Skip to content

A platform for community discussion. Free, open, simple.

FEATURE: Support OpenAI GPT-5.6 models (#41832)

Same1caf1d1d3d1
diff --git a/plugins/discourse-ai/app/models/llm_model.rb b/plugins/discourse-ai/app/models/llm_model.rb
index d9c4d56fd1e..a916e34078c 100644
--- a/plugins/discourse-ai/app/models/llm_model.rb
+++ b/plugins/discourse-ai/app/models/llm_model.rb
@@ -9,6 +9,8 @@ class LlmModel < ActiveRecord::Base
BEDROCK_PROVIDER_NAME = "aws_bedrock"
BEDROCK_CONVERSE_PROVIDER_NAME = "aws_bedrock_converse"
GOOGLE_VERTEX_AI_PROVIDER_NAME = "google_vertex_ai"
+ OPEN_AI_PROVIDER_NAME = "open_ai"
+ OPEN_AI_REASONING_MODES = %w[standard pro].freeze
# Interpolated into the Vertex AI hostname/path — must stay strict to avoid
# sending environment credentials to an attacker-controlled host.
GOOGLE_VERTEX_AI_REGION_FORMAT = /\A[a-z](?:[a-z0-9-]*[a-z0-9])?\z/
@@ -241,16 +243,22 @@ class LlmModel < ActiveRecord::Base
disable_native_tools: :checkbox,
reasoning_effort: {
type: :enum,
- values: %w[default none minimal low medium high xhigh],
+ values: %w[default none minimal low medium high xhigh max],
+ default: "default",
+ },
+ reasoning_mode: {
+ type: :enum,
+ values: ["default", *OPEN_AI_REASONING_MODES],
default: "default",
+ tooltip: "discourse_ai.llms.provider_field_hints.reasoning_mode",
},
disable_temperature: {
type: :checkbox,
- hidden_if: :reasoning_effort,
+ hidden_if: %i[reasoning_effort reasoning_mode],
},
disable_top_p: {
type: :checkbox,
- hidden_if: :reasoning_effort,
+ hidden_if: %i[reasoning_effort reasoning_mode],
},
disable_streaming: :checkbox,
service_tier: {
@@ -330,8 +338,9 @@ class LlmModel < ActiveRecord::Base
disable_native_tools: :checkbox,
reasoning_effort: {
type: :enum,
- values: %w[default none minimal low medium high xhigh],
+ values: %w[default none minimal low medium high xhigh max],
default: "default",
+ tooltip: "discourse_ai.llms.provider_field_hints.azure_reasoning_effort",
},
disable_temperature: {
type: :checkbox,
@@ -615,7 +624,12 @@ class LlmModel < ActiveRecord::Base
end
def required_provider_params
- if provider == BEDROCK_PROVIDER_NAME
+ reasoning_mode = lookup_custom_param("reasoning_mode")
+ if provider == OPEN_AI_PROVIDER_NAME && reasoning_mode == "pro"
+ if !url.to_s.include?("/v1/responses")
+ errors.add(:base, I18n.t("discourse_ai.llm_models.reasoning_mode_requirements"))
+ end
+ elsif provider == BEDROCK_PROVIDER_NAME
if lookup_custom_param("region").blank?
errors.add(:base, I18n.t("discourse_ai.llm_models.missing_provider_param", param: "region"))
end
diff --git a/plugins/discourse-ai/config/locales/client.en.yml b/plugins/discourse-ai/config/locales/client.en.yml
index 751f5947071..7128b9272ab 100644
--- a/plugins/discourse-ai/config/locales/client.en.yml
+++ b/plugins/discourse-ai/config/locales/client.en.yml
@@ -988,9 +988,9 @@ en:
anthropic-claude-haiku-4-5: "Fast and cost-effective"
google-gemini-3-1-pro: "Google's most intelligent model"
google-gemini-3-flash: "Fast and cost-efficient with multimodal reasoning"
- open_ai-gpt-5-5: "OpenAI's most capable model with extended context"
- open_ai-gpt-5-4: "OpenAI's flagship model, well suited for problem solving across domains"
- open_ai-gpt-5-4-nano: "Fastest, most cost-effective GPT-5 model"
+ open_ai-gpt-5-6-sol: "OpenAI's frontier model for complex professional work"
+ open_ai-gpt-5-6-terra: "Balances intelligence and cost for demanding workloads"
+ open_ai-gpt-5-6-luna: "Optimized for cost-sensitive, high-volume workloads"
open_router-moonshotai-kimi-k2-6: "Moonshot's multimodal model with vision capabilities"
open_router-deepseek-deepseek-v4-flash: "DeepSeek's latest high-performance model with 1M context"
open_router-minimax-minimax-m2-7: "MiniMax's model with large context"
@@ -1067,6 +1067,7 @@ en:
provider_quantizations: "Order of provider quantizations (comma delimited list eg: fp16,fp8)"
disable_streaming: "Disable streaming completions (convert streaming to non streaming requests)"
reasoning_effort: "Reasoning effort (only applicable to reasoning models)"
+ reasoning_mode: "Reasoning mode"
service_tier: "Service tier (flex = lower cost/higher latency, priority = faster processing)"
reasoning_parser: "vLLM reasoning parser"
thinking_override: "Thinking override"
@@ -1086,6 +1087,8 @@ en:
effort: "Effort level"
provider_field_hints:
+ reasoning_mode: "Standard mode uses the provider default. Pro mode requires the Responses API and support from the selected model, and increases latency and token usage."
