AI Proxy Advanced

Release date 2026/09/15

Feature

• Added cache_read_cost and cache_write_cost to model.options so cached and cache-write prompt tokens can be priced per 1M tokens, alongside input_cost / output_cost.

• Added filtering for empty Anthropic thinking blocks in request history.

• Added support for reporting prompt cache token metrics when using native (non-OpenAI) request/response formats.

• Added support for reporting prompt cache token metrics in OpenAI-compatible format.

• Added support for per-cache-TTL cache-write pricing, context-window pricing factors, and service-tier pricing factors via model.options (cache_write_cost_list, context_window_factor, service_tier_factor). Anthropic (cache TTL + service tier), OpenAI, and Gemini service tiers are collected from upstream responses.

Bugfix

• Fixed a bug where a response from Anthropic could fail to decode when the client’s request advertised Brotli support, since Anthropic’s edge may then return a Brotli-encoded body that Kong’s response decompression – which only understands gzip – can’t inflate. Kong now normalizes the outgoing Accept-Encoding request header before proxying to any LLM provider: gzip, identity when the client accepts gzip, and identity otherwise, so a client that cannot inflate gzip is never handed a compressed body on the paths that pass the upstream response through untouched.

• Fixed an issue where Gemini and Anthropic token usage metadata was lost in streaming responses.

• Fixed an issue where modern OpenAI output token limits were discarded or overridden during request normalization.

• Fixed an issue where some Claude Code MCP tools did not correctly handle empty ‘requires’ array.

• Fixed an issue where cache_creation_input_tokens and cache_read_input_tokens were excluded from prompt token counts, causing undercounting when Anthropic prompt caching was active.

• Fixed an issue where stream_options was forwarded to Databricks upstreams, causing a 400 error since Databricks does not support this field.

• Fixed an issue where native Anthropic streaming counted provisional message_start output tokens and summed cumulative message_delta usage.

• Fixed an issue where cache_control in Bedrock Anthropic format requests using ARN model IDs was lost.

• Fixed an issue where Bedrock native ConverseStream requests using ARN model IDs could fail while handling streamed responses.

• Fixed an issue where the context-management-2025-06-27 Anthropic beta flag was dropped when routing to a Bedrock target, causing Bedrock to reject requests carrying context_management with a 400 error on failover from an Anthropic target.

• Fixed an issue where output token usage was reported as zero for Bedrock native format request to AWS models like Nova.

• Fixed an issue where signed Anthropic thinking blocks with omitted text were removed from request history.

• Fixed an issue where Anthropic-format requests could report zero usage when it targets Bedrock’s own model via conversion.

• Fixed an issue where full-sync reconfigure reset balancer state even when plugin configs were unchanged.

• Fixed an issue where OpenAI, Azure, and Gemini realtime upstream URLs using wss were downgraded to ws.

• Fixed an issue where a credential header could appear unredacted in access logs when using AWS temporary credentials with Bedrock.

• Fixed an issue where Ollama rejected OpenAI message content arrays.

• Fixed an issue where OpenAI content array was sent as bedrock/gemini system message which only support string.

• Fixed an issue where the configured auth header (e.g. api-key, x-api-key) leaked the provider credential into access logs instead of being redacted.

• Fixed an issue where token usage for Anthropic requests using the compaction beta was undercounted. Kong now sums usage.iterations when present instead of relying solely on the top-level input_tokens / output_tokens, which only reflect non-compaction iterations.

• Fixed an issue where a malformed RSA private key in a Gemini provider’s service_account_json caused crash with SIGSEGV

• Fixed an issue where adapted Anthropic responses omitted cache and reasoning usage details.

• Fixed an issue where Claude Sonnet 5 and Opus 5 rejected OpenAI-format reasoning requests on Bedrock due to an unsupported thinking.type value.

• Fixed an issue where repeated cumulative usage in streaming responses inflated token counts.

• Fixed an issue where Gemini non-streaming responses that reached max_tokens reported a stop finish reason instead of length.

• Fixed an issue where a Gemini response with JSON null token counts in usageMetadata caused an HTTP 500 on non-streaming chat requests and an empty HTTP 200 response on streaming chat requests.

• Fixed an issue where provider-reported reasoning-token details were not correctly reported.

• Fixed an issue where xAI reasoning tokens were not folded into completion_tokens, breaking the OpenAI usage identity.

• Fixed an issue where Gemini’s tool-use token usage (billed when server-side tools such as grounding, code execution, or URL context are used) was not counted in the reported token usage or cost.

• Fixed Azure OpenAI responses support in deployment style.

• Fixed Gemini image generation token accounting to include reasoning tokens (thoughtsTokenCount) in output_tokens and report them as output_tokens_details.reasoning_tokens. Previously the reasoning tokens were dropped, which under-reported output tokens, analytics, and calculated cost for thinking-enabled image models.

