Add opper provider

[Felixkw12]

Adds Opper (https://opper.ai) — an EU-hosted AI gateway serving 700+ models from 30+ providers through an OpenAI-compatible endpoint at https://api.opper.ai/v3/compat. (Disclosure: I work at Opper.)

What's included

provider.toml — @ai-sdk/openai-compatible, OPPER_API_KEY
logo.svg (currentColor)
43 flagship models as base_model references (Anthropic, OpenAI, Gemini, xAI, Mistral, Moonshot, MiniMax, Meta, Perplexity)

Data sourcing

Pricing is mirrored from our public catalog endpoint GET /v3/compat/models (OpenAI models.list shape with per-token pricing and context length). Opper doesn't mark up provider token rates, so costs match upstream list prices.
reasoning_options were measured against POST /v3/compat/chat/completions on 2026-08-19: effort strings pass through to the upstream API unchanged, so each model carries its native set (details in the provider.toml comment). Budget-token values are not accepted on this surface.
The full live catalogue (612 models) is discoverable via /v3/compat/models; this PR curates current-generation flagships.

bun run validate passes locally.

[github-actions]

Action items

• [high] [violation] providers/opper/models/anthropic/claude-sonnet-5.toml:4 - Check: Every toggle must have a leading top-of-file wire comment for this host’s request path. Why: claude-sonnet-5, minimax/m3, and moonshot/kimi-k3 declare toggle, but none has a top-of-file wire comment, and provider.toml only documents passthrough reasoning_effort (no on/off field). Callers cannot tell how to disable reasoning on Opper’s OpenAI-compat surface. Action: For each toggle model, add a leading comment with the exact Opper wire path (e.g. body field/values), or remove toggle if Opper does not expose a real on/off control on /v3/compat.
• [high] [possible mistake] providers/opper/models/openai/gpt-5.3-codex.toml:10 - Check: Provider limit overrides must reflect this host’s real context window, not a different limit field. Why: Opper sets context = 272000, but lab/OpenAI gpt-5.3-codex uses context = 400_000 with input = 272_000. This looks like the input cap was written as total context, understating the window. Action: Verify against Opper’s catalog/API and set context (and input if needed) to the real values; do not use 272k as context unless Opper truly serves a 272k window.
• [high] [possible mistake] providers/opper/models/openai/gpt-5.3-chat-latest.toml:9 - Check: Provider context overrides should not invent a larger window than lab/first-party without evidence. Why: Opper sets context = 272000 while lab and OpenAI first-party use 128_000 for gpt-5.3-chat-latest. Same 272k figure as Codex suggests a copy/paste of the GPT input limit. Action: Confirm Opper’s served context for this model ID and correct limit.context (likely 128_000 unless Opper documents otherwise).
• [medium] [possible mistake] providers/opper/models/gemini/gemini-flash-latest.toml:8 - Check: Costs should match Opper’s catalog when claiming no markup / upstream list prices. Why: Opper prices gemini-flash-latest at 0.3 / 2.5 / 0.03, which matches Google gemini-3.5-flash-lite, not the Flash-latest alias peers (1.5 / 9 / 0.15 on Google/Vertex). Action: Re-check GET /v3/compat/models for this ID and align cost with the model actually served (or document if Opper intentionally prices the alias differently).
• [medium] [possible mistake] providers/opper/models/gemini/gemini-flash-lite-latest.toml:8 - Check: Same no-markup pricing claim vs first-party Flash-Lite latest. Why: Opper uses 0.1 / 0.4 / 0.01, while Google/Vertex Flash-Lite latest is 0.25 / 1.5 / 0.025. Large gap needs confirmation. Action: Verify Opper catalog pricing for this ID and correct cost if the values were mis-mapped.
• [low] [possible mistake] providers/opper/models/xai/grok-4.5.toml:10 - Check: Cache pricing consistency with first-party when claiming upstream rates. Why: Opper cache_read = 0.5 vs xAI first-party 0.3 (and several peers). Output/input match; only cache differs. Action: Confirm Opper’s published cache rate for grok-4.5 and fix if it should be 0.3.

[github-actions]

Action items

• [medium] [possible mistake] providers/opper/models/gemini/gemini-3.6-flash.toml:8 - Check: Costs must match the host’s real USD/MTok rates; PR claims no markup vs upstream list. Why: Opper authors input = 0.75 / output = 3.75 / cache_read = 0.075, but first-party Google/Vertex list this model at 1.50 / 7.50 / 0.15. That is half the lab rate and equals Gemini 3.7 Flash pricing in-repo, which is inconsistent with “match upstream list prices” unless Opper’s catalog truly bills the discounted rate. Action: Confirm against Opper GET /v3/compat/models for this ID and either keep the discounted figures with a brief note, or correct to the billed rates (including cache).
• [medium] [possible mistake] providers/opper/models/minimax/m3.toml:5 - Check: Context-based pricing should use [[cost.tiers]] when the upstream list is tiered. Why: First-party MiniMax-M3 is 0.30/1.20 (cache 0.06) with a 512_000 tier at 0.60/2.40 (cache 0.12). Opper publishes only the higher band as flat base cost, which overstates short-context price if Opper actually uses MiniMax’s tier schedule, or under-documents a flat high rate if that is intentional. Action: Verify Opper’s billed schedule; author base + [[cost.tiers]] if tiered, or keep a single band only if the catalog/bills that flat rate for all contexts.
• [medium] [possible mistake] providers/opper/models/openai/gpt-5.4.toml:5 - Check: When claiming costs match upstream list prices, include documented context tiers. Why: OpenAI list (and peers such as OpenRouter) use a 272_000 context tier for GPT-5.4 / 5.5 / 5.6-* (and pro variants); xAI Grok 4.3/4.5/4.6 and Google Gemini 3.1 Pro similarly have 200_000 tiers in first-party entries. Opper only publishes the base band for these families, so long-context cost is wrong if Opper truly passes through list pricing. Action: For each tiered model, confirm Opper billing and add matching [[cost.tiers]] (or document that Opper bills a single flat band).

