> For clean Markdown of any page, append .md to the page URL. > For a complete documentation index, see https://docs.mixlayer.com/llms.txt. > For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.mixlayer.com/_mcp/server. # Platform changelog > Major features, performance improvements, and bug fixes for the Mixlayer platform Major features, performance improvements, and bug fixes for Mixlayer’s APIs, agent workflows, and platform. ## October 30, 2026 * **Certificate authentication for workloads.** Exchange X.509 client certificates for short-lived Mixlayer access tokens. Organizations can use certificate identities and policy rules to control access to the Platform API and inference. * **Usage analytics by model and region.** Compare request volume and token usage over time by model, gateway region, or inference region. Filter regional traffic to understand where requests enter Mixlayer and where inference runs. * **Australia and UK gateway regions.** Production gateways are now available in Australia (`au-east`) and the United Kingdom (`uk-south`). Choose a regional endpoint to control where requests enter Mixlayer; inference model availability varies by region. * **Credit purchases in the console.** Select a saved card or add a new card and billing address directly in the credit purchase dialog. The selected card becomes the default for future purchases and automatic refills. * **Fix: Dependency security.** Fixed critical and high-severity vulnerabilities in platform dependencies. * **Fix: Service resilience.** Fixed a storage-exhaustion issue that caused service interruptions and spread production messaging services across availability zones to reduce dependence on a single zone. * **Fix: Final completion usage.** Fixed missing or incorrect completion usage totals when final token counts arrived after generation finished. Streaming and non-streaming completion responses now wait for those counts before finalizing usage. ## October 18, 2026 * **Reasoning token budgets.** Set `reasoning.max_tokens` in Chat Completions and Responses requests to cap reasoning on compatible models. Once that budget is reached, generation moves to the answer using the remaining overall output budget. * **Classification and decisions APIs.** Turn shared text into structured decisions with predicate, choice, and scoring questions. The new `/v1/decisions` endpoint accepts multiple questions in one request and returns answers in the same order, making it easier to build classification, triage, and routing workflows. * **Longer embedding inputs.** Process longer documents on compatible embedding models, up to their configured context limit. Inputs can span multiple processing batches, so the size of a single batch no longer limits those requests. * **Fix: Memory stability.** Fixed memory leaks that accumulated during sustained inference workloads. ## October 11, 2026 * **Workload identity federation.** Authenticate services and automated jobs using credentials from an OIDC identity provider. Exchange those credentials for short-lived Mixlayer tokens with organization-defined permissions for the Platform API and inference. * **Manage Mixlayer through MCP.** Connect MCP-compatible assistants and tools to the Platform API to manage organizations, members, and API keys, and access usage analytics and billing information. Actions use the authenticated caller’s permissions. ## October 4, 2026 * **Regional inference routing.** Choose where inference runs with organization routing policies and per-request region preferences. Set a preferred region order, require an explicit region, and control whether fallback stays within the same country or uses other eligible regions. HTTP response headers identify the gateway and inference regions. * **Fix: Prepaid credit access.** Fixed an issue that could ignore valid prepaid credits associated with an active billing contract and block inference access. ## September 27, 2026 * **Fix: Requests during restarts.** Fixed shutdown behavior that could interrupt active requests during service restarts. Services now stop accepting new work and allow active requests and connections time to finish. ## September 20, 2026 * **Structured answers with reasoning.** Combine reasoning with JSON Schema output constraints: the model can reason before producing a final answer in the requested structure. API usage details also report reasoning tokens, making it easier to understand how the output budget was used. * **Faster generation with mixed sampling settings.** Improved throughput for concurrent speculative generation requests with different sampling settings by keeping eligible requests on the GPU and reducing data transfer overhead. * **Faster embedding processing.** Optimized GPU operations reduce computation and memory-copy overhead for compatible embedding workloads. ## September 13, 2026 * **Faster structured JSON generation.** JSON Schema requests can now use speculative decoding on compatible models, accelerating generation when draft tokens are accepted while preserving the requested output constraints. * **Fix: Large prompts.** Fixed failures on large prompts caused by an unnecessary internal copy of the prompt exceeding a message-size limit. ## September 6, 2026 * **Image results from tools.** Tools can return images alongside text through the Responses API when using a model that accepts image input. Agents can use screenshots, charts, and other visual results in subsequent steps of a workflow. * **Promotional credits.** Redeem promotional codes and view active credits in the console. Eligible credits are applied to your billing account for use during their validity period. * **More responsive streaming under load.** Ongoing generation is scheduled ahead of incoming prompt processing, helping reduce pauses in streamed responses when long prompts compete for capacity. ## August 30, 2026 * **Fix: Sampling correctness.** Fixed random-number handling in batched generation so each request uses its own random-number sequence even when its position in a batch changes, including during speculative generation. ## August 23, 2026 * **Higher embedding throughput.