• Fixed Gemini realtime (Live API) token accounting to include reasoning tokens (thoughtsTokenCount) in the reported output tokens, and to report them as output_tokens_reasoning_tokens_count. Previously the reasoning tokens were dropped, which under-reported the output and total token counts of thinking-enabled realtime sessions.

• Fixed multipart video generation requests to OpenAI-compatible providers leaking internal multipart metadata into JSON upstream bodies, which caused the requests to fail.

• Fixed token usage statistics being silently dropped for audio/v1/audio/transcriptions requests. Token-billed OpenAI/Azure transcription models (e.g. gpt-4o-transcribe) now report input_tokens/output_tokens/total_tokens and the text_tokens/audio_tokens breakdown in the analytics log when logging.log_statistics is enabled.

• Fixed an issue where Gemini streaming analytics could over-report token usage. Gemini repeats its usage metadata on every stream chunk as a running request-level total, but the counts were summed across chunks instead of replaced, multiplying the reported usage by the number of chunks. Reasoning tokens were affected the most, because Gemini reports them in full on the first chunk and does not change them afterwards. A chunk that carried no usage metadata also reset the recorded output token count to zero.

• Fixed a 500 crash on Bedrock Converse requests when a cachePoint content block immediately followed a toolResult block in the same message.

• Fixed vector database indexes and cached embedding records being left behind when a plugin instance was deleted.

Release date 2026/09/16

Bugfix

• Fixed an issue where a Gemini response with JSON null token counts in usageMetadata caused an HTTP 500 on non-streaming chat requests and an empty HTTP 200 response on streaming chat requests.

• Fixed an issue where provider-reported reasoning-token details were not correctly reported.

• Fixed an issue where xAI reasoning tokens were not folded when visible completion tokens exceeded reasoning tokens.

• Fixed an issue where xAI reasoning tokens were not folded into completion_tokens, breaking the OpenAI usage identity.

• Fixed an issue where Gemini and Anthropic token usage metadata was lost in streaming responses.

• Fixed an issue where modern OpenAI output token limits were discarded or overridden during request normalization.

• Fixed an issue where Bedrock native ConverseStream requests using ARN model IDs could fail while handling streamed responses.

• Fixed an issue where full-sync reconfigure reset balancer state even when plugin configs were unchanged.

• Fixed an issue where an OpenAI content array was sent as a Bedrock/Gemini system message, which only supports strings.

• Fixed an issue where a malformed RSA private key in a Gemini provider’s service_account_json caused a crash with SIGSEGV.

• Fixed an issue where adapted Anthropic responses omitted cache and reasoning usage details.

• Fixed an issue where Claude Sonnet 5 and Opus 5 rejected OpenAI-format reasoning requests on Bedrock due to an unsupported thinking.type value.

• Fixed an issue where repeated cumulative usage in streaming responses inflated token counts.

• Fixed an issue where Gemini non-streaming responses that reached max_tokens reported a stop finish reason instead of length.

• Fixed an issue where cache_control in Bedrock Anthropic format requests using ARN model IDs was lost.

• Fixed multipart video generation requests to OpenAI-compatible providers leaking internal multipart metadata into JSON upstream bodies, which caused the requests to fail.

• Fixed a 500 crash on Bedrock Converse requests when a cachePoint content block immediately followed a toolResult block in the same message.

• Fixed vector database indexes and cached embedding records being left behind when a plugin instance was deleted.

Release date 2026/08/20

Feature

• Added cache_read_cost and cache_write_cost to model.options so cached and cache-write prompt tokens can be priced per 1M tokens, alongside input_cost / output_cost.

• Added filtering for empty Anthropic thinking blocks in request history.

• Added support for reporting prompt cache token metrics in OpenAI-compatible format.

• Added support for reporting prompt cache token metrics when using native (non-OpenAI) request/response formats.

• Usage cost now supports per-cache-TTL cache-write pricing, context-window pricing factors, and service-tier pricing factors via model.options (cache_write_cost_list, context_window_factor, service_tier_factor). Anthropic (cache TTL + service tier), OpenAI, and Gemini service tiers are collected from upstream responses.

Bugfix

• Fixed a bug where a response from Anthropic could fail to decode when the client’s request advertised Brotli support, since Anthropic’s edge may then return a Brotli-encoded body that Kong’s response decompression – which only understands gzip – can’t inflate. Kong now normalizes the outgoing Accept-Encoding request header before proxying to any LLM provider: gzip, identity when the client accepts gzip, and identity otherwise, so a client that cannot inflate gzip is never handed a compressed body on the paths that pass the upstream response through untouched.

• Fixed an issue where token usage for Anthropic requests using the compaction beta was undercounted. Kong now sums usage.iterations when present instead of relying solely on the top-level input_tokens / output_tokens, which only reflect non-compaction iterations.