+ azure_reasoning_effort: "Available effort levels depend on the model deployed in Azure. Unsupported values are rejected by Azure."
reasoning_parser: "Must match the vLLM server --reasoning-parser setting. Discourse cannot switch parsers per request; it uses this to shape request options."
thinking_override: "Optionally force supported vLLM chat templates to enable or disable thinking for each request. Unsupported parsers keep the server default."
reasoning_effort: "Sends vLLM's reasoning_effort request option for parsers that support it."
diff --git a/plugins/discourse-ai/config/locales/server.en.yml b/plugins/discourse-ai/config/locales/server.en.yml
index 57467712443..183e4a79310 100644
--- a/plugins/discourse-ai/config/locales/server.en.yml
+++ b/plugins/discourse-ai/config/locales/server.en.yml
@@ -1058,6 +1058,7 @@ en:
llm_models:
missing_provider_param: "%{param} can't be blank"
invalid_provider_param: "%{param} has an invalid format"
+ reasoning_mode_requirements: "Pro reasoning mode requires the Responses API"
secret_required: "A credential must be selected"
secret_not_found: "The selected credential does not exist"
bedrock_invalid_url: "Please complete all the fields to use this model."
diff --git a/plugins/discourse-ai/lib/completions/endpoints/open_ai_responses.rb b/plugins/discourse-ai/lib/completions/endpoints/open_ai_responses.rb
index b8b86840b4e..7dfaf1f6539 100644
--- a/plugins/discourse-ai/lib/completions/endpoints/open_ai_responses.rb
+++ b/plugins/discourse-ai/lib/completions/endpoints/open_ai_responses.rb
@@ -22,6 +22,7 @@ module DiscourseAi
reasoning_payload = { summary: "auto" }
reasoning_payload[:effort] = reasoning_effort if reasoning_effort
+ reasoning_payload[:mode] = reasoning_mode if reasoning_mode
payload.merge!(reasoning: reasoning_payload)
payload[:service_tier] = service_tier if service_tier
diff --git a/plugins/discourse-ai/lib/completions/endpoints/open_ai_shared.rb b/plugins/discourse-ai/lib/completions/endpoints/open_ai_shared.rb
index 4aa24f5e046..a7e735179ab 100644
--- a/plugins/discourse-ai/lib/completions/endpoints/open_ai_shared.rb
+++ b/plugins/discourse-ai/lib/completions/endpoints/open_ai_shared.rb
@@ -4,12 +4,27 @@ module DiscourseAi
module Completions
module Endpoints
module OpenAiShared
- OPEN_AI_REASONING_EFFORTS = %w[none minimal low medium high xhigh].freeze
+ OPEN_AI_REASONING_EFFORTS = %w[none minimal low medium high xhigh max].freeze
+ # Unknown and newer models receive max unchanged; only documented xhigh-ceiling
+ # models are downgraded.
+ OPEN_AI_MAX_TO_XHIGH_MODELS = %w[
+ gpt-5.1-codex-max
+ gpt-5.2
+ gpt-5.2-pro
+ gpt-5.2-codex
+ gpt-5.3-codex
+ gpt-5.4
+ gpt-5.4-mini
+ gpt-5.4-nano
+ gpt-5.4-pro
+ gpt-5.5
+ gpt-5.5-pro
+ ].freeze
+ OPEN_AI_MODEL_SNAPSHOT_SUFFIX = /-\d{4}-\d{2}-\d{2}\z/
OPEN_AI_REASONING_OUTPUT_RESERVATION = 25_000
- # "minimal" isn't a distinct tier on some OpenAI reasoning models (e.g.
- # gpt-5.5 rejects it outright) — collapse it into "low", same as vLLM's
- # own reasoning effort scale already does.
- OPEN_AI_EFFORT_ALIASES = { "max" => "xhigh", "minimal" => "low" }.freeze
+ # "minimal" isn't a distinct tier on every OpenAI reasoning model and can
+ # be rejected outright, so collapse it into "low".
