Reka
Listed at https://reka.ai
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | Reka Edge $0.10/M 输入；Reka Flash $0.80/M；Reka Core $2.00/M；Reka Flash Research 按请求计费（$25-$60/千次）。按量付费，无最低承诺。 | per million tokens |
| Output | Reka Edge $0.10/M；Reka Flash $2.00/M；Reka Core $6.00/M。多模态（图像、视频、音频）单独按"每分钟"计费，Flash 模型视频 $0.06/分钟、Core $0.08/分钟；图像按调用计费（Flash $0.01、Core $0.02）。 | per million tokens |
| Cache Read | Reka 文档未公布专用"缓存输入"价格——文本 token 计费不区分缓存与全新；多模态按分钟独立计量。 | Discounted |
🤖 Supported Models (5)
✨ Pros
- ✓Native multimodal: text + image + audio + video in a single API; a single message can mix modalities (Reka Flash/Core/Edge all natively support it)
- ✓Reka Flash pricing is ~1/5 of comparable closed flagships: $0.80/M input vs GPT-4.1-mini $0.40/M, Claude Sonnet 4.5 $3/M; multimodal usage billed per minute (video $0.06/min)
- ✓Reka Edge (reka-edge-2603) targets on-device deployment: vLLM-compatible open weights run 40-50 tok/s on a single GTX 3090, bypassing any cloud subscription
- ✓Reka Flash Research: agentic web-research model (reka-flash-research-20250708), $25/1K standard requests with parallel-thinking for multi-step reasoning
- ✓OpenAI-compatible API: base_url=https://api.reka.ai/v1 — drop-in for OpenAI Python SDK and other compatible clients
- ✓Three product lines — Reka Vision, Reka Research, Reka Speech — plus Reka Edge is now available on OpenRouter for zero-code model switching
- ✓Five co-founders (Dani Yogatama, Cyprien de Masson d'Autume, Mikel Artetxe, Qi Liu, Yi Tay) come from Google DeepMind, Meta AI, and other top labs
⚠️ Cons
- ×Lower awareness: adoption rate is still early vs OpenAI/Anthropic/Google; the third-party SDK, tutorial, and community ecosystems are thinner
- ×Reka Core pricing ($2.00/$6.00) is comparable to Claude Sonnet 4.5 but the brand trust is not yet established — large enterprise contracts still default to incumbent vendors
- ×No mainland-China direct endpoint; api.reka.ai is a single global endpoint and production traffic from China needs a proxy or relay
- ×Reka Flash Research bills per 1K requests ($25-$60), so heavy agent users face unpredictable costs — a single multi-step research task can trigger multiple parallel-thinking charges
- ×Reka Spark appears on the reka.ai product catalog but is not on the public API yet — only Edge / Flash / Core plus Flash Research are exposed
- ×No Chinese-language docs: official site, pricing, and API reference are all English-only, so mainland China teams need strong English-language engineering capacity
- ×No explicit prompt-cache pricing — text tokens are not differentiated by cache hit/miss, so repeat-query cost reduction over long context is limited
🎯 Best For
Teams that need text + image + audio + video behind a single API: they want native multimodal without stitching together OpenAI for vision and Gemini for video, want lower per-token pricing than Claude Sonnet 4.5, and are open to evaluating a less-famous vendor with a strong technical background (co-founders from Google DeepMind / Meta AI).
💰 Pricing & Plans
| Model | Input ($/M) | Output ($/M) | Image (per call) | Video (per min) | Audio (per min) | Notes |
|---|---|---|---|---|---|---|
| Reka Edge | $0.10 | $0.10 | $0.005 | $0.03 | N/A | Ultra-compact; vLLM-compatible open weights; ideal for on-device / edge |
| Reka Flash | $0.80 | $2.00 | $0.01 | $0.06 | $0.015 | Cost-efficient daily-driver; multimodal text+image+audio+video |
| Reka Core | $2.00 | $6.00 | $0.02 | $0.08 | $0.02 | Top-tier multimodal model for complex reasoning over mixed inputs |
| Reka Flash Research (reka-flash-research-20250708) | $25 / 1K requests (standard) | $35 / 1K (parallel-low) | $60 / 1K (parallel-high) | — | — | Agentic web-research model; per-request pricing; multi-step parallel thinking |
🔧 API & Developer Experience
- •API Style: OpenAI-compatible REST API at https://api.reka.ai/v1 — the base_url drop-in lets the official OpenAI Python SDK, curl, Go, Java, and JS clients work with no glue code.
- •Model Surface: Three chat models (Reka Edge, Reka Flash, Reka Core) accessible via /v1/chat/completions and /v1/models; Reka Flash Research is exposed under the /v1/agents-style research endpoint with per-request billing.
- •Native Multimodal: Single message can mix text + image + audio + video; you do not need separate OCR/STT/diarization APIs — Reka's chat completions accept multimodal content blocks directly.
- •Open Weights + vLLM: Reka Edge runs locally via vLLM 0.15.x on NVIDIA GPUs with ≥24GB VRAM (GTX 3090 verified at 40-50 tok/s); the self-hosted path bypasses any api.reka.ai rate limits and is suitable for offline / private deployments.
- •MCP Server: A managed MCP server at https://docs.reka.ai/_mcp/server exposes Reka Vision + chat capabilities to Claude Code, Cursor, and other MCP-aware clients — directly comparable to the Mixedbread Vision MCP.
- •OpenRouter Listing: Reka Edge is available on OpenRouter (per reka.ai/news/switch-models-zero-code-changes-reka-edge-now-available-on-openrouter), so teams can route via OpenRouter for cross-vendor failover without code changes.
- •Auth & Onboarding: Sign up at platform.reka.ai to obtain an X-Api-Key, top up credits, and use the chat completions endpoint immediately; pay-as-you-go with no minimum commitment.
🌐 Native Multimodal: One API for Text, Image, Audio, and Video
Most LLM APIs split multimodal into separate endpoints (OpenAI /v1/audio + /v1/images, Google separate Vision/Audio, Anthropic has no native audio). Reka's bet is that multimodal is the model itself. Reka Edge, Flash, and Core all accept image, video, and audio blocks in one chat completion message — no separate OCR/STT/frame-extraction needed. Pricing mirrors the modality mix: text per million (Edge $0.10/$0.10, Flash $0.80/$2.00, Core $2.00/$6.00), images per call ($0.005-$0.02), video/audio per minute ($0.03-$0.08 and $0.015-$0.02). Flash Research layers agentic web-research on top ($25 per 1K standard; $35-$60 with parallel-thinking), competing with Perplexity Sonar and OpenAI deep-research. Combined with vLLM-deployable Reka Edge open weights (40-50 tok/s on a GTX 3090), the result is a production-ready multimodal stack not locked to any single cloud vendor.
🌍 Regional Availability & Latency
Reka is headquartered at 530 Lawrence Expressway, Sunnyvale, California (Reka, Inc., founded 2022), with engineering in the Bay Area. api.reka.ai is a single global endpoint, so North America and Europe see low first-byte latency; mainland China direct access is not guaranteed and production traffic typically needs a proxy or relay with ~150-250ms cross-Pacific latency. Reka Edge's open weights are a strong workaround: a single NVIDIA GPU with ≥24GB VRAM (verified on GTX 3090) runs the model at 40-50 tokens/sec via vLLM 0.15.x, so teams needing China residency or sub-50ms latency can self-host. Reka Vision has separate per-minute video pricing and an MCP server endpoint. Reka Edge is also on OpenRouter for cross-vendor failover; Flash Research adds agentic search with US-hosted inference.