feat(provider): add Y-API

providers/y-api/models/deepseek/deepseek-v4-flash-0731.toml

@@ -0,0 +1,14 @@
+# Toggle: thinking.type = enabled|disabled
+# Effort: reasoning_effort = low|high|max
+base_model = "deepseek/deepseek-v4-flash-0731"
+reasoning_options = [
+  { type = "toggle" },
+  { type = "effort", values = ["low", "high", "max"] },
+]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.0075
+output = 0.015

providers/y-api/models/deepseek/deepseek-v4-flash.toml

@@ -0,0 +1,15 @@
+# Toggle: thinking.type = enabled|disabled
+# Effort: reasoning_effort = low|high|max
+# Cost: free tier — no credit is deducted (https://y-api.bestvirtualgoods.com/pricing).
+base_model = "deepseek/deepseek-v4-flash"
+reasoning_options = [
+  { type = "toggle" },
+  { type = "effort", values = ["low", "high", "max"] },
+]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0
+output = 0

providers/y-api/models/deepseek/deepseek-v4-pro.toml

@@ -0,0 +1,22 @@
+# Toggle: thinking.type = enabled|disabled
+# Effort: reasoning_effort = high|max
+#
+# Interleaved field: this host does NOT emit reasoning_content for this model
+# (0 of 14 samples, 2026-09-16). It emits reasoning (string) plus
+# reasoning_details (list of one {type: "reasoning.text"}) on all 14 samples,
+# which is why the field below is reasoning_details and not the lab's
+# reasoning_content. The first-party DeepSeek V4 Pro entry uses
+# reasoning_content; this host's response shape differs, so the override
+# describes the host.
+base_model = "deepseek/deepseek-v4-pro"
+reasoning_options = [
+  { type = "toggle" },
+  { type = "effort", values = ["high", "max"] },
+]
+
+[interleaved]
+field = "reasoning_details"
+
+[cost]
+input = 0.025
+output = 0.05

providers/y-api/models/deepseek/deepseek-v4.1-flash.toml

@@ -0,0 +1,33 @@
+# Toggle: thinking.type = enabled|disabled
+# Effort: reasoning_effort = low|high|max
+#
+# This host ACCEPTS medium and xhigh (no 400) but does not act on them
+# distinctly, so they are not authored. Measured on this host, 2026-09-16:
+#
+#   this model, usage.completion_tokens_details.reasoning_tokens, n=3/level
+#     low 84/116/78   medium 108/108/104   high 109/139/104
+#     xhigh 131/87/132   max 140/134/128
+#   Adjacent levels overlap: medium's whole range (104-108) sits inside
+#   high's (104-139), and xhigh's (87-132) is wider than the gap to either
+#   neighbour. Nothing here separates medium from high, or xhigh from high.
+#
+#   control, sibling deepseek-v4-flash-0731 (same family, same host),
+#   reasoning_content length, n=5/level, means
+#     low 826 / medium 442 / high 453 / xhigh 648 / max 1801
+#   which is not monotonic at all (max ranged 414-6790).
+#
+# An effort level a caller cannot observe is worse than not offering it, so
+# this file authors the lab baseline set, matching the other Y-API DeepSeek
+# Flash entries.
+base_model = "deepseek/deepseek-v4.1-flash"
+reasoning_options = [
+  { type = "toggle" },
+  { type = "effort", values = ["low", "high", "max"] },
+]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.01
+output = 0.05

providers/y-api/models/moonshotai/kimi-k3.toml

@@ -0,0 +1,15 @@
+# Toggle: thinking.type = enabled|disabled
+# Effort: reasoning_effort = low|high|max
+# Reasoning text arrives as reasoning_details, not reasoning_content (live API, 2026-09-16).
+base_model = "moonshotai/kimi-k3"
+reasoning_options = [
+  { type = "toggle" },
+  { type = "effort", values = ["low", "high", "max"] },
+]
+
+[interleaved]
+field = "reasoning_details"
+
+[cost]
+input = 0.15
+output = 0.75

providers/y-api/models/openai/gpt-5.6-luna.toml

@@ -0,0 +1,8 @@
+# Effort: reasoning_effort = none|low|medium|high|xhigh
+# The live API rejects `max` for this model; no reasoning text is exposed.
+base_model = "openai/gpt-5.6-luna"
+reasoning_options = [{ type = "effort", values = ["none", "low", "medium", "high", "xhigh"] }]
+
+[cost]
+input = 0.015
+output = 0.065

providers/y-api/models/openai/gpt-5.6-sol.toml

@@ -0,0 +1,8 @@
+# Effort: reasoning_effort = none|low|medium|high|xhigh
+# The live API rejects `max` for this model; no reasoning text is exposed.
+base_model = "openai/gpt-5.6-sol"
+reasoning_options = [{ type = "effort", values = ["none", "low", "medium", "high", "xhigh"] }]
+
+[cost]
+input = 0.25
+output = 1.5

providers/y-api/models/openai/gpt-5.6-terra.toml

@@ -0,0 +1,8 @@
+# Effort: reasoning_effort = none|low|medium|high|xhigh
+# The live API rejects `max` for this model; no reasoning text is exposed.
+base_model = "openai/gpt-5.6-terra"
+reasoning_options = [{ type = "effort", values = ["none", "low", "medium", "high", "xhigh"] }]
+
+[cost]
+input = 0.1
+output = 0.6