+ OPEN_AI_EFFORT_ALIASES = { "minimal" => "low" }.freeze
def default_options
{ model: llm_model.name }
@@ -67,9 +82,13 @@ module DiscourseAi
)
effort ||= legacy_reasoning_effort
- return DiscourseAi::Completions::ThinkingConfig.disabled if effort.blank?
+ if effort.blank?
+ return default_reasoning_mode_thinking_config(model_params) if reasoning_mode
+ return DiscourseAi::Completions::ThinkingConfig.disabled
+ end
provider_effort = OPEN_AI_EFFORT_ALIASES[effort] || effort
+ provider_effort = "xhigh" if provider_effort == "max" && max_effort_maps_to_xhigh?
if !OPEN_AI_REASONING_EFFORTS.include?(provider_effort)
return DiscourseAi::Completions::ThinkingConfig.unsupported(canonical_effort: effort)
end
@@ -103,6 +122,36 @@ module DiscourseAi
thinking_config&.provider_effort
end
+ def reasoning_mode
+ return @reasoning_mode if defined?(@reasoning_mode)
+
+ @reasoning_mode = llm_model.lookup_custom_param("reasoning_mode")
+ if !LlmModel::OPEN_AI_REASONING_MODES.include?(@reasoning_mode) ||
+ !supports_reasoning_mode?
+ @reasoning_mode = nil
+ end
+ @reasoning_mode
+ end
+
+ def default_reasoning_mode_thinking_config(model_params)
+ DiscourseAi::Completions::ThinkingConfig.new(
+ enabled: true,
+ reserved_output_tokens: open_ai_reasoning_output_reservation(model_params, "medium"),
+ strip_temperature: true,
+ strip_top_p: true,
+ )
+ end
+
+ def supports_reasoning_mode?
+ llm_model.provider == LlmModel::OPEN_AI_PROVIDER_NAME &&
+ llm_model.url.to_s.include?("/v1/responses")
+ end
+
+ def max_effort_maps_to_xhigh?
+ model_name = llm_model.name.to_s.sub(OPEN_AI_MODEL_SNAPSHOT_SUFFIX, "")
+ OPEN_AI_MAX_TO_XHIGH_MODELS.include?(model_name)
+ end
+
def legacy_reasoning_effort
return @legacy_reasoning_effort if defined?(@legacy_reasoning_effort)
@legacy_reasoning_effort = llm_model.lookup_custom_param("reasoning_effort")
@@ -153,6 +202,7 @@ module DiscourseAi
log.request_tokens = processor.prompt_tokens if processor.prompt_tokens
log.response_tokens = processor.completion_tokens if processor.completion_tokens
log.cache_read_tokens = processor.cache_read_tokens if processor.cache_read_tokens
+ log.cache_write_tokens = processor.cache_write_tokens if processor.cache_write_tokens
end
def decode(response_raw)
diff --git a/plugins/discourse-ai/lib/completions/llm_presets.rb b/plugins/discourse-ai/lib/completions/llm_presets.rb
index 6bef911b567..15806bfd483 100644
--- a/plugins/discourse-ai/lib/completions/llm_presets.rb
+++ b/plugins/discourse-ai/lib/completions/llm_presets.rb
@@ -177,35 +177,38 @@ module DiscourseAi
id: "open_ai",
models: [
model(
- name: "gpt-5.5",
+ name: "gpt-5.6-sol",
tokens: 1_050_000,
- display_name: "GPT-5.5",
+ display_name: "GPT-5.6 Sol",
max_output_tokens: 128_000,
input_cost: 5.0,
cached_input_cost: 0.50,
+ cache_write_cost: 6.25,
output_cost: 30.0,
vision_enabled: true,
endpoint: "https://api.openai.com/v1/responses",
),
model(
- name: "gpt-5.4",
- tokens: 400_000,
- display_name: "GPT-5.4",
+ name: "gpt-5.6-terra",
+ tokens: 1_050_000,
+ display_name: "GPT-5.6 Terra",
max_output_tokens: 128_000,
input_cost: 2.50,
cached_input_cost: 0.25,
+ cache_write_cost: 3.125,
output_cost: 15.0,
vision_enabled: true,
endpoint: "https://api.openai.com/v1/responses",
),
model(
- name: "gpt-5.4-nano",
- tokens: 400_000,
- display_name: "GPT-5.4 Nano",
+ name: "gpt-5.6-luna",
+ tokens: 1_050_000,
+ display_name: "GPT-5.6 Luna",
max_output_tokens: 128_000,
- input_cost: 0.20,
- cached_input_cost: 0.02,
- output_cost: 1.25,
+ input_cost: 1.0,
+ cached_input_cost: 0.10,
+ cache_write_cost: 1.25,
+ output_cost: 6.0,
vision_enabled: true,
endpoint: "https://api.openai.com/v1/responses",
),
diff --git a/plugins/discourse-ai/lib/completions/open_ai_message_processor.rb b/plugins/discourse-ai/lib/completions/open_ai_message_processor.rb
