Gateway

Open and frontier models behind one endpoint, each call on the cheapest model that clears the bar

Content checks, search, drafts, metadata and dubs each get a primary model, an escalation and a fallback. Routing by genre and language becomes a rule you can read, price and re-test, not a habit.

One endpoint, every model

OpenAI-compatible. Change the base URL; existing code and custom GPT prompts keep working.
Open and frontier together. Open models served by Dvar from India, priced in rupees; frontier models through the same call.
Routes decided by evals. Each workflow, language or genre goes to the cheapest model that clears its eval. When a new model ships, the eval re-runs and the route changes without a code change.
Escalation and fallback. Unsure answers go up a tier; a provider outage falls back; search degrades to keyword results instead of failing.
Every call traced. Cost, latency and errors per call, with personal details masked, feeding the evals and later post-training.

from openai import OpenAI

client = OpenAI(
    base_url="https://portal.dvarlabs.ai/v1",     # was the provider's URL
    api_key=os.environ["DVAR_API_KEY"],
)
client.chat.completions.create(
    model="auto",                                 # the gateway picks per workflow
    extra_headers={                               # routing hints (illustrative)
                   "X-Dvar-Workflow": "content-checks",
                   "X-Dvar-Language": "hi", "X-Dvar-Genre": "horror"},
    messages=[...],
)

The monthly bill, per workflow

List prices times a token shape per call. Volumes and token shapes are written for the demo; type your own. The content-check escalation share comes from the frozen run.

Workflow	Route	Calls a month	Per call	Frontier only	This setup
Content checks Five checks on every scene of every draft, dub and subtitle file	Primary GLM 5.3 on Dvar · escalate 1% to claude-sonnet-5.5 (Dvar model confidence under 0.7 (share from the run)) · fallback gpt-oss-120b on Dvar Short, repeated, structured: a small open model with code checks in front, frontier only when unsure.	about 2,000 shows × 80 episodes × 25 scenes × 5 passes (drafts, regenerations, dubs)	650 in · 40 out ₹0.14 frontier ₹0.08 primary	₹28.6 lakh	₹16.6 lakh 1.7× lower
Search query reading Turn a typed query into genre, tags, language and a title guess	Primary gpt-oss-120b on Dvar · fallback keyword results Must answer inside the search latency budget; on timeout the gateway returns keyword results, never an error.	searches a month that need more than keyword matching (assumed)	420 in · 45 out ₹0.11 frontier ₹0.0060 primary	₹32.5 lakh	₹1.80 lakh 18× lower
Episode script drafts First drafts of episode scripts from a beat sheet and the show bible	Primary claude-sonnet-5.5 · fallback GLM 5.3 on Dvar Creative work stays on frontier until the generator gate shows an open or post-trained model clears the bar.	about 2,000 shows × 80 episodes × 2.5 drafts	3,000 in · 1,200 out ₹1.51 frontier ₹1.51 primary	₹6.05 lakh	₹6.05 lakh same
Titles, synopses, thumbnail and push copy Short marketing text per episode, in several variants	Primary gpt-oss-120b on Dvar · fallback GLM 5.3 on Dvar High volume, short output, easy to check: open model first.	160,000 episodes × 10 variants	1,200 in · 250 out ₹0.41 frontier ₹0.02 primary	₹6.59 lakh	₹35,200 19× lower
Translation for dubs and subtitles Hindi scripts into other Indian languages and English	Primary GLM 5.3 on Dvar · escalate 10% to claude-sonnet-5.5 (language check fails on the output) · fallback gpt-oss-120b on Dvar Open multilingual model, with the content checks run on the output and frontier for retries.	160,000 episodes × 3 languages	1,500 in · 1,700 out ₹1.68 frontier ₹0.74 primary	₹8.06 lakh	₹4.34 lakh 1.9× lower
All workflows				₹81.8 lakh	₹29.1 lakh 2.8× lower

Episode drafts stay on frontier here on purpose: the generator gate decides when an open or post-trained model clears the bar for each language and genre, and the route changes then.

A simulated day through the gateway

One day at the default volumes, sampled 1 in 1,000 and scaled back up. Turn on an incident to see where calls go.

Frontier provider returns errors, 14:00–15:00Search model slow (900 ms), 21:00–22:00

00:0012:0023:00

PrimaryEscalatedFallbackDegraded, no model

Workflow	Calls	Escalated	Cost (if frontier only)
Content checks	6,64,000	7,000	₹55,348 (₹94,819)
Search query reading	9,97,000	0	₹5,982 (₹1.08 lakh)
Episode script drafts	11,000	0	₹16,632 (₹16,632)
Titles, synopses, thumbnail and push copy	60,000	0	₹1,320 (₹24,696)
Translation for dubs and subtitles	11,000	0	₹8,091 (₹18,480)