Review AI Engine models and costs¶
The AI Engine separates model discovery from model selection. Refresh model lists reads the configured providers' catalogues. It updates available choices and retirement status; it does not replace a saved choice or rewrite recommendations. A provider refresh failure preserves its previous models. An incomplete paginated response is a failure, rather than evidence that models retired.
The Priority list selects chat models. Speech and image models belong on Tasks, where Custom filters choices to the task's capabilities and product rules. A provider default can select different models for drafting, analysis, images as input, and short answers. Its eligibility model is not necessarily the model a particular operation calls.
Choose by workload¶
The existing recommendations implement the owner's provider and budget rules, rather than a universal ranking. Their reasons and review dates are displayed on Tasks. The following is the current recommendation policy, not a claim that each model is the newest available or wins every benchmark.
| Workload | Existing recommendation | Reason and constraints |
|---|---|---|
| Inspector, critique, grading, calibration, private assistant | Foundry GPT-5 mini | Structured responses and longer inputs; private assistant stays on Foundry. |
| Drafts, studio assistance, study scripts, podcast scripts | Foundry GPT-5 mini | Longer content generation under the existing subscription budget. |
| Captions and Telegram replies | Foundry GPT-5 nano | Short responses; lower published table rate than mini. |
| Alt text | Foundry GPT-5 mini | Image input; compare quality on representative images before changing. |
| Public explanations | Gemini Flash-Lite | Existing cached public path; NVIDIA trial tier is excluded. |
| Source grounding | Gemini Flash | Existing web-page and YouTube adapter. |
| Listen & Learn speech | Gemini 3.8 Flash TTS | Current long-form multi-speaker model; Flash-Lite is the cheaper Custom option. |
| Podcast voice | ElevenLabs v3 | Product is restricted to the existing ElevenLabs dialogue adapter. |
| Cover and manual images | Replicate Imagen 4 Fast | Existing image adapter and per-image pricing. |
Compare candidates using the same real workload: response validity, output quality, latency, and cost for the expected input and output sizes. A one-word connection Test establishes connectivity; it does not establish task quality or which model is cheapest. Media task Test resolves routing without making paid audio or images.
Recording transcripts, Page OCR, and Embeddings are planned registry entries without connected implementations. Changing a model or key cannot enable them. They do not expose routing controls until their adapters and call sites exist.
Gemini speech review, 2026-10-27¶
Google's official TTS cookbook
names gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts. It states that these
models speak text verbatim and take speaker information in speech_metadata
annotations. The adapter supports that request shape while retaining the
legacy prompt shape for older models. Both raw PCM and complete WAV replies
are converted to the existing MP3 output.
The new models are adapter-supported choices, not proof that the configured account can use them. A provider listing confirms availability. Google's official pricing and speech guide were verified on 2026-10-27 once documentation access became available:
| Model | Input / audio output per 1M tokens through 2026-12-31 | From 2027-01-01 |
|---|---|---|
| Gemini 3.8 Flash TTS | $0.50 / $9 | $1 / $18 |
| Gemini 3.8 Flash-Lite TTS | $0.50 / $6 | $1 / $12 |
Flash TTS is the reviewed recommendation for long-form multi-speaker stability. Flash-Lite is the lower-cost option for bulk production. Existing explicit model selections survive migration. The adapter's legacy fallback remains available for callers that do not use task routing.
Generated speech uses 25 audio tokens per second when the provider supplies no token counts. The earlier input-audio approximation of 32 tokens per second overestimated output usage. Duration-derived counts remain marked estimated. Published introductory prices have explicit expiration dates, shared by the catalogue and usage recorder; they cannot silently remain permanent discounts.
The same review corrected Claude Opus 4.6 ($5 input / $25 output) and Haiku 4.5 ($1 / $5), and added published standard token rates for current OpenAI GPT-6 models, Claude Opus/Sonnet 5.5, Gemini 3.8 Flash, and Gemini 4 Argon. Sources: OpenAI pricing and Claude pricing. Foundry deployments, explicit model choices, and public-route provider rules remain in force; a newer public model is not evidence of an available Azure deployment. Token estimates do not include provider-specific cache discounts, long-context surcharges, grounding/tool fees, or account discounts. Live production account configuration and provider invoices were not accessed.
Interpret usage costs¶
Usage & Cost shows the last 200 recorded calls, not a complete monthly bill. Missing costs and unpriced models are excluded from subtotals and explicitly marked unknown. A confirmed zero-cost trial call remains distinct from an unknown price. Small nonzero costs retain enough precision to remain visible.
New calls require a model-specific confirmed rate rather than an unrelated provider default. Image callers supply their per-image cost explicitly. Speech estimates return unknown when the selected model has no confirmed rate. Existing stored cost rows are not rewritten by a refresh or a rate-table change; historical unknowns need separate reconciliation against provider usage or billing evidence.
Before revising a recommendation, verify its current API support and pricing, update the dated reason and model-specific rate together, then validate both the recommendation and the configured provider's availability. Documentation access alone does not establish account authorization or a live model response.