• Fixed an issue where Gemini’s tool-use token usage (billed when server-side tools such as grounding, code execution, or URL context are used) was not counted in the reported token usage or cost.

• Fixed an issue where signed Anthropic thinking blocks with omitted text were removed from request history.

• Fixed an issue where cache_creation_input_tokens and cache_read_input_tokens were excluded from prompt token counts, causing undercounting when Anthropic prompt caching was active.

• Fixed an issue where native Anthropic streaming counted provisional message_start output tokens and summed cumulative message_delta usage.

• Fixed an issue where the context-management-2025-06-27 Anthropic beta flag was dropped when routing to a Bedrock target, causing Bedrock to reject requests carrying context_management with a 400 error on failover from an Anthropic target.

• Fixed an issue where output token usage was reported as zero for Bedrock native format request to AWS models like Nova.

• Fixed an issue where Anthropic-format requests could report zero usage when it targets Bedrock’s own model via conversion.

• Fixed an issue where OpenAI, Azure, and Gemini realtime upstream URLs using wss were downgraded to ws.

• Fixed an issue where a credential header could appear unredacted in access logs when using AWS temporary credentials with Bedrock.

• Fixed an issue where Ollama rejected OpenAI message content arrays.

• Fixed an issue where some Claude Code MCP tools did not correctly handle empty ‘requires’ array.

• Fixed Gemini image generation token accounting to include reasoning tokens (thoughtsTokenCount) in output_tokens and report them as output_tokens_details.reasoning_tokens. Previously the reasoning tokens were dropped, which under-reported output tokens, analytics, and calculated cost for thinking-enabled image models.

• Fixed Gemini realtime (Live API) token accounting to include reasoning tokens (thoughtsTokenCount) in the reported output tokens, and to report them as output_tokens_reasoning_tokens_count. Previously the reasoning tokens were dropped, which under-reported the output and total token counts of thinking-enabled realtime sessions.

• Fixed token usage statistics being silently dropped for audio/v1/audio/transcriptions requests. Token-billed OpenAI/Azure transcription models (e.g. gpt-4o-transcribe) now report input_tokens/output_tokens/total_tokens and the text_tokens/audio_tokens breakdown in the analytics log when logging.log_statistics is enabled.

• Fixed an issue where Gemini streaming analytics could over-report token usage. Gemini repeats its usage metadata on every stream chunk as a running request-level total, but the counts were summed across chunks instead of replaced, multiplying the reported usage by the number of chunks. Reasoning tokens were affected the most, because Gemini reports them in full on the first chunk and does not change them afterwards. A chunk that carried no usage metadata also reset the recorded output token count to zero.

Release date 2026/07/02

Bugfix

• Fixed an issue where we didn’t support the Azure GA version of Realtime API.

• Fixed an issue where OpenAI-format chat reasoning requests were not mapped consistently across Anthropic, Gemini, Bedrock, and other providers.

• Fixed an issue where system prompts from Claude Code were truncated when doing non-passthrough proxy.

• Fixed an issue where we didn’t support Anthropic format for the Databricks provider.

• Fixed an issue where we didn’t use the Huggingface’s token usage if available.

• Fixed an issue where we didn’t collect the model name in OpenAI streaming response from Azure.

• Fixed an issue where we didn’t correctly ungzip streaming responses.

• Fixed an issue where dashscope replied with a 500 error to inference requests.

• Fixed an issue where GenAI spans were not finished sometimes.

• Fixed an issue where observability errors occurred when gpt-image-1.5 returned image token usage details.

• Fixed an issue where Azure OpenAI authentication failed during load balancer failover due to token refresh attempts in balancer_by_lua. Tokens are now pre-fetched in the access phase.

• Fixed an issue where concurrent worker processes could race to initialize the semantic routing vectorDB index, causing "Index already exists" errors on startup or config reload.

• Fixed an issue where we treated plaintext Anthropic response as gzipped.

• Fixed an issue where output_config was rejected by Vertex AI in Anthropic-native requests.

• Fixed an issue where the configured model name may never be used for the Azure provider.

• Fixed an issue where Claude Code (or CLI) statistics were missing when using an Anthropic model via the Bedrock provider.

Release date 2026/09/17

Bugfix

• Fixed an issue where non-legacy OpenAI output token limits were discarded or overridden during request normalization.

• Fixed an issue where adapted Anthropic responses omitted cache and reasoning usage details.

• Fixed an issue where repeated cumulative usage in streaming responses inflated token counts.

• Fixed an issue where Gemini non-streaming responses that reached max_tokens reported a stop finish reason instead of length.

• Fixed an issue where a Gemini response with JSON null token counts in usageMetadata caused an HTTP 500 on non-streaming chat requests and an empty HTTP 200 response on streaming chat requests.

