the-grid-ai: enrich inline provider models (reasoning options, limits, interleaved)

[Cahl-Dee]

Summary

Updates the existing providers/the-grid-ai/models/*.toml definitions in place. Per @rekram1-node's feedback, The Grid AI is a model aggregator (it serves other labs' models under tier aliases, not its own models), so the files stay fully inline under providers/ — no models/ metadata layer and no base_model references.

What changed

• reasoning_options now reflect the actual API surface (source: GET https://api.thegrid.ai/v1/models + the contract spec):
  • *-standard / *-prime → [{ type = "effort", values = ["low", "medium", "high"] }] (reasoning is mandatory on these tiers; no toggle).
  • *-max → [{ type = "toggle" }, { type = "effort", values = ["low", "medium", "high", "xhigh", "max"] }] (reasoning can be turned off, and the wider effort set is accepted).
• limit.input added to all 9 models (previously only context + output).
• [interleaved] field = "reasoning_content" added to all 9 models — every instrument returns reasoning tokens in the reasoning_content field.
• Descriptions refreshed to describe each tier, with a note that any model meeting the contract spec can serve the request.
• Pricing: The Grid is a live market, so token cost is variable and not a fixed number. Each file carries a comment explaining this and linking to the instrument's live market page (in lieu of a static cost). A contract-spec link sits under each description.
• provider.toml: refreshed the HTTP reasoning-controls comment to match the current API (reasoning_effort enum, per-tier mandatory vs. toggle behavior).

Data source

• Catalog / contract spec: https://thegrid.ai/docs/instrument-specifications/current-instruments (accessed 2026-07-06)
• Per-instrument live pricing: https://app.thegrid.ai/trading/market-<id> (linked in each file)

Validation

• bun validate → green (all 9 files)
• bun run compare:migrations → clean
• bun --cwd packages/web run build → green

[rekram1-node]

afaik grid ai isnt a model lab, you are providing other models under certain aliases right? So this would be incorrect it should stay as it was

[Cahl-Dee]

Understood!

Some tools don't understand the nuance between a "model" and what we offer. To them, the functional experience is what matters. You send a request to an endpoint with a "model" parameter, and get a response. The value passed in that parameter, to them, is a model.

Totally get your clear distinction for models.dev / opencode. Addressing now.

[Cahl-Dee]

Addressed. Reverted the file-structure change — everything stays inline under providers/the-grid-ai/ with no models/ metadata layer or base_model references, since The Grid is an aggregator rather than a model lab.

Kept the substantive data fixes in place: corrected reasoning_options per tier (effort-only on *-standard/*-prime, toggle + wider effort set on *-max), added limit.input and [interleaved] field = "reasoning_content" to all 9 instruments, and documented the variable/market-based pricing via comments linking each instrument's live market page. bun validate and the web build are green.