index 59fa91f6b25..7dc6083d8a9 100644
--- a/plugins/discourse-ai/lib/completions/open_ai_message_processor.rb
+++ b/plugins/discourse-ai/lib/completions/open_ai_message_processor.rb
@@ -1,7 +1,7 @@
# frozen_string_literal: true
module DiscourseAi::Completions
class OpenAiMessageProcessor
- attr_reader :prompt_tokens, :completion_tokens, :cache_read_tokens
+ attr_reader :prompt_tokens, :completion_tokens, :cache_read_tokens, :cache_write_tokens
def initialize(partial_tool_calls: false)
@tool = nil
@@ -9,6 +9,7 @@ module DiscourseAi::Completions
@prompt_tokens = nil
@completion_tokens = nil
@cache_read_tokens = nil
+ @cache_write_tokens = nil
@partial_tool_calls = partial_tool_calls
end
@@ -125,14 +126,17 @@ module DiscourseAi::Completions
usage = json.dig(:usage)
return if !usage
- cached_tokens = usage.dig(:prompt_tokens_details, :cached_tokens).to_i
+ token_details = usage[:prompt_tokens_details] || {}
+ cached_tokens = token_details[:cached_tokens].to_i
+ cache_write_tokens = token_details[:cache_write_tokens].to_i
prompt_tokens = usage[:prompt_tokens].to_i
completion_tokens = usage[:completion_tokens].to_i
- @prompt_tokens = prompt_tokens - cached_tokens
+ @prompt_tokens = prompt_tokens - cached_tokens - cache_write_tokens
@completion_tokens = completion_tokens if completion_tokens.positive?
@cache_read_tokens = cached_tokens
+ @cache_write_tokens = cache_write_tokens
end
end
end
diff --git a/plugins/discourse-ai/lib/completions/open_ai_responses_message_processor.rb b/plugins/discourse-ai/lib/completions/open_ai_responses_message_processor.rb
index 419a046380f..6c05692fac8 100644
--- a/plugins/discourse-ai/lib/completions/open_ai_responses_message_processor.rb
+++ b/plugins/discourse-ai/lib/completions/open_ai_responses_message_processor.rb
@@ -3,7 +3,7 @@ module DiscourseAi::Completions
class OpenAiResponsesMessageProcessor
PROVIDER_KEY = :open_ai_responses
- attr_reader :prompt_tokens, :completion_tokens, :cache_read_tokens
+ attr_reader :prompt_tokens, :completion_tokens, :cache_read_tokens, :cache_write_tokens
def initialize(partial_tool_calls: false, output_thinking: false)
@tool = nil # currently streaming ToolCall
@@ -11,6 +11,7 @@ module DiscourseAi::Completions
@prompt_tokens = nil
@completion_tokens = nil
@cache_read_tokens = nil
+ @cache_write_tokens = nil
@partial_tool_calls = partial_tool_calls
@streaming_parser = nil # JsonStreamingTracker, if used
@has_new_data = false
@@ -281,14 +282,17 @@ module DiscourseAi::Completions
end
def update_usage(json)
- usage = json.dig(:response, :usage)
+ usage = json.dig(:response, :usage) || json[:usage]
return if !usage
- cached_tokens = usage.dig(:input_tokens_details, :cached_tokens).to_i
+ token_details = usage[:input_tokens_details] || {}
+ cached_tokens = token_details[:cached_tokens].to_i
+ cache_write_tokens = token_details[:cache_write_tokens].to_i
- @prompt_tokens ||= usage[:input_tokens] - cached_tokens
+ @prompt_tokens ||= usage[:input_tokens] - cached_tokens - cache_write_tokens
@completion_tokens ||= usage[:output_tokens]
@cache_read_tokens ||= cached_tokens
+ @cache_write_tokens ||= cache_write_tokens
end
def build_partial_reasoning_delta(json, prefix: nil)
diff --git a/plugins/discourse-ai/spec/lib/completions/endpoints/open_ai_responses_api_spec.rb b/plugins/discourse-ai/spec/lib/completions/endpoints/open_ai_responses_api_spec.rb
index 705c353f96a..c250dbc3a56 100644
--- a/plugins/discourse-ai/spec/lib/completions/endpoints/open_ai_responses_api_spec.rb
+++ b/plugins/discourse-ai/spec/lib/completions/endpoints/open_ai_responses_api_spec.rb
@@ -343,6 +343,40 @@ data: {"type":"response.reasoning_summary_part.added","sequence_number":3,"item_
)
end
+ it "records cache-write tokens and configured cost for non-streaming responses" do
+ model.update!(
