Summary

Updates the Neon provider to match what the Neon AI Gateway actually serves today. Neon's gateway is powered by Databricks Foundation Model APIs, so I took the current Databricks supported-models list (docs updated 2026-07-02) as the candidate set and probed every model live against a real Neon AI Gateway branch (us-east-2), then live-checked capabilities. The current Neon entries here were missing 11 working models and listed one (gpt-5-5) that the gateway rejects.

Neon catalog id = Databricks pay-per-token endpoint name minus the databricks- prefix (e.g. claude-sonnet-4-6 → underlying global.anthropic.claude-sonnet-4-6).

Added (12) — all verified working via live calls

Proprietary (pass-through provider list pricing), via base_model:

• gpt-5-3-codex, gpt-5-2-codex, gpt-5-1-codex-max, gpt-5-1-codex-mini (OpenAI Responses API)
• claude-opus-4-8
• gemini-3-5-flash

Open-weight (priced from Databricks pay-per-token DBU rate × $0.070/DBU):

• llama-4-maverick — $0.50/$1.50
• meta-llama-3-3-70b-instruct — $0.50/$1.50
• meta-llama-3-1-8b-instruct — $0.15/$0.45
• qwen35-122b-a10b — $0.22/$2.20
• qwen3-next-80b-a3b-instruct — $0.15/$1.20
• gemma-3-12b — $0.15/$0.50

Removed (1)

• gpt-5-5 — the gateway returns 400 unknown model "gpt-5-5" (any prefix / dialect). Not currently served by Neon.

How this was verified

For each candidate I sent real requests to the gateway and recorded the result:

• works? — minimal completion (chat-completions dialect, or the OpenAI Responses dialect for the codex models). All 12 above returned 200.
• image input (attachment) — sent a generated 32×32 PNG data-URI + "what color?". 200 ⇒ supported; 400 does not support image ⇒ not.
• tool calling (tool_call) — sent a get_weather function tool and checked for an emitted tool call. All 12 support tool calling.

Capability corrections applied so the entries match the gateway (not just the canonical base model):

• qwen35-122b-a10b: canonical Qwen3.5 is multimodal, but Databricks/Neon serve it text-only (live: no image input). Overridden to attachment = false, modalities.input = ["text"], and limit.output = 8000 per the Databricks doc.
• meta-llama-3-3-70b-instruct: overridden to attachment = false (text-only on the gateway; live-confirmed).

Notes / open questions for maintainers

• The databricks:generate script currently IGNOREs llama / qwen / gemma prefixes, so the Databricks provider omits them. These do work on Neon, so I've added them here with Databricks DBU-derived pricing — happy to drop them if you'd prefer Neon to mirror that ignore-list.
• Open-weight pricing uses the published Databricks pay-per-token DBU rates × $0.070/DBU (the rate that reproduces the headline GPT-OSS-120B $0.15/$0.60). The pre-existing gpt-oss-120b/gpt-oss-20b Neon entries appear to use OpenRouter pricing instead (e.g. gpt-oss-120b = $0.072/$0.28 rather than Databricks' $0.15/$0.60) — left unchanged here, but flagging the inconsistency.
• Also verified not available on Neon and intentionally not added: claude-sonnet-5, claude-fable-5, meta-llama-3-1-405b-instruct (unknown model), the Gemini image-output models, and all embedding models (the gateway exposes no /embeddings route).

Test plan

• bun run validate-equivalent: parsed each new TOML, resolved base_model merges, and confirmed required fields (name, attachment, reasoning, tool_call, open_weights, release_date, last_updated, cost.{input,output}, limit.{context,output}, modalities.{input,output}). All 36 Neon models valid.
• CI schema validation (GitHub Action).

Update — reasoning controls & fast mode corrected (live re-verified)

A maintainer asked (via DM) whether Neon actually serves Opus fast mode and how the qwen35 reasoning controls work. I re-probed the live gateway (us-east-2) and corrected two entries in this PR:

claude-opus-4-8

• Removed [experimental.modes.fast]. The gateway rejects fast mode on every request shape:
  • speed:"fast" + anthropic-beta: fast-mode-2026-02-01 on /ai-gateway/anthropic/v1/messages → 400 "speed: Extra inputs are not permitted"
  • beta header only → 400 "invalid beta flag"
  • same via /ai-gateway/mlflow/v1/chat/completions → 400 "speed: Extra inputs are not permitted"
  • (Opus 4.7 and 4.6 behave identically — fast mode is not exposed by Neon at all.)
• reasoning_options: toggle + budget_tokens → toggle + effort. thinking.type:"enabled" + budget_tokens returns 400 '"thinking.type.enabled" is not supported for this model. Use "thinking.type.adaptive" and "output_config.effort" to control thinking behavior.' The gateway accepts thinking.type:"disabled" (toggle off, 200) and thinking.type:"adaptive" + output_config.effort. Its own validation error enumerates the accepted effort values: low, medium, high, xhigh, max (negative control effort:"banana" → 400 "unknown variant 'banana', expected one of 'low','medium','high','xhigh','max'"). none/minimal/default are rejected.

qwen35-122b-a10b

• reasoning_options: toggle + budget_tokens → effort (none/low/medium/high). On the mlflow chat-completions route, thinking, enable_thinking, and any reasoning-budget field all return 400 "unknown field". The real control is reasoning_effort; the gateway's validation error enumerates the accepted values (parameter "reasoning_effort" must be "none", "low", "medium" or "high"). reasoning_effort:"none" returns a plain, non-reasoning completion (reasoning off); the other values return type:"reasoning" blocks. There is no separate budget or enable_thinking toggle — none is the off state.

All checks were minimal live requests against a real Neon AI Gateway branch; no budget_tokens control exists for either model on Neon today.

claude-opus-4-7 and claude-opus-4-6 (pre-existing entries, corrected here too)

Both carried the same [experimental.modes.fast] block, and the gateway 400s fast mode on both (speed: Extra inputs are not permitted) — removed from both.

Reasoning differs between the two:

• claude-opus-4-7 behaves like 4-8: thinking.type:"enabled" + budget_tokens → 400 (must use adaptive + output_config.effort); disabled toggles off; effort enum low/medium/high/xhigh/max (verified, incl. effort:"max" producing thinking tokens). → reasoning_options changed from toggle + budget_tokens to toggle + effort.
• claude-opus-4-6 still supports the classic API: thinking.type:"enabled" + budget_tokens=1024 returns thinking blocks (200), and disabled toggles off. → reasoning_options left unchanged (toggle + budget_tokens is correct); only the unsupported fast mode block was removed.