• Fixed an issue where provider-reported reasoning-token details were not correctly reported.

• Fixed an issue where xAI reasoning tokens were not folded when visible completion tokens exceeded reasoning tokens.

• Fixed an issue where xAI reasoning tokens were not folded into completion_tokens, breaking the OpenAI usage identity.

• Fixed an issue where Gemini and Anthropic token usage metadata was lost in streaming responses.

• Fixed an issue where a malformed RSA private key in a Gemini provider’s service_account_json caused a crash with SIGSEGV.

• Fixed multipart video generation requests to OpenAI-compatible providers leaking internal multipart metadata into JSON upstream bodies, which caused the requests to fail.

• Fixed a 500 crash on Bedrock Converse requests when a cachePoint content block immediately followed a toolResult block in the same message.

• Fixed an issue that failed to delete vector database indexes and cached embedding records when a plugin instance was deleted.

Release date 2026/08/19

Feature

• Added cache_read_cost and cache_write_cost to model.options so cached and cache-write prompt tokens can be priced per 1M tokens, alongside input_cost / output_cost.

• Added filtering for empty Anthropic thinking blocks in request history.

• Added support for reporting prompt cache token metrics in OpenAI-compatible format.

• Added support for reporting prompt cache token metrics when using native (non-OpenAI) request/response formats.

• Usage cost now supports per-cache-TTL cache-write pricing, context-window pricing factors, and service-tier pricing factors via model.options (cache_write_cost_list, context_window_factor, service_tier_factor). Anthropic (cache TTL + service tier), OpenAI, and Gemini service tiers are collected from upstream responses.

Bugfix

• Fixed a bug where a response from Anthropic could fail to decode when the client’s request advertised Brotli support, since Anthropic’s edge may then return a Brotli-encoded body that Kong’s response decompression – which only understands gzip – can’t inflate. Kong now normalizes the outgoing Accept-Encoding request header before proxying to any LLM provider: gzip, identity when the client accepts gzip, and identity otherwise, so a client that cannot inflate gzip is never handed a compressed body on the paths that pass the upstream response through untouched.

• Fixed an issue where native Anthropic streaming counted provisional message_start output tokens and summed cumulative message_delta usage.

• Fixed an issue where output token usage was reported as zero for Bedrock native format request to AWS models like Nova.

• Fixed an issue where Anthropic-format requests could report zero usage when it targets Bedrock’s own model via conversion.

• Fixed an issue where OpenAI, Azure, and Gemini realtime upstream URLs using wss were downgraded to ws.

• Fixed an issue where token usage for Anthropic requests using the compaction beta was undercounted. Kong now sums usage.iterations when present instead of relying solely on the top-level input_tokens / output_tokens, which only reflect non-compaction iterations.

• Fixed an issue where Gemini’s tool-use token usage (billed when server-side tools such as grounding, code execution, or URL context are used) was not counted in the reported token usage or cost.

• Fixed an issue where signed Anthropic thinking blocks with omitted text were removed from request history.

• Fixed an issue where some Claude Code MCP tools did not correctly handle empty ‘requires’ array.

• Fixed Gemini image generation token accounting to include reasoning tokens (thoughtsTokenCount) in output_tokens and report them as output_tokens_details.reasoning_tokens. Previously the reasoning tokens were dropped, which under-reported output tokens, analytics, and calculated cost for thinking-enabled image models.

• Fixed Gemini realtime (Live API) token accounting to include reasoning tokens (thoughtsTokenCount) in the reported output tokens, and to report them as output_tokens_reasoning_tokens_count. Previously the reasoning tokens were dropped, which under-reported the output and total token counts of thinking-enabled realtime sessions.

• Fixed token usage statistics being silently dropped for audio/v1/audio/transcriptions requests. Token-billed OpenAI/Azure transcription models (e.g. gpt-4o-transcribe) now report input_tokens/output_tokens/total_tokens and the text_tokens/audio_tokens breakdown in the analytics log when logging.log_statistics is enabled.

• Fixed an issue where Gemini streaming analytics could over-report token usage. Gemini repeats its usage metadata on every stream chunk as a running request-level total, but the counts were summed across chunks instead of replaced, multiplying the reported usage by the number of chunks. Reasoning tokens were affected the most, because Gemini reports them in full on the first chunk and does not change them afterwards. A chunk that carried no usage metadata also reset the recorded output token count to zero.

Release date 2026/07/29

Bugfix

• Fixed an issue where the OpenAI developer role was not treated as a system prompt when transforming requests to Bedrock and Gemini providers.

• Fixed an issue where Ollama rejected OpenAI message content arrays.

• Fixed an issue where the OpenAI content array was sent as a Bedrock/Gemini system message which only support string.

• Fixed an issue where the configured auth header (e.g. api-key, x-api-key) leaked the provider credential into access logs instead of being redacted.

