Updating Requesty models

[requesty-JohnCosta27]

• Adding various models
• Using @requesty/ai-sdk instead of openai-compatible

[github-actions]

Action items

• [critical] [violation] providers/requesty/provider.toml:3 - Check: Provider schema allows api only for @ai-sdk/openai-compatible, @openrouter/ai-sdk-provider, and optionally @ai-sdk/openai / @ai-sdk/anthropic / kiro-acp-ai-provider; other npm packages must omit api. Why: This PR sets npm = "@requesty/ai-sdk" while keeping api = "https://router.requesty.ai/v1". That combination fails the provider refine in packages/core/src/schema.ts and will not validate. Action: Either keep npm = "@ai-sdk/openai-compatible" with api, or switch to @requesty/ai-sdk and remove api (matching custom-SDK providers like aihubmix/merge-gateway). Do not ship both as written.
• [high] [possible mistake] providers/requesty/models/xiaomi/mimo-v2.5.toml:2 - Check: Provider entries must not incorrectly clear reasoning when the underlying model reasons. Why: Both entries set reasoning = false while models/xiaomi/mimo-v2.5.toml and mimo-v2.5-pro.toml define reasoning = true, and peer providers (OpenRouter, Vercel, Kilo, etc.) keep them as reasoning models with toggle/[]. This publishes a false non-reasoning capability. Action: Drop the reasoning = false overrides; keep inherited reasoning = true and set reasoning_options = [] (or a verified Requesty control). Apply the same fix to providers/requesty/models/xiaomi/mimo-v2.5-pro.toml.
• [medium] [possible mistake] providers/requesty/models/openai/gpt-5.5.toml:2 - Check: reasoning_options must match Requesty’s verified request surface, not upstream OpenAI enums alone. Why: New OpenAI/Azure/Bedrock GPT entries claim effort values including xhigh, but providers/requesty/provider.toml documents OpenAI mapping as only low|medium|high passthrough plus max/none/min translations, and every existing Requesty GPT entry omits xhigh. Action: Verify whether Requesty accepts/forwards xhigh for these model IDs; if not, remove it and align with the documented Requesty effort set (and existing Requesty GPT peers).
• [medium] [possible mistake] providers/requesty/models/minimaxi/MiniMax-M2.7.toml:2 - Check: Provider overrides of attachment / modalities must match what Requesty actually serves. Why: MiniMax-M2.5, MiniMax-M2.7, and MiniMax-M2.7-highspeed set attachment = true and input = ["text", "image"], but their models/minimax/* bases and the native MiniMax provider are text-only (attachment = false, input = ["text"]). Action: Confirm vision support on these Requesty routes; if unsupported, remove the attachment/image overrides and inherit the base modalities.
• [medium] [possible mistake] providers/requesty/models/nvidia/nemotron-3-super-120b-a12b.toml:5 - Check: Material cost claims need provider evidence. Why: Super/ultra are priced input = 0 / output = 0, while the NVIDIA provider lists non-zero rates (e.g. super 0.20/0.80, ultra 0.50/2.50). Zero pricing is plausible only if Requesty truly serves them free. Action: Cite Requesty pricing for these IDs, or correct costs to the billed USD/1M rates. Same check for nemotron-3-ultra-550b-a55b.toml.
• [low] [possible mistake] providers/requesty/models/openai/gpt-5.5.toml:4 - Check: Long-context tier pricing should match billed rates when base list prices are copied. Why: New GPT-5.4/5.5/5.6 entries copy OpenAI base rates (e.g. GPT-5.5 5/30/0.50) but omit the OpenAI 272_000 context tier that OpenAI and existing Requesty gpt-5.4 publish at 2×. That understates cost above the tier threshold. Action: Add matching [[cost.tiers]] where Requesty bills the higher tier, or cite evidence that Requesty is flat-rate.
• [low] [violation] PR body - Check: Data-changing PRs should cite direct provider sources for material claims. Why: Body only says models were added and the SDK changed; no Requesty pricing/docs/API links mapped to costs, limits, modalities, or reasoning controls across ~80 new models. Action: Add first-party citations (Requesty model/pricing pages and reasoning docs) stating what each supports.

