the response contract" doctrine effectively mandates anyway), plus fresh provenance proof. The abstraction is a one-time cost that converts every future model question from an engineering project into an eval run.
Three viable shapes, in ascending complexity: a router (a cheap model classifies each turn and routes it to the best provider), specialist delegation (one model owns the guest-facing voice; others handle bounded subtasks), and ensembles (multiple models answer or cross-check). For the Guest Agent, delegation is the only shape that fits the tone doctrine — guests must hear one consistent "she," and raw output from three vendors will not sound like one person. Ensembles double cost and latency for a surface whose hard guarantees already come from the deterministic spine, not the model.
The honest version of "one mind" for Lucia: the mind is the Engine's deterministic truth plus the doctrine; the model is a replaceable voice-and-reasoning component — which is precisely what the current architecture already asserts. Multi-vendor's real value here is leverage and resilience (certified fallback across providers, pricing leverage, no lock-in), not blended intelligence.
First, trial GPT-5.6 Terra and Luna against the existing guest-facing scenarios — zero migration, immediate signal on whether Sol is over-provisioned. Second, build the thin provider abstraction inside the Guest Agent's model layer, preserving the response contract and meta.model_runtime provenance semantics per doctrine. Third, run the guest-facing eval battery — the one the canon already requires before launch readiness — across a shortlist of Claude Sonnet 5, Gemini 3.7 Flash, and the best OpenAI tier from step one, scoring all four dimensions plus measured latency and cost. Fourth, pick the primary voice model on evidence, and keep a second provider wired as certified fallback: that alone captures most of the multi-vendor benefit at a fraction of the "one mind" complexity. Defer router or ensemble ambitions until the battery is green on one model.