• Fixed an issue where the context-management-2025-06-27 Anthropic beta flag was dropped when routing to a Bedrock target, causing Bedrock to reject requests carrying context_management with a 400 error on failover from an Anthropic target.

Release date 2026/07/14

Bugfix

• Fixed an issue where we didn’t handle SSE event headers correctly.

• Fixed an issue where native Bedrock request parsing dropped string message content in Kong’s internal format, causing other plugins to receive invalid data.

• Fixed Azure OpenAI Responses API support for deployment-style routing.

Release date 2026/07/02

Bugfix

• Fixed an issue where Gemini service-account authentication probed metadata.google.internal when a service account JSON was configured.

• Fixed an issue where cache_control was lost in Bedrock Anthropic format requests using ARN model IDs.

• Fixed an issue where Bedrock native ConverseStream requests using ARN model IDs could fail when handling streamed responses.

• Fixed an issue where balancer retries could continue using the last selected targetafter all priority peers were exhausted.

• Fixed an issue where cache_creation_input_tokens and cache_read_input_tokens were excluded from prompt token counts, causing undercounting when Anthropic prompt caching was active.

• Fixed intermittent upstream 401s caused by reused expired Azure managed-identity tokens.

Release date 2026/06/23

Bugfix

• Fixed an issue where Anthropic streaming finish reasons are not mapped in OpenAI to Anthropic conversion.

• Fixed an issue where the Anthropic driver dropped the top_p, top_k, and stop parameters when translating OpenAI-format requests. The driver now also omits temperature, top_p, and top_k for models that reject them (Claude Opus 4.7, Opus 4.8, and Fable), preventing upstream 400 errors.

• Fixed an issue where unsafe model names could construct upstream AI provider paths.

• Fixed an issue where unsafe batch id could cause upstream AI provider path traversal.

• Fixed an issue where an unsafe model name in native format adapters could cause upstream AI provider path traversal.

• Fixed an issue where the xAI driver nested reasoning_effort under reasoning.effort and did not normalize reasoning_content in responses.

• Fixed an issue where native-format requests using a model alias were incorrectly proxying to upstream model.

Release date 2026/06/04

Bugfix

• Fixed an issue where stream_options was forwarded to Databricks upstreams, causing a 400 error since Databricks does not support this field.

• Fixed an issue where full-sync reconfigure reset balancer state even when plugin configs were unchanged.

• Fixed an issue where driver transformers returned errors or stream metadata in the wrong return position.

• Fixed an issue where OpenAI-compatible Anthropic chat requests only preserved the last system message.

• Fixed an issue where configured model name may never be used for Azure provider.

• Fixed an issue where we didn’t collect the model name in OpenAI streaming response from Azure.

Release date 2026/05/25

Bugfix

• Fixed an issue where dashscope replied 500 to inference request.

• Fixed an issue where GenAI spans were not finished sometimes.

• Fixed an issue where we didn’t support Azure GA version of Realtime API.

• Fixed an issue where OpenAI-format chat reasoning requests were not mapped consistently across Anthropic, Gemini, Bedrock and more providers.

• Fixed an issue where system prompt of Claude Code was truncated when doing non-passthrough proxy.

• Fixed an issue where we didn’t use the huggingface’s token usage if available.

Release date 2026/04/28

Bugfix

• Fixed an issue where Claude Code (or CLI) statistics were missing when using an Anthropic model via the Bedrock provider.

• Fixed an issue where observability errors occurred when gpt-image-1.5 returned image token usage details.

• Fixed an issue where Azure OpenAI authentication failed during load balancer failover due to token refresh attempts in balancer_by_lua. Tokens are now pre-fetched in the access phase.

• Fixed an issue where we treated plaintext Anthropic response as gzipped.

• Fixed an issue where we didn’t correctly ungzip streaming response.

Release date 2026/04/07

Feature

• Added native passthrough support for Vertex AI multimodal embeddings. Introduced new llm/v1/multimodal-embeddings route type.

• Added support for databricks provider.

• Added support for DeepSeek provider.

• Added the Ollama provider, for native support for all Ollama-hosted models.

• Added ACL support that blocks requests matching deny rules, requires allow rules to pass when present, and resolves consumer/model values once per request.

• Updated to use /v2/chat when proxying text generation request to cohere provider.

• Added the ability to select target models by request models in same AI Proxy Advanced plugin.

• Added support for proxying cohere embeddings models on bedrock.

• Added support for vLLM provider.

Bugfix

• Fixed structured output and tool calls for Ollama (also Llama2 and Mistral) providers.

• Updated mappings for Ollama (also Llama2 and Mistral) providers.

• Fixed an issue where attributes for GenAI spans were incorrect according to https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-spans.