+ name: "gpt-5.6-luna",
+ input_cost: 1.0,
+ cached_input_cost: 0.1,
+ cache_write_cost: 1.25,
+ output_cost: 6.0,
+ )
+ stub_request(:post, "https://api.openai.com/v1/responses").to_return(
+ status: 200,
+ body: {
+ output: [{ type: "message", content: [{ type: "output_text", text: "ok" }] }],
+ usage: {
+ input_tokens: 2_006,
+ output_tokens: 300,
+ input_tokens_details: {
+ cached_tokens: 1_920,
+ cache_write_tokens: 64,
+ },
+ },
+ }.to_json,
+ )
+
+ result = model.to_llm.generate("hello", user: Discourse.system_user)
+
+ expect(result).to eq("ok")
+ log = AiApiAuditLog.last
+ expect(log.request_tokens).to eq(22)
+ expect(log.response_tokens).to eq(300)
+ expect(log.cache_read_tokens).to eq(1_920)
+ expect(log.cache_write_tokens).to eq(64)
+ expect(log.estimated_cost).to eq(BigDecimal("0.002094"))
+ end
+
it "accepts xhigh as reasoning_effort for responses API" do
model.update!(provider_params: { reasoning_effort: "xhigh" })
@@ -368,6 +402,118 @@ data: {"type":"response.reasoning_summary_part.added","sequence_number":3,"item_
expect(parsed_body[:reasoning]).to include(effort: "xhigh", summary: "auto")
end
+ it "sends native max effort and pro mode for GPT-5.6" do
+ model.update!(
+ name: "gpt-5.6-sol",
+ provider_params: {
+ reasoning_effort: "max",
+ reasoning_mode: "pro",
+ },
+ )
+
+ parsed_body = nil
+ stub_request(:post, "https://api.openai.com/v1/responses").with(
+ body:
+ proc do |req_body|
+ parsed_body = JSON.parse(req_body, symbolize_names: true)
+ true
+ end,
+ ).to_return(status: 200, body: { output: [] }.to_json)
+
+ prompt =
+ DiscourseAi::Completions::Prompt.new(
+ "You are a bot",
+ messages: [type: :user, content: "hello"],
+ )
+ dialect = DiscourseAi::Completions::Dialects::OpenAiResponses.new(prompt, model)
+
+ endpoint.perform_completion!(dialect, Discourse.system_user, { temperature: 0.7, top_p: 0.9 })
+
+ expect(parsed_body[:reasoning]).to eq(summary: "auto", effort: "max", mode: "pro")
+ expect(parsed_body).not_to have_key(:temperature)
+ expect(parsed_body).not_to have_key(:top_p)
+ end
+
+ it "uses the provider's default effort for pro mode" do
+ model.update!(name: "gpt-5.6-sol", provider_params: { reasoning_mode: "pro" })
+
+ parsed_body = nil
+ stub_request(:post, "https://api.openai.com/v1/responses").with(
+ body:
+ proc do |req_body|
+ parsed_body = JSON.parse(req_body, symbolize_names: true)
+ true
+ end,
+ ).to_return(status: 200, body: { output: [] }.to_json)
+
+ prompt =
+ DiscourseAi::Completions::Prompt.new(
+ "You are a bot",
+ messages: [type: :user, content: "hello"],
+ )
+ dialect = DiscourseAi::Completions::Dialects::OpenAiResponses.new(prompt, model)
+
+ expect(endpoint.prepare_model_params({})[:reserved_output_tokens]).to eq(25_000)
+ endpoint.perform_completion!(dialect, Discourse.system_user, { temperature: 0.7, top_p: 0.9 })
+
+ expect(parsed_body[:reasoning]).to eq(summary: "auto", mode: "pro")
+ expect(parsed_body).not_to have_key(:temperature)
+ expect(parsed_body).not_to have_key(:top_p)
+ end
+
+ it "sends standard reasoning mode for custom model names" do
+ model.update!(
+ name: "future-reasoning-deployment",
+ provider_params: {
+ reasoning_mode: "standard",
+ },
+ )
+
+ parsed_body = nil
+ stub_request(:post, "https://api.openai.com/v1/responses").with(
+ body:
+ proc do |req_body|
+ parsed_body = JSON.parse(req_body, symbolize_names: true)
+ true
+ end,
+ ).to_return(status: 200, body: { output: [] }.to_json)
+
+ prompt =
+ DiscourseAi::Completions::Prompt.new(
+ "You are a bot",
+ messages: [type: :user, content: "hello"],
+ )
+ dialect = DiscourseAi::Completions::Dialects::OpenAiResponses.new(prompt, model)
+
+ endpoint.perform_completion!(dialect, Discourse.system_user)
+
+ expect(parsed_body[:reasoning]).to eq(summary: "auto", mode: "standard")
+ end
+