providers/y-api/models/openai/gpt-6-astra.toml

@@ -0,0 +1,9 @@
+# Effort: reasoning_effort = low|medium|high|xhigh — no `none` and no `max`, per the
+# live API error text; no reasoning text is exposed.
+# This model also rejects `max_tokens` and requires `max_completion_tokens`.
+base_model = "openai/gpt-6-astra"
+reasoning_options = [{ type = "effort", values = ["low", "medium", "high", "xhigh"] }]
+
+[cost]
+input = 0.5
+output = 2.5

providers/y-api/models/qwen/qwen3.8-flash.toml

@@ -0,0 +1,26 @@
+# Effort: reasoning_effort accepts low|medium|xhigh. An unlisted value returns the
+# supported set verbatim, so this is read off the wire rather than from Alibaba's docs:
+#   "Unexpected reasoning effort minimal. Supported types are xhigh (default),
+#    medium, and low."
+# `none` also answered 200 on this host (5/5), but it is not in that enumeration, so
+# it is not authored here.
+#
+# No [interleaved]: this host returns the trace as a plain `reasoning` string
+# (`message` keys are content/reasoning/role/tool_calls), which is not a
+# schema-valid interleaved.field. Same reason the four OpenAI models omit it.
+#
+# [limit] is overridden. The Alibaba Cloud entry advertises a 1M context, but this
+# host caps this model well below that:
+#   "The input (2000052 tokens) is longer than the model's context length (262144 tokens)."
+# 262144 = 256K, so inheriting the base limit would overstate the window by ~4x.
+#
+# Cost: USD cash price (credit / 20), per the header in provider.toml.
+base_model = "alibaba/qwen3.8-flash"
+reasoning_options = [{ type = "effort", values = ["low", "medium", "xhigh"] }]
+
+[limit]
+context = 262_144
+
+[cost]
+input = 0.01
+output = 0.025

providers/y-api/models/tencent/hy3.toml

@@ -0,0 +1,11 @@
+# Effort: reasoning_effort = none|low|high
+# Cost: free tier — no credit is deducted (https://y-api.bestvirtualgoods.com/pricing).
+base_model = "tencent/hy3"
+reasoning_options = [{ type = "effort", values = ["none", "low", "high"] }]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0
+output = 0

providers/y-api/models/xiaomi/mimo-v2.5.toml

@@ -0,0 +1,11 @@
+# Toggle: thinking.type = enabled|disabled
+# Cost: free tier — no credit is deducted (https://y-api.bestvirtualgoods.com/pricing).
+base_model = "xiaomi/mimo-v2.5"
+reasoning_options = [{ type = "toggle" }]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0
+output = 0

providers/y-api/models/z-ai/glm-5.2.toml

@@ -0,0 +1,15 @@
+# Effort: reasoning_effort = high|max; thinking is always on.
+#
+# Interleaved field: reasoning_details, not the lab's reasoning_content. This
+# host returned reasoning_content on 0 of 10 calls (2026-09-16); it returned
+# reasoning_details ({type: "reasoning.text"}) on 8 of 10, the other 2 carrying
+# no side channel at all. The field that exists is the one authored.
+base_model = "zhipuai/glm-5.2"
+reasoning_options = [{ type = "effort", values = ["high", "max"] }]
+
+[interleaved]
+field = "reasoning_details"
+
+[cost]
+input = 0.07
+output = 0.22

providers/y-api/models/z-ai/glm-5.3-flash.toml

@@ -0,0 +1,10 @@
+# Effort: reasoning_effort = low|high|max; thinking is always on.
+base_model = "zhipuai/glm-5.3-flash"
+reasoning_options = [{ type = "effort", values = ["low", "high", "max"] }]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.0075
+output = 0.025

providers/y-api/models/z-ai/glm-5.3.toml

@@ -0,0 +1,10 @@
+# Effort: reasoning_effort = low|high|max; thinking is always on.
+base_model = "zhipuai/glm-5.3"
+reasoning_options = [{ type = "effort", values = ["low", "high", "max"] }]
+
+[interleaved]
+field = "reasoning_content"
+
+[cost]
+input = 0.07
+output = 0.25

providers/y-api/provider.toml

@@ -0,0 +1,20 @@
+# Cost basis: USD cash price. The published catalog quotes account credit, and
+# credit is topped up at a promotional 20x rate, so every cost value here is
+# the credit price / 20.
+#
+# The 20x rate is stated by the service itself, in two places, both read
+# 2026-09-16:
+#   https://y-api.bestvirtualgoods.com/about
+#     "Top up $1 to receive $20 of account credit (limited-time rate; reverts
+#      to 1:10 afterwards)."
+#   https://y-api.bestvirtualgoods.com/pricing
+#     "limited-time top up $1 for $20 in credit (reverts to 1:10 afterwards)"
+#
+# Both pages also say the rate is limited-time and reverts to 1:10. When it
+# does, every cost below doubles. The service is not reachable from mainland
+# China and quotes no non-USD price, so no FX conversion is involved.
+name = "Y-API"
+env = ["Y_API_API_KEY"]
+npm = "@ai-sdk/openai-compatible"
+api = "https://api.y-api.bestvirtualgoods.com/v1"
+doc = "https://y-api.bestvirtualgoods.com/docs"