> ## Documentation Index > Fetch the complete documentation index at: https://docs.apiyi.com/llms.txt > Use this file to discover all available pages before exploring further. # Does APIYI Support Cache Billing? > Yes. Claude, OpenAI, Gemini, DeepSeek, Qwen, Grok, and other major channels all support cache billing — hit fields are returned untouched and bills follow official discount rates ## Short Answer **Yes.** APIYI's **Claude, OpenAI, Gemini, DeepSeek, Qwen, and Grok** channels all support cache billing: cache-related request parameters are forwarded upstream as-is, cache-hit fields come back to you untouched, and the billing dashboard lists cached usage as separate line items at the official discount rates — no middleware-specific adaptation needed in your code. **Claude and OpenAI cache hits are stable and reliable** (both have dedicated guides on this site — linked below). DeepSeek, Qwen, and Grok all use fully automatic prefix caching and hit normally with a stable prefix — though **Grok's upstream explicitly makes no guarantee of a hit**, so budget at the uncached price. Gemini's implicit caching is supported, but its **hit rate is mediocre** — don't build your cost budget around Gemini caching. ## The Three Channels at a Glance | | OpenAI (gpt-5 series) | Claude | Gemini | | - | - | - | - | | Trigger | **Fully automatic**, zero code | Manual `cache_control` markers | Implicit caching, auto-enabled | | Minimum threshold | 1024 tokens | 1024–4096 tokens by model | 4096 (3 series) / 2048 (2.5 series) | | Write fee | Free | 1.25× (5 min) / 2× (1 hour) | Free | | Hit price | 0.1× of input | 0.1× of input | Per Google's official discount | | Real-world experience | ✅ Stable hits | ✅ Stable hits | ⚠️ Mediocre hit rate | | Full guide | [OpenAI Cache Billing](/en/api-capabilities/openai/prompt-caching) | [Claude Cache Billing](/en/api-capabilities/claude-prompt-caching) | [Gemini Cache Billing](/en/api-capabilities/gemini/prompt-caching) | ## Channel Notes ### OpenAI: fully automatic, zero effort Keep a stable prefix of at least 1024 tokens and hits happen automatically — the matched portion bills at **10% of the input price**, with no write fee, so the 2nd request is already pure savings. How to write requests that hit and how to use `prompt_cache_key`: see the [OpenAI Prompt Caching Billing Guide](/en/api-capabilities/openai/prompt-caching). ### Claude: manual markers, biggest savings Add `cache_control` to the content blocks you want cached; hits bill at **0.1×** (writes cost 1.25× / 2×). Essential for Claude Code, Cline, Cursor, and other heavy workloads. Note it **only works in the Anthropic native format (`/v1/messages`)** — calling Claude through the OpenAI-compatible format gets no cache discount. See the [Claude Prompt Caching Billing Guide](/en/api-capabilities/claude-prompt-caching). ### Gemini: supported, but keep expectations low APIYI auto-enables implicit context caching for the Gemini native format, and hits bill at Google's official discount. In practice, however, **Gemini's cache hit rate is clearly below Claude / OpenAI** (upstream implicit caching behavior is not controllable). Our advice: * Treat the cache discount as a nice-to-have bonus — **estimate costs at the uncached price** * For cache-sensitive workloads with long, frequent prefixes, prefer the OpenAI or Claude channels ## Other Channels: DeepSeek / Qwen / Grok Caching on these channels is **fully automatic** (no markers needed), works normally through APIYI, and performs well in practice: | Channel | Trigger | Hit discount (official) | Hit reliability | | - | - | - | - | | **DeepSeek** | Automatic, prefix matching | Hits save **over 90%** — the steepest discount of the lot | Stable | | **Qwen** | Implicit caching, auto-enabled, prefix of at least 1024 tokens | Hits bill at the official discounted rate | Stable | | **Grok (xAI)** | Automatic, prefix matching | Hits save roughly **75%** (`grok-4.6`; varies by model tier) — see the [Grok cache billing guide](/en/api-capabilities/grok/prompt-caching) | Deterministic when they happen, but no upstream guarantee | Raising hit rates works the same way as with OpenAI: **stable content first, volatile content last** — keep timestamps and random IDs out of the start of the prompt. The "stable prefix" playbook in the [OpenAI Cache Billing Guide](/en/api-capabilities/openai/prompt-caching) applies directly. ## How to Confirm a Hit Check the cache fields in the response `usage`: | Channel | Hit field | | - | - | | OpenAI `/v1/chat/completions` | `usage.prompt_tokens_details.cached_tokens` | | OpenAI `/v1/responses` | `usage.input_tokens_details.cached_tokens` | | Claude `/v1/messages` | `usage.cache_read_input_tokens` | | Gemini native format | `usageMetadata.cachedContentTokenCount` | | DeepSeek | `usage.prompt_cache_hit_tokens` / `prompt_cache_miss_tokens` | | Qwen / Grok `/v1/chat/completions` | `usage.prompt_tokens_details.cached_tokens` | | Grok `/v1/responses` | `usage.input_tokens_details.cached_tokens` | A value above 0 means a hit. In the dashboard call logs, cached usage appears as separate discounted line items you can verify directly. ## Things to Watch **Caching follows the call format**: calling Claude models through the OpenAI-compatible format (`/v1/chat/completions`) cannot get Claude's cache discount — use the native `/v1/messages` format for heavy Claude usage. * Caches are **isolated per model**: switching models (even within the same series) shares nothing * For vendors not listed above (Kimi, etc.), go by the cache fields actually returned in your call logs * Official mechanism details: `platform.openai.com/docs/guides/prompt-caching`, `docs.claude.com/en/docs/build-with-claude/prompt-caching`, `ai.google.dev/gemini-api/docs/caching`, `api-docs.deepseek.com/quick_start/pricing`, `docs.x.ai/developers/advanced-api-usage/prompt-caching` ## Related Docs Automatic caching: 1024-token threshold, 90%-off hits, prompt\_cache\_key routing 75% off on hits, 128-token block granularity, and which endpoint suits long conversations Implicit caching thresholds and what to expect Where to place cache\_control, break-even math, multi-turn techniques How console group multipliers convert to USD prices Inspect per-request token usage and cache billing details ## Contact Us WeCom support QR code Scan to add, or [contact support](https://work.weixin.qq.com/kfid/kfc9adfd5810ece25ec) Cache billing questions and technical support **Support**: [support@apiyi.com](mailto:support@apiyi.com) **Business**: [business@apiyi.com](mailto:business@apiyi.com) This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.