[github-actions]

Action items

• [high] [possible mistake] providers/requesty/models/minimaxi/MiniMax-M2.5.toml:778 - Check: Provider overrides must not invent modalities the underlying model does not support. Why: These MiniMax entries set attachment = true and modalities.input = ["text", "image"], but models/minimax/MiniMax-M2.5 / M2.7 / M2.7-highspeed and the native MiniMax provider are text-only (attachment = false). Clients would be told vision is supported when it is not. Same issue in MiniMax-M2.7.toml and MiniMax-M2.7-highspeed.toml. Action: Drop the attachment/image overrides unless you can cite Requesty evidence that these routes accept images; otherwise inherit the text-only base modalities.
• [high] [possible mistake] providers/requesty/models/zai/GLM-5.toml:2162 - Check: Provider overrides must not invent modalities the underlying model does not support. Why: GLM-5, glm-5.1, and glm-5.2 set attachment = true and image input, but models/zhipuai/glm-5* and native Z.ai glm-5 are text-only (vision is a separate glm-5v* family). Action: Remove the attachment/image overrides unless Requesty documents image input for these exact model IDs.
• [medium] [possible mistake] providers/requesty/models/openai/gpt-5.5.toml:1496 - Check: reasoning_options must match Requesty’s verified effort surface, not native OpenAI alone. Why: New OpenAI/Azure/Bedrock GPT entries claim xhigh (and often none|…|xhigh|max), but existing Requesty peers (e.g. openai/gpt-5.4.toml) use none|low|medium|high|max, and providers/requesty/provider.toml documents OpenAI mapping for none|min|low|medium|high|max with no xhigh. Also, native GPT-5.5 Pro is medium|high|xhigh only, so the proposed Pro enum is inconsistent both ways. Affected paths include openai/gpt-5.4-*, gpt-5.5*, gpt-5.6-*, and matching azure/ / bedrock/ files. Action: Align effort values with Requesty’s documented/gateway mapping (or update the provider reasoning note with evidence that xhigh is accepted), and special-case Pro if its allowed set differs.
• [medium] [possible mistake] providers/requesty/models/xiaomi/mimo-v2.5.toml:2125 - Check: Reasoning controls should match this provider surface and comparable gateways for the same model. Why: Entries claim OpenAI-style effort + budget_tokens, but native Xiaomi and OpenRouter (OpenAI-compatible gateway peer) expose toggle only for MiMo V2.5. Action: Change to toggle (or [] if Requesty does not forward Xiaomi’s toggle) and drop unverified effort/budget claims; apply the same fix to mimo-v2.5-pro.toml.
• [medium] [possible mistake] providers/requesty/models/azure/gpt-5.4-mini.toml:363 - Check: Multimodal capability overrides should match the underlying model unless the route truly strips them. Why: Overrides attachment = false and modalities.input = ["text"] while models/openai/gpt-5.4-mini is image-capable (attachment = true, input = ["text", "image"]). Action: Verify Azure-via-Requesty actually rejects images; if not, remove these overrides.
• [medium] [possible mistake] providers/requesty/models/tensorx/deepseek-v4-flash.toml:1683 - Check: Same-provider routes for the same underlying model should not disagree on reasoning controls without evidence. Why: tensorx/deepseek-v4-* claim effort = high|xhigh, while deepseek/deepseek-v4-* and sference/deepseek-v4-flash use reasoning_options = [] on the same Requesty API. OpenRouter’s DeepSeek V4 entries use high|xhigh, so the empty arrays may underclaim or tensorx may overclaim. Action: Verify Requesty’s DeepSeek reasoning field once and align all DeepSeek V4 routes to the same options.
• [medium] [possible mistake] providers/requesty/models/nvidia/nemotron-3-super-120b-a12b.toml:961 - Check: Published costs should reflect what Requesty charges. Why: Super and Ultra are priced input = 0 / output = 0, but native NVIDIA lists non-zero rates (e.g. Super ~$0.20/$0.80, Ultra ~$0.50/$2.50). Free is plausible for nano; free Super/Ultra needs confirmation. Action: Confirm Requesty pricing for these IDs and correct non-zero rates if they are not free.
• [low] [possible mistake] providers/requesty/models/openai/gpt-5.5.toml:1498 - Check: Long-context GPT pricing tiers when peers document them. Why: Existing Requesty openai/gpt-5.4.toml and native OpenAI GPT-5.4/5.5 entries include cost.tiers at 272K context; new gpt-5.4*, gpt-5.5*, and Azure/Bedrock copies omit tiers, so >272K pricing may be wrong. Action: Add matching tiers where Requesty bills the higher long-context rates, or cite that Requesty is flat-rate.
• [low] [violation] .pr-review/pull-request.json:4 - Check: Data-changing PRs should cite first-party sources in the PR body. Why: Dozens of new models, prices, limits, and modalities are introduced with only a two-line summary and no pricing/docs/API links mapped to claims, so reviewers cannot validate material facts. Action: Add Requesty model catalog/pricing/API citations and state what each link supports (prices, modalities, reasoning controls, deletions).