+ it "omits unsupported reasoning modes" do
+ model.update!(provider_params: { reasoning_mode: "unsupported" })
+
+ parsed_body = nil
+ stub_request(:post, "https://api.openai.com/v1/responses").with(
+ body:
+ proc do |req_body|
+ parsed_body = JSON.parse(req_body, symbolize_names: true)
+ true
+ end,
+ ).to_return(status: 200, body: { output: [] }.to_json)
+
+ prompt =
+ DiscourseAi::Completions::Prompt.new(
+ "You are a bot",
+ messages: [type: :user, content: "hello"],
+ )
+ dialect = DiscourseAi::Completions::Dialects::OpenAiResponses.new(prompt, model)
+
+ endpoint.perform_completion!(dialect, Discourse.system_user)
+
+ expect(parsed_body[:reasoning]).to eq(summary: "auto")
+ end
+
it "accepts none as reasoning_effort for responses API" do
model.update!(provider_params: { reasoning_effort: "none" })
diff --git a/plugins/discourse-ai/spec/lib/completions/endpoints/open_ai_spec.rb b/plugins/discourse-ai/spec/lib/completions/endpoints/open_ai_spec.rb
index cc545b4f7a7..ff1b924fe4f 100644
--- a/plugins/discourse-ai/spec/lib/completions/endpoints/open_ai_spec.rb
+++ b/plugins/discourse-ai/spec/lib/completions/endpoints/open_ai_spec.rb
@@ -1197,8 +1197,8 @@ TEXT
expect(parsed_body).not_to have_key(:top_p)
end
- it "uses per-call thinking_effort before model reasoning_effort" do
- model.update!(provider_params: { reasoning_effort: "low" })
+ it "uses per-call thinking effort and maps max to xhigh for allowlisted models" do
+ model.update!(name: "gpt-5.5-2026-04-24", provider_params: { reasoning_effort: "low" })
parsed_body = nil
stub_request(:post, "https://api.openai.com/v1/chat/completions").with(
@@ -1215,6 +1215,65 @@ TEXT
expect(parsed_body).not_to have_key(:temperature)
end
+ it "uses native max effort for GPT-5.6" do
+ model.update!(name: "gpt-5.6-sol")
+
+ parsed_body = nil
+ stub_request(:post, "https://api.openai.com/v1/chat/completions").with(
+ body:
+ proc do |req_body|
+ parsed_body = JSON.parse(req_body, symbolize_names: true)
+ true
+ end,
+ ).to_return(status: 200, body: { choices: [{ message: { content: "test" } }] }.to_json)
+
+ endpoint.perform_completion!(dialect, user, { thinking_effort: "max" })
+
+ expect(parsed_body[:reasoning_effort]).to eq("max")
+ end
+
+ it "uses native max effort for future OpenAI models by default" do
+ model.update!(name: "gpt-5.7")
+
+ parsed_body = nil
+ stub_request(:post, "https://api.openai.com/v1/chat/completions").with(
+ body:
+ proc do |req_body|
+ parsed_body = JSON.parse(req_body, symbolize_names: true)
+ true
+ end,
+ ).to_return(status: 200, body: { choices: [{ message: { content: "test" } }] }.to_json)
+
+ endpoint.perform_completion!(dialect, user, { thinking_effort: "max" })
+
+ expect(parsed_body[:reasoning_effort]).to eq("max")
+ end
+
+ it "passes max through for Azure deployments with opaque names" do
+ azure_url = "https://test.openai.azure.com/openai/v1/chat/completions"
+ model.update!(
+ provider: "azure",
+ name: "production-deployment",
+ url: azure_url,
+ provider_params: {
+ reasoning_effort: "max",
+ },
+ )
+
+ parsed_body = nil
+ stub_request(:post, azure_url).with(
+ body:
+ proc do |req_body|
+ parsed_body = JSON.parse(req_body, symbolize_names: true)
+ true
+ end,
+ ).to_return(status: 200, body: { choices: [{ message: { content: "test" } }] }.to_json)
+
+ endpoint.perform_completion!(dialect, user)
+
+ expect(parsed_body[:reasoning_effort]).to eq("max")
+ end
+
it "reserves OpenAI recommended output room for reasoning when no token cap is configured" do
prepared_params = endpoint.prepare_model_params(thinking_effort: "high")
@@ -1235,6 +1294,24 @@ TEXT
expect(prepared_params).not_to have_key(:reserved_output_tokens)
end
+ it "treats standard reasoning mode as a no-op for Chat Completions" do
+ model.update!(provider_params: { reasoning_mode: "standard" })
+