[github-actions]

Action items

• [high] [possible mistake] providers/opper/models/minimax/m3.toml:11 - Check: Context-tier pricing must be reachable under the resolved context window, and Opper rates should match the claimed no-markup catalog schedule. Why: After patch 3 dropped the limit.context override, this entry inherits lab context = 512_000, but [[cost.tiers]] starts at 524_288, so the higher band can never apply. Base rates (0.6 / 2.4) also match first-party MiniMax’s extended band, not its base (0.3 / 1.2 at providers/minimax/models/MiniMax-M3.toml), with a further doubled tier above that. Action: Reconcile against GET /v3/compat/models: restore a real context override if Opper serves >512k, and/or fix base + tier thresholds/prices so the schedule is consistent and usable.
• [medium] [possible mistake] providers/opper/models/openai/gpt-5.5-pro.toml:5 - Check: Context-tier pricing for GPT-5.5 family models Opper claims to bill on the upstream schedule. Why: Patch 4 adds [[cost.tiers]] at 272_000 for gpt-5.4, gpt-5.4-pro, gpt-5.5, and the gpt-5.6-* entries, and first-party providers/openai/models/gpt-5.5-pro.toml has the same 272k band (60 / 270). This file only has flat 30 / 180. Action: Add the matching [[cost.tiers]] band if Opper bills it, or document in a leading comment why Pro is flat while sibling GPT-5.4/5.5/5.6 Opper entries are tiered.

[github-actions]

Action items

• [medium] [possible mistake] providers/opper/models/minimax/m3.toml:5 - Check: Provider costs should match this host’s billed rates; PR claims no markup over upstream list prices. Why: Opper’s flat band is 0.6 / 2.4 / 0.12 with a higher band at 1.2 / 4.8 / 0.24, while first-party MiniMax is 0.30 / 1.20 / 0.06 with the upper band 0.60 / 2.40 / 0.12 at ~512k. That looks like the upper MiniMax band was used as the base and then doubled again. Action: Confirm against Opper’s live GET /v3/compat/models (and billing) that these are the real charged rates; if not, correct the base/[[cost.tiers]] values (and note any real markup).
• [low] [possible mistake] providers/opper/models/gemini/gemini-3.6-flash.toml:7 - Check: Same no-markup claim vs first-party Google pricing. Why: Opper lists 0.75 / 3.75 / 0.075, but first-party providers/google/models/gemini-3.6-flash.toml is 1.50 / 7.50 / 0.15 (Vertex matches that). Several relays use the lower figure, so this may be intentional host pricing, but it conflicts with the PR’s “matches upstream list prices” statement. Action: Verify the Opper catalog/billing figure and either keep it with a brief leading note that it differs from Google list, or correct it if the catalog was misread.

[github-actions]

Action items

• [medium] [possible mistake] providers/opper/models/xai/grok-4.3.toml:7 - Check: Context-tier pricing when claiming no markup vs upstream. Why: First-party xAI and established peers bill a 200k band (2.5/5, cache 0.4) on top of 1.25/2.5; this entry is flat-only while grok-4.6 on Opper already carries the upstream tier. That conflicts with “no gateway markup / match list prices” unless Opper’s catalog truly flattens only some Grok IDs. Action: Confirm against GET /v3/compat/models; add [[cost.tiers]] at 200k if billed, or add a leading comment documenting flat billing (as on minimax/m3 / gemini-3.6-flash).
• [medium] [possible mistake] providers/opper/models/xai/grok-4.5.toml:7 - Check: Context-tier pricing vs first-party/peers. Why: Lab/OpenRouter/xAI use a 200k tier (4/12, cache 0.6) with base cache_read 0.3; this file is flat at cache_read 0.5 with no tier, while Opper’s own grok-4.6 includes the tier schedule. Action: Verify Opper’s live schedule; add the 200k band (and correct cache rates) if present, or document intentional flat billing in a leading comment.
• [medium] [possible mistake] providers/opper/models/xai/grok-build-0.1.toml:5 - Check: Context-tier pricing vs first-party/peers. Why: xAI and peers bill 2/4 (cache 0.4) above 200k on this always-on reasoner; Opper lists only base 1/2/0.2. Same inconsistency as the other Grok gaps under a no-markup claim. Action: Confirm catalog tiers and add [[cost.tiers]] or a leading flat-billing note.
• [medium] [possible mistake] providers/opper/models/gemini/gemini-3.6-flash.toml:1 - Check: Model ID ↔ billed rate consistency. Why: Opper’s 0.75/3.75 (cache 0.075) matches first-party Gemini 3.7 Flash, not 3.6 Flash (1.50/7.50). The PR already dropped gemini-flash-latest for alias/price mismatch; shipping 3.6 under 3.7 list rates risks a wrong catalog mapping. Action: Re-check that this Opper model ID is actually 3.6 (not a mislabeled 3.7 route). If the catalog is wrong, omit the entry until fixed; if the discounted 3.6 rate is real, keep the leading note and cite the catalog field that pins ID + price together.

[github-actions]

No actionable findings.

[github-actions]

No actionable findings.