• Fixed excessive DNS warning logs for template variable endpoints like $(headers.x_gcp_endpoint). These warnings appeared on every configuration change because template variables cannot be resolved until request time.

• Fixed usage.cost for OpenAI llm/v1/files/{file_id}/content batch output analytics.

• Fixed an issue where metrics were missing and v2 API was unavailable in Cohere native format.

• Fixed an issue where id/created fields were missing in multiple drivers.

• Fixed an issue where gzip streaming response in OpenAI format was broken.

• Fixed an issue where content-type header, from non-200 response of streaming request, was overridden.

• Fixed an issue where GCP tokens would not be proactively refreshed before expiration.

• Fixed zero usage statistics for OpenAI llm/v1/responses in streaming mode.

Performance

• decoded OpenAI response only when necessary.

Release date 2026/08/20

Bugfix

• Fixed a bug where a response from Anthropic could fail to decode when the client’s request advertised Brotli support, since Anthropic’s edge may then return a Brotli-encoded body that Kong’s response decompression – which only understands gzip – can’t inflate. Kong now normalizes the outgoing Accept-Encoding request header before proxying to any LLM provider: gzip, identity when the client accepts gzip, and identity otherwise, so a client that cannot inflate gzip is never handed a compressed body on the paths that pass the upstream response through untouched.

Release date 2026/07/27

Bugfix

• Fixed an issue where unsafe model names could construct upstream AI provider paths.

• Fixed an issue where the configured auth header (api-key or x-api-key, for example) leaked the provider credential into access logs instead of being redacted.

• Fixed an issue where unsafe batch IDs could cause upstream AI provider path traversal.

• Fixed an issue where unsafe model names in native format adaptors could cause upstream AI provider path traversal.

• Fixed an issue where the Anthropic driver dropped the top_p, top_k, and stop parameters when translating OpenAI-format requests. The driver now also omits temperature, top_p, and top_k for models that reject them (Claude Opus 4.7, Opus 4.8, and Fable), preventing upstream 400 errors.

• Fixed Azure OpenAI responses support in deployment style.

Release date 2026/06/30

Bugfix

• Fixed an issue where observability errors occurred when gpt-image-1.5 returned image token usage details.

Release date 2026/05/05

Bugfix

• Fixed excessive DNS warning logs for template variable endpoints like $(headers.x_gcp_endpoint). These warnings appeared on every configuration change because template variables cannot be resolved until request time.

Release date 2026/04/08

Bugfix

• Fixed an issue where attributes for GenAI spans were incorrect according to https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-spans.

• Fixed an issue where Azure OpenAI authentication failed during load balancer failover due to token refresh attempts in balancer_by_lua. Tokens are now pre-fetched in the access phase.

• Fixed usage.cost for OpenAI llm/v1/files/{file_id}/content batch output analytics.

• Fixed an issue where metrics were missing and v2 API was unavailable in Cohere native format.

• Fixed an issue where id/created fields were missing in multiple drivers.

• Fixed an issue where gzip streaming response in OpenAI format was broken.

• Fixed an issue where content-type header, from non-200 response of streaming request, was overridden.

Release date 2026/02/01

Bugfix

• Fixed an issue where GCP tokens would not be proactively refreshed before expiration.

Release date 2025/12/18

Feature

• Supported health checks and circuit breaker for the load balancer. Added two new fields max_fails and fail_timeout.

• Added missing least-connections load balancing algorithm to the AI Proxy Advanced plugin which was missed in the previous release.

• Added support for Gemini live websocket in the ai-proxy-advanced plugin.

• Added load balance, failover and circuit breaker feature for semantic routing.

Bugfix

• Fixed balancer retry failures caused by expired DNS entries by preloading DNS for targets.

• Fixed an issue where the token count for Gemini Vertex embeddings API in native format was incorrect.

• Fixed an issue where the token count for Gemini Vertex embeddings API in OpenAI format was incorrect.

• Fixed intermittent 500 responses from the AI Proxy Advanced plugin when using Azure OpenAI

• Fixed an issue where the native format option did not work correctly for non-openai formats.

• Fixed an issue where the semantic load balancing with pgvector namespace was not functioning correctly.

• Fixed an issue when using responses API and background mode.

• Fixed an issue where Files content analytics extraction is not handled properly for Azure.

• Fixed an issue where requests to Anthropic Claude models via Azure Foundry were not being processed correctly.

• Fixed an issue where AWS Bedrock invoke command is not properly proxied.

• Fixed an issue where the Gemini image generation model responses were not being processed correctly.

• Fixed missing id and created fields in certain drivers.

• Fixed an issue where the max_completion_tokens parameter was not being set correctly for O1 series models (e.g., o1, o3, o4, gpt-5).