+ parsed_body = nil
+ stub_request(:post, "https://api.openai.com/v1/chat/completions").with(
+ body:
+ proc do |req_body|
+ parsed_body = JSON.parse(req_body, symbolize_names: true)
+ true
+ end,
+ ).to_return(status: 200, body: { choices: [{ message: { content: "test" } }] }.to_json)
+
+ endpoint.perform_completion!(dialect, user)
+
+ expect(parsed_body).not_to have_key(:reasoning)
+ expect(parsed_body).not_to have_key(:reasoning_effort)
+ end
+
it "omits reasoning parameters when not configured" do
parsed_body = nil
stub_request(:post, "https://api.openai.com/v1/chat/completions").with(
diff --git a/plugins/discourse-ai/spec/lib/completions/llm_presets_spec.rb b/plugins/discourse-ai/spec/lib/completions/llm_presets_spec.rb
index 08b3b83a53d..c2dd337f926 100644
--- a/plugins/discourse-ai/spec/lib/completions/llm_presets_spec.rb
+++ b/plugins/discourse-ai/spec/lib/completions/llm_presets_spec.rb
@@ -30,6 +30,51 @@ RSpec.describe DiscourseAi::Completions::LlmPresets do
expect(model[:provider_params]).to include(region: "global")
end
+ it "includes the current GPT-5.6 family with pricing and capabilities" do
+ preset = described_class.find_provider("open_ai")
+
+ expect(preset[:models]).to eq(
+ [
+ {
+ name: "gpt-5.6-sol",
+ tokens: 1_050_000,
+ display_name: "GPT-5.6 Sol",
+ max_output_tokens: 128_000,
+ input_cost: 5.0,
+ cached_input_cost: 0.50,
+ cache_write_cost: 6.25,
+ output_cost: 30.0,
+ vision_enabled: true,
+ endpoint: "https://api.openai.com/v1/responses",
+ },
+ {
+ name: "gpt-5.6-terra",
+ tokens: 1_050_000,
+ display_name: "GPT-5.6 Terra",
+ max_output_tokens: 128_000,
+ input_cost: 2.50,
+ cached_input_cost: 0.25,
+ cache_write_cost: 3.125,
+ output_cost: 15.0,
+ vision_enabled: true,
+ endpoint: "https://api.openai.com/v1/responses",
+ },
+ {
+ name: "gpt-5.6-luna",
+ tokens: 1_050_000,
+ display_name: "GPT-5.6 Luna",
+ max_output_tokens: 128_000,
+ input_cost: 1.0,
+ cached_input_cost: 0.10,
+ cache_write_cost: 1.25,
+ output_cost: 6.0,
+ vision_enabled: true,
+ endpoint: "https://api.openai.com/v1/responses",
+ },
+ ],
+ )
+ end
+
it "marks every Anthropic model known to require adaptive-only thinking" do
%w[claude-opus-4-7 claude-opus-4-8 claude-sonnet-5].each do |name|
model = described_class.find_model("anthropic", name)
diff --git a/plugins/discourse-ai/spec/lib/completions/open_ai_message_processor_spec.rb b/plugins/discourse-ai/spec/lib/completions/open_ai_message_processor_spec.rb
index b6995a42fdc..9470827434d 100644
--- a/plugins/discourse-ai/spec/lib/completions/open_ai_message_processor_spec.rb
+++ b/plugins/discourse-ai/spec/lib/completions/open_ai_message_processor_spec.rb
@@ -144,4 +144,25 @@ RSpec.describe DiscourseAi::Completions::OpenAiMessageProcessor do
processor.process_streamed_message(chunk(delta: {}, finish_reason: "tool_calls"))
}.to raise_error(JSON::ParserError)
end
+
+ it "separates uncached, cache-read, and cache-write input tokens" do
+ processor = described_class.new
+
+ processor.process_message(
+ choices: [],
+ usage: {
+ prompt_tokens: 2_006,
+ completion_tokens: 300,
+ prompt_tokens_details: {
+ cached_tokens: 1_920,
+ cache_write_tokens: 64,
+ },
+ },
+ )
+
+ expect(processor.prompt_tokens).to eq(22)
+ expect(processor.completion_tokens).to eq(300)
+ expect(processor.cache_read_tokens).to eq(1_920)
+ expect(processor.cache_write_tokens).to eq(64)
+ end
end
diff --git a/plugins/discourse-ai/spec/lib/completions/open_ai_responses_message_processor_spec.rb b/plugins/discourse-ai/spec/lib/completions/open_ai_responses_message_processor_spec.rb
index cc5ab62f1cc..cec6088671a 100644
--- a/plugins/discourse-ai/spec/lib/completions/open_ai_responses_message_processor_spec.rb
+++ b/plugins/discourse-ai/spec/lib/completions/open_ai_responses_message_processor_spec.rb
