APIRank
EN 中文

Reka Multimodal API Pricing 2026: Edge, Flash, Core Models & Verified Costs

The conversation about what an "AI API" should look like in 2026 has split along a fault line that nobody is talking about openly: most providers now bolt multimodal on as a separate endpoint (OpenAI ships /v1/audio + /v1/images as adjacent products, Google splits Vision and Audio, and Anthropic still has no native audio). Reka (legal entity Reka, Inc., Sunnyvale, California; founded 2022 by five researchers from Google DeepMind and Meta AI including Yi Tay and Mikel Artetxe) has taken the opposite bet: that multimodal is not a feature flag, it is the model. Reka Edge, Reka Flash, and Reka Core all natively accept image, video, and audio blocks in one chat completion message, so a developer does not need to wire OCR, STT, or video frame extraction on top of the call. This review covers the verified per-token and per-minute pricing on docs.reka.ai/pricing (captured 2026-08-26), the vLLM-deployable Reka Edge open weights (40-50 tok/s on a GTX 3090 per docs.reka.ai/quickstart), and where Reka fits between GPT-4.1, Claude Sonnet 4.5, and Gemini.

🌍 Quick verdict: Reka is a Bay Area multimodal foundation model vendor with five native-modality chat models (Reka Edge $0.10/$0.10 per 1M, Reka Flash $0.80/$2.00, Reka Core $2.00/$6.00), per-minute video and audio pricing on top, and an agentic web-research model (Reka Flash Research) at $25 per 1K standard requests. Pricing is pure pay-as-you-go with no minimum commitment. Reka Edge is the differentiator: vLLM-compatible open weights run locally on a single NVIDIA GPU (≥24GB VRAM, GTX 3090 verified at 40-50 tok/s), bypassing the cloud entirely. The trade-off: Reka is less famous than OpenAI/Anthropic/Google, so the third-party SDK and tutorial ecosystem is thinner; there is no mainland-China direct endpoint; and Reka Spark appears on the product catalog but is not yet exposed via the public API.

What Reka actually is

Reka was founded in 2022 by five researchers from Google DeepMind, Meta AI, and other labs (co-founders: Dani Yogatama, Cyprien de Masson d'Autume, Mikel Artetxe, Qi Liu, Yi Tay). The company is headquartered at 530 Lawrence Expressway, Sunnyvale, California, and is registered as Reka, Inc. (verified via the Organization JSON-LD on reka.ai). Reka's product catalog splits into two groups:

  • Foundation models: Reka Spark (ultra-compact, edge device-targeted), Reka Edge (compact, on-device execution), Reka Flash (cost-efficient daily driver), Reka Core (top-tier complex reasoning), and Reka Flash Research (agentic web-research).
  • Vertical platforms: Reka Vision (video + image search, Q&A, and generation), Reka Research (complex multi-step research), Reka Speech (audio understanding beyond transcription).

Reka Spark is currently listed on the reka.ai product catalog but is not exposed via the public API (verified by listing /v1/models on the public endpoint — only reka-flash and reka-edge appear in the baseline set, per docs.reka.ai/chat/models.md). The Chat Completions surface is the production path: base_url=https://api.reka.ai/v1 with X-Api-Key as the auth header. Reka Vision has a separate per-minute pricing tier (covered below) and exposes a managed MCP server at https://docs.reka.ai/_mcp/server for Claude Code, Cursor, and other MCP-aware clients.

Reka pricing 2026: verified from docs.reka.ai

The pricing structure on docs.reka.ai/pricing.md (captured 2026-08-26) is per model + modality. Reka Chat tokens are billed per million; multimodal usage (image, video, audio) is billed per call or per minute, separately from text tokens:

  • Reka Edge: input $0.10 / output $0.10 per 1M tokens. Image $0.005 per call. Video $0.03 per minute. No audio. Ultra-compact model targeted at on-device execution; vLLM-compatible open weights.
  • Reka Flash: input $0.80 / output $2.00 per 1M tokens. Image $0.01 per call. Video $0.06 per minute. Audio $0.015 per minute. Cost-efficient model for most everyday tasks.
  • Reka Core: input $2.00 / output $6.00 per 1M tokens. Image $0.02 per call. Video $0.08 per minute. Audio $0.02 per minute. Top-tier model for complex reasoning tasks.
  • Reka Flash Research (reka-flash-research-20250708): $25 per 1K standard requests, $35 per 1K with parallel-thinking (low), $60 per 1K with parallel-thinking (high). Agentic web-research model that supports complex multi-step reasoning across the web.