Release date 2026/09/02

Bugfix

• Fixed a bug where a response from Anthropic could fail to decode when the client’s request advertised Brotli support, since Anthropic’s edge may then return a Brotli-encoded body that Kong’s response decompression, which only understands gzip, can’t inflate. Kong now normalizes the outgoing Accept-Encoding request header before proxying to any LLM provider: gzip, identity when the client accepts gzip, and identity otherwise, so a client that cannot inflate gzip is never handed a compressed body on paths where the upstream response passes through unmodified.

Release date 2026/07/22

Bugfix

• Fixed an issue where unsafe batch id could cause upstream AI provider path traversal.

• Fixed an issue where an unsafe model name in native format adapters could cause upstream AI provider path traversal.

• Fixed an issue where unsafe model names could construct upstream AI provider paths.

Release date 2026/04/15

Bugfix

• Fixed an issue where id/created fields were missing in multiple drivers.

Release date 2026/02/26

Bugfix

• Fixed balancer retry failures caused by expired DNS entries by preloading DNS for targets.

• Fixed excessive DNS warning logs for template variable endpoints like $(headers.x_gcp_endpoint). These warnings appeared on every configuration change because template variables cannot be resolved until request time.

• Fixed an issue where metrics were missing and v2 API was unavailable in Cohere native format.

• Fixed an issue where GCP tokens would not be proactively refreshed before expiration.

Release date 2025/12/10

Bugfix

• Fixed intermittent 500 responses from the AI Proxy Advanced plugin when using Azure OpenAI

• Fixed an issue where Files content analytics extraction is not handled properly for Azure.

Release date 2025/11/18

Bugfix

• Fixed an issue where the token count for Gemini Vertex embeddings API in native format was incorrect.

• Fixed an issue where the token count for Gemini Vertex embeddings API in OpenAI format was incorrect.

• Fixed an issue where the native format option did not work correctly for non-openai formats.

• Fixed an issue where the semantic load balancing with pgvector namespace was not functioning correctly.

• Fixed an issue where AWS Bedrock invoke command is not properly proxied.

Release date 2025/10/01

Feature

• Added latency and cost observability to realtime API.

• Added support for Gemini Rerank in native format.

• Exposed input_tokens_details.cached_tokens_details.image_tokens_count in observability metrics.

Bugfix

• Fixed an issue where the AI Proxy Advanced plugin’s lowest latency load balancer could fail to select a target when only one target was available, leading to 500 errors.

• Fixed an issue where the ai-retry phase was not correctly setting the namespace in the kong.plugin.ctx, causing the ai-proxy-advanced balancer to retry the first target more than once.

• Fixed an issue where the Gemini provider didn’t support Model Garden.

• Fixed an issue where managed identity could be cached incorrectly.

• Fixed an issue where Mistral models would return Unsupported field: seed when using some inference libraries.

• Fixed an issue where array input wasn’t being properly validated for certain providers. Note that Bedrock, Gemini Public, and Mistral providers don’t support array input, as before.

• Fixed an issue where some Titan embeddings models reported malformed requests.

Release date 2026/02/26

Bugfix

• Fixed balancer retry failures caused by expired DNS entries by preloading DNS for targets.

• Fixed an issue where the token count for Gemini Vertex embeddings API in native format was incorrect.

• Fixed an issue where the token count for Gemini Vertex embeddings API in OpenAI format was incorrect.

• Fixed excessive DNS warning logs for template variable endpoints like $(headers.x_gcp_endpoint). These warnings appeared on every configuration change because template variables cannot be resolved until request time.

• Fixed intermittent 500 responses from the AI Proxy Advanced plugin when using Azure OpenAI

• Fixed an issue where the native format option did not work correctly for non-openai formats.

• Fixed an issue where Files content analytics extraction is not handled properly for Azure.

• Fixed an issue where AWS Bedrock invoke command is not properly proxied.

• Fixed missing id and created fields in certain drivers.

• Fixed an issue where the max_completion_tokens parameter was not being set correctly for O1 series models (e.g., o1, o3, o4, gpt-5).

Release date 2025/09/05

Feature

• Support for Gemini Rerank in native format

Bugfix

• Fix panic in LLM observability when populating token usage for AI Proxy token usage details

• AI Proxy Advanced: Fixed an issue where gemini provider not support model garden

• Fixed an issue where Mistral models would return Unsupported field: seed when using some inference libraries.

• Skip unknown cached details key from o11y observation.

• Fixed an issue where array input is not recognized as valid request for some providers. Bedrock, Gemini Public and Mistral provider don’t accept array input as before.

• Fixed an issue where some Titan embeddings model reported malformed request.

Release date 2025/07/28

Bugfix

• Fixed an issue where ai-retry phase was not correctly setting the namespace in the kong.plugin.ctx, causing ai-proxy-advanced balancer retry first target more than once.

• Fixed an issue where AI Proxy and AI Proxy Advanced can’t properly failover to a Bedrock target.