** Compatible embedding requests are now processed together, improving throughput when multiple requests arrive at the same time. ## August 16, 2026 * **Fix: Prompt cache correctness.** Fixed a prompt-cache issue that could reuse stale context after cache memory was recycled. Cached prefixes are now validated across the entire chain before reuse. ## August 9, 2026 * **More control over agent tools.** Group related functions into namespaces in the Responses API. Use `parallel_tool_calls: false` in Responses or Chat Completions requests to limit each generation step to one tool call. ## August 2, 2026 * **Hosted web search.** Give agents access to current web information through Mixlayer’s hosted `web_search` tool. Mixlayer runs the search and feeds the results back into the conversation, reducing the integration work needed to build search-enabled assistants. * **Faster text generation.** Token selection now runs in batches on the GPU, reducing data transfers during generation. The same acceleration applies to JSON Schema outputs on compatible models. * **Fix: Detecting truncated output.** Fixed token-limited outputs being reported as complete. Completions now return `finish_reason: "length"`, and Responses returns `status: "incomplete"`, so applications can detect truncated output. ## July 19, 2026 * **Images and PDFs in inference requests.** Send images and PDFs alongside text through the Responses API on compatible models. Combine images and documents with tool calling or structured output to build document analysis and extraction workflows. * **Organization retention controls.** Choose how long newly stored responses are retained, up to 90 days, or disable new response storage. Configure the policy at the organization level to control conversation persistence. * **Organization audit logs.** Authorized administrators can review API key, membership, billing, and organization changes through the Platform API. Audit entries identify the actor, action, affected resource, and source IP when available. ## July 5, 2026 * **Fix: Request cancellation.** Fixed streaming Responses requests continuing to wait for output after a client disconnected. Disconnections now trigger request cleanup without waiting for the next output chunk. ## June 28, 2026 * **Embeddings API.** Generate vectors for semantic search, retrieval, and clustering through an OpenAI-compatible `/v1/embeddings` endpoint. Send individual inputs or batches and choose the output dimensions and encoding. * **Scoped API keys for platform automation.** Use API keys to automate Platform API operations as well as inference. Assign read, write, or administrative API permissions independently of inference access. * **Fix: Recovery after allocation failures.** Fixed an issue where a failed oversized request could leave available inference memory inaccessible to subsequent requests. Available capacity is now restored after the failed allocation. ## June 21, 2026 * **Responses API and conversation continuity.** Build agent workflows with the Responses API over HTTP or WebSockets, including streaming text and tool calls. Opt in to response storage and continue a conversation using `previous_response_id`. * **Prepaid billing and spend controls.** Purchase credits, configure automatic refills, and review purchase history. Set low-balance warnings, usage-spend alerts, and spend cutoffs to help manage your account’s usage. * **Fix: Accurate token usage.** Fixed missing or incorrect tool-token and cached-token counts in completion usage. Streaming and non-streaming endpoints now report the inference service’s authoritative totals. ## June 14, 2026 * **Streaming tool calls.** Receive tool names and arguments incrementally through streaming Chat Completions responses. Applications can show tool activity as it develops instead of waiting for the complete call. * **Fix: Long chat histories.** Fixed request failures when large chat histories filled the gateway’s request queue. Processing now waits for available queue capacity instead of failing immediately. ## May 31, 2026 * **Fix: Billing account setup.** Fixed an account setup error that prevented billing contracts from being created. ## May 24, 2026 * **Automatic prompt prefix caching.** Reuse cached computation when requests share the same prompt prefix, reducing repeated prompt processing when a cache entry is available. API usage details report cached input tokens so you can track reuse. ## May 17, 2026 * **Team and organization management.** Invite teammates, manage membership and owner roles, and update organization settings through the Platform API. Transfer ownership or leave an organization as responsibilities change. ## May 3, 2026 * **Request and usage analytics.** Explore request volume, token usage, latency, and errors through the analytics API. Filter traffic, view aggregate trends, and inspect individual request records to understand application behavior. * **API key rotation and expiration.** Set expiration times and rotate existing API keys to replace credentials. Manage credential lifetimes alongside key names and active status. ## April 26, 2026 * **Fix: ModelSocket sequence hangs.** Fixed a hang when an empty append added no new tokens to a ModelSocket sequence. Empty appends now complete so the sequence can continue. ## April 19, 2026 * **Structured JSON output.** Request a JSON object or provide a JSON Schema through Chat Completions’ `response_format`. Constrain the generated answer to the structure your application expects for extraction and other structured workflows. ## February 1, 2026 * **Fix: Abrupt disconnections.** Fixed cleanup after abruptly closed Completions and Chat Completions connections, releasing the associated inference sequence instead of leaving work running. ## November 9, 2025 * **Fix: Completion usage metering.** Fixed duration metering for Completions and Chat Completions requests so sequence time is excluded from their billable usage.