The price positioning relative to closed flagships is striking. Reka Flash at $0.80/M input is roughly 5x cheaper than Claude Sonnet 4.5 ($3/M input) and on par with GPT-4.1-mini ($0.40/M input) — but Reka Flash natively accepts image, video, and audio blocks in one message, whereas Anthropic has no native audio and OpenAI splits modalities across separate endpoints. Reka Core at $2.00/$6.00 is comparable to Claude Sonnet 4.5 on text pricing, but Reka Core includes image and video in the same call surface. The pay-as-you-go model has no minimum commitment: sign up at platform.reka.ai, top up credits, and use the API. There is no permanent free tier.

Multimodal pricing: per-minute video and audio

The per-minute pricing for video and audio is unusual in the industry. Most LLM APIs either bundle multimodal into per-token pricing (Gemini) or charge per image/file upload (OpenAI Vision). Reka's per-minute pricing is calibrated for video and audio workloads where the content length drives the cost, not the prompt size:

  • Video processing: Reka Edge $0.03 per minute, Reka Flash $0.06 per minute, Reka Core $0.08 per minute. A 10-minute video clip analyzed by Reka Core costs $0.80.
  • Audio processing: Reka Flash $0.015 per minute, Reka Core $0.02 per minute. A 30-minute podcast analyzed by Reka Core costs $0.60. (Reka Edge does not support audio.)
  • Image processing: Reka Edge $0.005 per image, Reka Flash $0.01 per image, Reka Core $0.02 per image — billed per call regardless of image resolution.

The pricing structure rewards the right model for the workload: use Reka Edge for short clips or low-resolution frames (the same call surface handles image, short video, and text), Reka Flash for typical multimodal workloads, and Reka Core only when complex multi-step reasoning over long video or high-fidelity audio is required. The Reka Vision platform has separate per-minute video pricing tiers (docs.reka.ai/vision/pricing.md) for video search, Q&A, and generation use cases.

Reka Edge open weights: vLLM at 40-50 tok/s on a single GPU

Reka Edge is the differentiator that justifies a deeper look. Per docs.reka.ai/quickstart.md, Reka Edge runs locally via vLLM 0.15.x on NVIDIA GPUs with ≥24GB VRAM — the docs explicitly cite GTX 3090 GPUs at 40-50 tokens/sec for local serving. This is the path for teams that need:

  • China residency or strict data sovereignty — self-host the model on your own infrastructure, no api.reka.ai calls required.
  • Sub-50ms latency — local inference on a single GPU is faster than any cloud roundtrip.
  • Offline operation — Reka Edge is small enough to run on a workstation with one consumer GPU.
  • No per-token billing surprises — pay for the GPU, not the API calls.

The hardware requirement (NVIDIA GPU, ≥24GB VRAM, Linux; macOS is not supported for serving) is the main constraint. The supported stack is NVIDIA CUDA + vLLM 0.15.x (the docs explicitly reject vLLM 0.16.0 and Python outside 3.10-3.14). For teams that already have NVIDIA hardware in their stack, Reka Edge is a strong option for multimodal workloads that would otherwise require OpenAI Vision + GPT-4 + Whisper stitched together.

OpenAI compatibility and the OpenRouter listing

Reka's Chat API is OpenAI-compatible at https://api.reka.ai/v1. The official OpenAI Python SDK works as a drop-in — change base_url to the Reka endpoint, pass your X-Api-Key as api_key, and the chat completions call goes through unchanged. Reka Edge is also listed on OpenRouter (per the reka.ai news post titled "Switch models with zero code changes: Reka Edge now available on OpenRouter"), so teams that route through OpenRouter can add Reka Edge to their failover rotation without touching application code.

The combination — OpenAI-compatible API + OpenRouter listing + vLLM-deployable open weights — makes Reka unusually portable. A team can start with the hosted API for prototyping, switch to self-hosted Reka Edge for production, and route through OpenRouter for cross-vendor failover, all without changing the application code that calls the model.

Reka Flash Research: agentic web-research at $25 per 1K requests

Reka Flash Research is Reka's entry in the agentic-search category that Perplexity Sonar and OpenAI deep-research occupy. The pricing is per 1K requests: $25 standard, $35 parallel-thinking (low), $60 parallel-thinking (high). The model is reka-flash-research (version reka-flash-research-20250708) and is documented as supporting "complex, multi-step research across the web."