@@ -86,4 +86,27 @@ RSpec.describe DiscourseAi::Completions::OpenAiResponsesMessageProcessor do
expect(result.last).to eq("HN summary")
end
end
+
+ describe "usage" do
+ it "separates uncached, cache-read, and cache-write input tokens" do
+ processor = described_class.new
+
+ processor.process_message(
+ output: [],
+ usage: {
+ input_tokens: 2_006,
+ output_tokens: 300,
+ input_tokens_details: {
+ cached_tokens: 1_920,
+ cache_write_tokens: 64,
+ },
+ },
+ )
+
+ expect(processor.prompt_tokens).to eq(22)
+ expect(processor.completion_tokens).to eq(300)
+ expect(processor.cache_read_tokens).to eq(1_920)
+ expect(processor.cache_write_tokens).to eq(64)
+ end
+ end
end
diff --git a/plugins/discourse-ai/spec/requests/admin/ai_llms_controller_spec.rb b/plugins/discourse-ai/spec/requests/admin/ai_llms_controller_spec.rb
index 1f2854ccaf4..b0e0f549eba 100644
--- a/plugins/discourse-ai/spec/requests/admin/ai_llms_controller_spec.rb
+++ b/plugins/discourse-ai/spec/requests/admin/ai_llms_controller_spec.rb
@@ -59,6 +59,31 @@ RSpec.describe DiscourseAi::Admin::AiLlmsController do
expect(capabilities.dig("anthropic", "requires_configured_url")).to eq(true)
end
+ it "includes OpenAI reasoning controls metadata" do
+ get "/admin/plugins/discourse-ai/ai-llms.json"
+ expect(response).to be_successful
+
+ open_ai_params = response.parsed_body.dig("meta", "provider_params", "open_ai")
+ expect(open_ai_params["reasoning_effort"]).to include(
+ "type" => "enum",
+ "values" => %w[default none minimal low medium high xhigh max],
+ "default" => "default",
+ )
+ expect(open_ai_params["reasoning_mode"]).to include(
+ "type" => "enum",
+ "values" => %w[default standard pro],
+ "default" => "default",
+ )
+
+ azure_params = response.parsed_body.dig("meta", "provider_params", "azure")
+ expect(azure_params["reasoning_effort"]).to include(
+ "type" => "enum",
+ "values" => %w[default none minimal low medium high xhigh max],
+ "default" => "default",
+ )
+ expect(azure_params).not_to have_key("reasoning_mode")
+ end
+
it "includes vLLM reasoning controls metadata" do
get "/admin/plugins/discourse-ai/ai-llms.json"
expect(response).to be_successful
@@ -274,6 +299,30 @@ RSpec.describe DiscourseAi::Admin::AiLlmsController do
)
end
+ it "stores reasoning configuration for custom model names" do
+ provider_params = { reasoning_effort: "max", reasoning_mode: "pro" }
+ attrs =
+ valid_attrs.merge(
+ name: "future-reasoning-deployment",
+ url: "https://api.openai.com/v1/responses",
+ provider_params: provider_params,
+ )
+
+ post "/admin/plugins/discourse-ai/ai-llms.json", params: { ai_llm: attrs }
+
+ expect(response.status).to eq(201)
+ expect(LlmModel.last.provider_params).to eq(provider_params.stringify_keys)
+ end
+
+ it "allows standard reasoning mode with Chat Completions" do
+ attrs = valid_attrs.merge(provider_params: { reasoning_mode: "standard" })
+
+ post "/admin/plugins/discourse-ai/ai-llms.json", params: { ai_llm: attrs }
+
+ expect(response.status).to eq(201)
+ expect(LlmModel.last.provider_params["reasoning_mode"]).to eq("standard")
+ end
+
it "does not store nested hash values in provider_params" do
provider_params = { organization: { nested: "injected_value" } }
@@ -302,6 +351,17 @@ RSpec.describe DiscourseAi::Admin::AiLlmsController do
end
context "with invalid attributes" do
+ it "rejects pro reasoning mode without a Responses API configuration" do
+ attrs = valid_attrs.merge(provider_params: { reasoning_mode: "pro" })
+
+ post "/admin/plugins/discourse-ai/ai-llms.json", params: { ai_llm: attrs }
+
+ expect(response.status).to eq(422)
+ expect(response.parsed_body["errors"]).to contain_exactly(
+ I18n.t("discourse_ai.llm_models.reasoning_mode_requirements"),
+ )
+ end
+
it "doesn't create a model" do
post "/admin/plugins/discourse-ai/ai-llms.json",
params: {