• Fixed an issue where AI Proxy and AI Proxy Advanced might produce duplicate content in the response when the SSE event was truncated.

• Fixed an issue where AI Proxy and AI Proxy Advanced might drop content in the response when the SSE event was truncated.

• Fixed an issue where managed identity may not be cached properly.

Release date 2025/07/16

Bugfix

• Fixed an issue where the llm license migration could fail if the license counter contained more than one week of data.

Release date 2025/07/03

Deprecation

• ai-proxy, ai-proxy-advanced: Deprecated the preserve route_type. You are encouraged to use new route_types added in version 3.11.x.x and onwards.

Bugfix

• Add tried_targets field in serialized analytics logs for record of all tried ai targets.

• Fixed an issue where some of ai metrics was missed in analytics

• Fixed an issue where AI Proxy Advanced can’t failover from other provider to Bedrock.

• Fix consistent-hashing algorithm not using correct value to hash.

• Fixed an issue where the stale semantic key vector may not be refreshed after the plugin config is updated.

• Fixed an issue where AI Proxy and AI Proxy Advanced would use corrupted plugin config.

Known Issues

• If any AI Gateway plugin has been enabled in a self-managed Kong Gateway deployment for more than a week, upgrades from 3.10 versions to 3.11.0.0 will fail due to a license migration issue. This does not affect Konnect deployments.

  A fix will be provided in 3.11.0.1.

  See breaking changes in 3.11 for a temporary workaround.

Release date 2025/10/10

Bugfix

• Fixed an issue where AI Proxy Advanced can’t failover from other provider to Bedrock.

• Fixed model name escape or bedrock from request

Release date 2025/08/07

Bugfix

• Fixed an issue where the Gemini provider could not use Anthropic ‘rawPredict’ endpoint models hosted in Vertex.

• Fixed an issue where ai-retry phase was not correctly setting the namespace in the kong.plugin.ctx, causing ai-proxy-advanced balancer retry first target more than once.

• Fixed an issue where AI Proxy and AI Proxy Advanced can’t properly failover to a Bedrock target.

• Fix consistent-hashing algorithm not using correct value to hash.

Release date 2025/07/06

Bugfix

• Fixed an issue where the stale semantic key vector may not be refreshed after the plugin config is updated.

Release date 2025/04/15

Bugfix

• Fixed an issue where AI Proxy and AI Proxy Advanced would use corrupted plugin config.

Release date 2025/03/27

Breaking Change

• Changed the serialized log key of AI metrics from ai.ai-proxy to ai.proxy to avoid conflicts with metrics generated from plugins other than AI Proxy and AI Proxy Advanced. If you are using logging plugins (for example, File Log, HTTP Log, etc.), you will have to update metrics pipeline configurations to reflect this change.

Deprecation

• Deprecated preserve mode in config.route_type. Use config.llm_format instead. The preserve mode setting will be removed in a future release.

Feature

• Added support for boto3 SDKs for Bedrock provider, and for Google GenAI SDKs for Gemini provider.

• Added new priority balancer algorithm, which allows setting apriority group for each upstream model.

• Added the failover_criteria configuration option, which allows retrying requests to the next upstream server in case of failure.

• Added cost to tokens_count_strategy when using the lowest-usage load balancing strategy.

• Added the huggingface, azure, vertex, and bedrock providers to embeddings. They can be used by the ai-proxy-advanced, ai-semantic-cache, ai-semantic-prompt-guard, and ai-rag-injector plugins.

• Allow authentication to Bedrock services with assume roles in AWS.

• Added the ability to set a catch-all target in semantic routing.

Bugfix

• Fixed an issue where AI upstream URL trailing would be empty.

• Fixed an issue where the ai-proxy-advanced plugin failed to failover between providers of different formats.

• Fixed an issue where the ai-proxy-advanced plugin identity running failed in retry scenarios.

Release date 2025/07/07

Bugfix

• Fixed an issue where AI Proxy and AI Proxy Advanced would use corrupted plugin config.

Release date 2024/12/12

Feature

• Added support for streaming responses to the AI Proxy Advanced plugin.

• Made the embeddings.model.name config field a free text entry, enabling use of a self-hosted (or otherwise compatible) model.

Bugfix

• Fixed an issue where stale plugin config was not updated in dbless and hybrid mode.

• Fixed an issue where lowest-usage and lowest-latency strategy did not update data points correctly.

Release date 2024/11/04

Bugfix

• Fixed an issue where stale plugin config was not updated in dbless and hybrid mode.

• Fixed an issue where lowest-usage and lowest-latency strategy did not update data points correctly.

Release date 2024/09/11

Feature

• allow AI plugin to read request from buffered file

• Added the ai-proxy-advanced plugin that supports advanced load balancing between LLM services.