The per-request pricing differs from token-based models. A single multi-step research task can trigger multiple parallel-thinking charges if the agent takes several reasoning steps. For predictable cost, the application should cap the number of parallel-thinking steps per request. Perplexity Sonar Pro's API is priced at $3 input / $15 output per 1M tokens, which can be cheaper or more expensive than Reka Flash Research depending on the prompt size and reasoning depth — the comparison is not straightforward.

Reka Vision and Reka Speech

Beyond the Chat API, Reka ships three vertical platforms:

  • Reka Vision (docs.reka.ai/vision): video + image search, Q&A, captioning, detection, embeddings, and highlight-clip generation. Has its own per-minute video pricing tier and an MCP server endpoint for AI clients.
  • Reka Research: complex multi-step research workflows, available as part of Reka Flash Research.
  • Reka Speech: audio understanding beyond transcription — extracts meaning, context, and insights from audio sources.

The Vision API is the most production-ready of the three. It supports video upload, search, Q&A, highlight-clip generation, metadata tagging, and image management, and it exposes all of these capabilities through a managed MCP server that Claude Code, Cursor, and other AI clients can connect to. For teams building AI video workflows, the Reka Vision MCP is a useful comparison point against the Mixedbread Vision MCP.

China access and regional availability

Reka is a US-based company with a single global endpoint at api.reka.ai. There is no mainland-China direct endpoint, so production traffic from China requires a proxy or relay — typical cross-Pacific first-byte latency is ~150-250ms, similar to other US-based AI platforms. Reka Edge's open weights are the workaround for China residency requirements: 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 that need strict data sovereignty can self-host the model on their own infrastructure.

For North America and Europe, the hosted API at api.reka.ai sees low first-byte latency. The Reka Vision API runs on separate infrastructure with per-minute video pricing. Reka Edge on OpenRouter routes through OpenRouter's standard infrastructure, so the latency profile depends on the OpenRouter route.

Who Reka is for

Reka is a strong pick for three specific use cases:

  1. Multimodal-in-one-call applications: a single API call handles text + image + video + audio without stitching together OpenAI for vision, Whisper for audio, and a separate video frame extractor. The per-minute video and audio pricing is calibrated for content-length workloads.
  2. Cost-sensitive multimodal workloads: Reka Flash at $0.80/M input is ~5x cheaper than Claude Sonnet 4.5 and on par with GPT-4.1-mini, with native multimodal in the same call.
  3. Self-hosted multimodal inference: Reka Edge's open weights via vLLM (40-50 tok/s on a GTX 3090) are one of the few production-ready multimodal paths that does not require cloud API calls — useful for China residency, data sovereignty, or sub-50ms latency requirements.

Reka is a weaker pick when:

  • You need maximum brand trust: Reka is less well-known than OpenAI, Anthropic, or Google, so enterprise contracts often default to incumbent vendors. The five co-founders' credentials (Yi Tay at Google DeepMind, Mikel Artetxe at Meta AI, etc.) help, but the brand has not yet reached the same recognition.
  • You need Chinese-language support: all docs, pricing, and API references are English-only.
  • You need prompt caching at discounted rates: Reka's pricing does not differentiate cached vs. fresh text tokens, so long-context repeat-query cost optimization is limited compared to OpenAI or Anthropic.

Bottom line

Reka is the most credible multimodal-first foundation model vendor in 2026. The pricing on docs.reka.ai/pricing.md is verified (captured 2026-08-26): Reka Edge $0.10/$0.10 per 1M tokens, Reka Flash $0.80/$2.00, Reka Core $2.00/$6.00, plus per-minute video and audio pricing and per-request Reka Flash Research ($25-$60 per 1K). The unique differentiator is Reka Edge's open weights, which run on a single NVIDIA GPU at 40-50 tokens/sec via vLLM — the strongest workaround for teams that need multimodal AI without cloud API dependency. The trade-off is exactly the right shape: Reka competes with OpenAI and Anthropic on the same multimodal workloads but ships the model itself openly and supports self-hosting, so a team can start with the hosted API and migrate to on-prem as cost or compliance demands. The authoritative sources are the pricing page, the quickstart, the reka.ai product catalog, and the Reka Edge-on-OpenRouter news post.

If you are routing Reka behind a single OpenAI-compatible key with cross-region failover, FreeModel is the simplest handoff: one dashboard, one billing relationship, and routing controls without glue code.