Changelog

Context Compression automatically summarizes conversation history using a smaller, cost-efficient model before sending requests to your primary model. This significantly reduces input token costs while preserving conversation context.

Highlights

• Up to 78% token savings on long multi-turn conversations
• Three compression strategies to balance quality vs. savings:
• conservative — compresses after 8+ turns (minimal context loss)
• on — compresses after 6+ turns (balanced)
• aggressive — compresses after 3+ turns (maximum savings)
• All endpoints supported:
• POST /v1/chat/completions
• POST /v1/messages
• POST /v1/responses

How to Enable

Option 1: API Key Dashboard (Zero Code Changes)

Go to API Key Management → Edit your API key → Enable Context Compression and select your preferred strategy. All requests using that key will automatically apply compression.

Option 2: Per-Request via Request Body

  {
    "model": "gpt-4.1",
    "messages": [...],
    "compression": {
      "enabled": true,
      "strategy": "on",
      "model": "gpt-4.1-mini"
    }
  }

Option 3: Per-Request via HTTP Headers

X-Context-Compression: on X-Compression-Model: gpt-4.1-mini

SDK Support

Compression examples are now available for all supported SDKs:

• Python SDK (OpenAI, Anthropic, Responses API)
• TypeScript / Vercel AI SDK (@apertis/ai-sdk-provider)
• LangChain (via default_headers)
• LlamaIndex (via additional_kwargs)
• LiteLLM (via extra_headers)

Priority

Request body params > HTTP headers > API key defaults. Per-request settings always override key-level defaults.

See more on **Documentation**

Grok 4.2 is the next major iteration of xAI's Grok series, advancing the model's reasoning, coding, and multimodal capabilities with architectural improvements over Grok 4 and 4.1. It is positioned as a more powerful and general-purpose frontier AI model in the Grok family with stronger deep reasoning and real-world task performance.

The full Qwen 3.5 series is provided at **Apertis Coding Plan** as well, Enjoy it.

Gemini 3.1 Flash Image Preview (also known as "Nano Banana 2") is Google's latest state-of-the-art image generation and editing model, delivering Pro-level visual quality at Flash-level speed. It combines strong contextual understanding with fast, cost-efficient inference, enabling high-quality image generation and seamless iterative editing. Optimized for both performance and accessibility, it makes advanced visual creation workflows faster and more scalable.

New feature: Cached responses now support streaming (SSE) delivery, covering ~80% of API traffic that uses stream: true.

• On cache hit, the system emits synthetic SSE chunks from the stored response — no upstream API call needed
• Content is split on rune boundaries (50 runes/chunk, 10ms intervals) to preserve multi-byte characters
• Proper X-Cache-Hit, X-Cached-Tokens, and X-Actual-Model headers on streaming cache hits
• Non-streaming cache hits continue to work as before (direct JSON response)

Cache Correctness Hardening

• Temperature guard: Only caches requests where temperature: 0 is explicitly present in the raw JSON body. Omitted temperature (Go zero value 0.0) is no longer falsely treated as

cacheable — providers default to ~1.0 for omitted values

• SSE error safety: If synthetic SSE emission fails mid-stream, the handler returns immediately instead of falling through to normal processing, preventing HTTP double-write

corruption

• Tool call exclusion: Responses containing tool_calls are excluded from cache storage since the SSE emitter only supports text content replay

Cache TTL & Infrastructure

• Default prompt cache TTL extended from 10 → 30 minutes

Enjoy it.