feat(levelcode): Gemini 4 Argon joins the catalog at Pro, at its introductory rate

[ndemianc]

Adds google/gemini-4-argon to Levelcode::ModelCatalog (D62). The editor's model picker reads GET /account/models live, so once this is deployed every Pro user can pick Argon. No editor release is needed.

The row

Price	$2/M in · $10/M out · $0.10/M cached read
Context	1,048,576 tokens
Tier	Pro and up; the free plan cannot reach it
Cost order	after Kimi K2.7 Code, before Sonnet 5: 2.61× against 2.67×
Turns on Pro	about 99 reference turns, against 260 on Kimi K2.7

The price was checked against the OpenRouter endpoints API on 2026-10-14. Google Vertex (global) and Google AI Studio both list exactly $2 / $10 / $0.10.

Re-price by 2026-11-30

This is Google's introductory rate, 50% off, and it ends on 2026-11-30. The standard rate is $4 / $20 / $0.20.

The row is also the OpenRouter max_price ceiling. If nobody re-prices it, OpenRouter refuses every Argon request from 1 December. That failure is loud and never a loss, and the fix is a one-line edit to the three prices plus the multiplier spec.

Until then, the ceiling keeps requests off Vertex US ($2.20 / $11) and both Priority endpoints ($3.60 / $18). The cheaper AI Studio Flex endpoint ($1 / $5) stays eligible; a request routed there is still billed at the catalog rate.

Why it can go in without an editor release

Argon always reasons. The editor's OpenAI-compatible agent path drops reasoning between tool calls, and OpenRouter says tool use needs reasoning preserved. So I ran the existing spike against google/gemini-4-argon on 2026-10-14 (extensions/levelcode-ai/scripts/kimi-k3-spike.js).

It exited 0. Turn 1 called the tool; turn 2 answered correctly with no reasoning carried back. A pass proves only that the request is accepted, so one real multi-step agent task in the editor is still due before announcing.

Tests

• What the specs pin: the context, the rates, the 2.61 multiplier, Pro reach, the free plan's lack of it (gateway_model('free', argon)), the $2 / $10 ceiling, and Argon in a Pro account's live roster (live: true, about 99 turns).
• Each was seen to fail when it should:
  • row removed: 7 failures
  • row moved after Sonnet 5: 1 failure (the monotonic-multiplier spec)
  • row priced at the standard rate: 5 failures
• Runs:
  • the four affected files: 81 examples, 0 failures
  • spec/models, spec/services/levelcode, spec/requests/api/levelcode and spec/requests/levelcode: 1,088 examples, 0 failures
  • rubocop: clean

Not done here

• Deploy. Rollback is deleting the row; anyone who picked Argon falls back to the plan default through gateway_model.
• The after-deploy checks:
  • Argon listed live: true for a Pro account and not for a free one
  • a plain chat and a screenshot in gateway mode
  • one multi-step agent task
  • one prompt above 200K tokens, to confirm no long-context price tier trips the ceiling
  • OpenRouter activity showing $2 / $10
  • BYOK with an OpenRouter key
• The pricing page's feature line still names only Kimi, Opus 5 and Fable 5 / 5.1. Adding Argon there is a copy decision, left for you.

[Copilot]

🔵 Needs a closer look

Multi-step agent compatibility and long-context pricing checks remain outstanding and require human validation.

0 open findings

What changed in this PR

Adds Gemini 4 Argon to the shared model catalog, making it available to Pro-and-up users through the live editor picker without an editor release.

Changes:

• Defines introductory pricing, context size, and Pro-tier access.
• Adds coverage for catalog values, plan eligibility, routing ceilings, and the live account roster.

File	Description
spec/​services/​levelcode/​open_router_adapter_spec.rb	Pins Argon’s introductory routing ceiling.
spec/​requests/​api/​levelcode/​v1/​account_spec.rb	Checks live availability, multiplier, and estimated turns.
spec/​models/​levelcode_plans_spec.rb	Verifies Pro access and free-plan fallback.
spec/​models/​levelcode_model_catalog_spec.rb	Pins context, pricing, multiplier, and entitlement.
app/​services/​levelcode/​model_catalog.rb	Adds Argon and documents the repricing deadline.

🧠 Review effort: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.