Short answer
Use ChatGPT to generate a tightly‑prompted 90‑second script, then synthesize it with ElevenLabs. Orchestrate with a scheduled serverless job (or no‑code automation) that writes, synthesizes, stores, and publishes. Expect per‑brief costs in the low cents (example below) and typical end‑to‑end latency under 30 seconds if you synthesize on demand — much lower if you pre‑generate.

Recommended orchestration (practical pipeline)
1) Scheduler: Cron on a small server, Cloud Run / Lambda with EventBridge, or a no‑code scheduler (Make, n8n, Zapier).
2) Script generation: Call ChatGPT API with a concise prompt template and strict length constraints (target ~225 words for 90s). Validate/readability check.
3) TTS: Send the cleaned script to ElevenLabs TTS API (select voice, adjust prosody/SSML if needed).
4) Postprocess: Normalize audio (LUFS), convert to MP3/AAC, add metadata/ID3, save to S3 or CDN.
5) Publish: Push to CMS, podcast feed, or websocket/endpoints for apps.
6) Monitoring: Log generation times, API errors, and cost metrics; keep a fallback pre‑recorded intro in case TTS fails.

Prompt tips for ChatGPT
- Use a fixed template: context + headline + 3 bullets + “Produce a neutral, 90‑second spoken script ~225 words.”
- Enforce “Do not exceed X words/characters.”
- Ask for plain text only (no markup), then run a sanitizer before TTS.

Latency expectations (typical ranges)
- ChatGPT text generation: ~0.5–6 s depending on model and network.
- ElevenLabs TTS synthesis for ~90s script: ~2–20 s (model and concurrency matter).
- Total on‑demand: typically 3–30 s. If you pre‑generate nightly, user‑facing latency = CDN delivery time (sub‑second).

Cost per brief — how to estimate + worked example
Formula to compute:
- tokens ≈ characters / 4 (English) — include prompt + response.
- chat_cost = (tokens/1000) × model_price_per_1k_tokens.
- eleven_cost = (characters / 1,000,000) × eleven_price_per_1M_characters.
Example (assumptions for illustration):
- Script ≈225 words ≈ 1,200 characters ≈ 300 tokens; prompt adds 100 tokens → total 400 tokens.
- If model = $0.03 per 1k tokens → chat_cost ≈ 0.4×$0.03 = $0.012.
- If ElevenLabs = $10 per 1M chars → eleven_cost ≈ 1,200/1,000,000×$10 ≈ $0.012.
- Total ≈ $0.024 per brief. Adjust up if you use premium models or higher TTS rates.

Decision criteria (choose based on):
- Budget: cheaper models + batch generation overnight for lowest cost.
- Output quality: premium ChatGPT + ElevenLabs voices for highest naturalness.
- Latency: pre‑generate for near zero user latency; synthesize on demand for freshness.
- Team size / workflow stage: use no‑code for single operator; build serverless/queue architecture for teams or scale.

Best‑for / Avoid‑if
- Best for: daily briefs, push notifications, newsletter audio, low-latency updates with high voice quality.
- Avoid if: you need real‑time multi‑speaker conversation (not a single script), or strict legal/regulatory guarantees for spoken claims without human review.

Practical checklist to launch
- [ ] Define prompt template and word cap (~225 words).
- [ ] Implement scheduled function (Lambda/Cloud Run or Make/n8n) to call ChatGPT.
- [ ] Sanitize output and call ElevenLabs TTS with chosen voice/SSML.
- [ ] Normalize, store to S3/HTTP CDN, add metadata.
- [ ] Monitor latency, errors, and cost; add retry/fallback audio.

Recommendation
Start with a serverless scheduled job (once daily or hourly) using ChatGPT for script + ElevenLabs for TTS, pre‑generate the audio and host it on a CDN. Tune model and voice for your budget and quality targets, monitor costs, and move to on‑demand generation only if real‑time freshness is required.