Reka AI pricing

About

Reka AI is a foundation-model lab that builds natively multimodal (omni) models — single models that ingest and generate text, images, video and audio inside one architecture rather than bolting vision onto a text core. The public API catalog spans a compact on-device tier (Reka Edge), a fast cost-efficient workhorse (Reka Flash), and a flagship for complex tasks (Reka Core), all billed per million tokens. On top of the base models, Reka sells two product-shaped surfaces with their own meters: Reka Research, a multi-step web-research agent billed per 1,000 requests, and Reka Vision, a video- and image-intelligence API billed per input minute and per image on a credit-based Developer tier, with a sales-quoted Enterprise tier.

As of mid-2026 the company positions itself as “building models and infrastructure for the physical AI era” — an AI research lab of 50+ researchers and engineers from Google DeepMind and Meta building foundational intelligence for the physical world, aimed at intelligent robots and next-generation wearables. Its ecosystem is now four named lines: Labs (research), Infer (the inference engine and multimodal API), Vision (video processing infrastructure, available via API, MCP or app), and Claru (data infrastructure — egocentric video, robotics trajectories, world-model footage and expert human judgment for training frontier AI). Only the Infer/Chat, Research and Vision surfaces carry public prices; everything else routes through a Request a Demo motion.

Reka was founded in 2022–2023 by a group of senior researchers from Google DeepMind, Google Brain and Meta FAIR — CEO Dani Yogatama, chief scientist Yi Tay, plus Cyprien de Masson d’Autume, Qi Liu and Mikel Artetxe. The company’s most consequential lifecycle beat came in May 2024, when Snowflake’s reported $1 billion-plus talks to acquire Reka broke down with no deal. Rather than fold into a data-cloud platform, Reka stayed independent — and in July 2025 raised a $110M round that lifted its valuation above $1B, with Nvidia and, notably, Snowflake itself joining as investors. The would-be acquirer became a strategic backer.

That independence shapes the pricing. Reka monetizes like a developer-first inference provider: raw, public per-million-token rates across the model family, per-action meters for Research and Vision, and a sales-gated enterprise track that removes rate limits and shifts to monthly invoicing with bulk discounts. It competes with multimodal-first peers and the closed-weight majors — see Mistral AI and OpenAI for adjacent two-track and per-token models — but leans harder on enterprise on-prem deployment for security-sensitive verticals (defense, media, security).

Pricing summary : a public multimodal usage meter with three units

Reka AI runs a pure usage-based model across three distinct meters, plus a sales-led on-prem track:

• Reka Chat API (per-token + per-multimodal-unit) — separate input/output rates per 1M tokens by model: Reka Edge $0.10 / $0.10, Reka Flash $0.80 / $2.00, Reka Core $2.00 / $6.00. Multimodal inputs carry their own charges — images per image ($0.005–$0.02), video per minute ($0.03–$0.08/min), audio per minute ($0.015–$0.02/min, N/A on Edge).
• Reka Research (per-request) — billed per 1,000 web-research requests, not per token: $25.00 standard, $35.00 parallel-thinking low, $60.00 parallel-thinking high on reka-flash-research. The price scales with how much parallel reasoning a query triggers.
• Reka Vision (per-video-minute + per-image) — 3 free hours (180 indexed minutes), then $0.05 per input video minute to index, $0.005 per video search, $2 per 1M QA/tagging output tokens, and $0.04 per input minute for clip generation (extra $0.01 per input minute for 1080p). Images are a separate meter: $10 per 1M uploaded, $50 per 1M images per month stored, $0.003 per image search. Video storage auto-deletes after 30 days and is free.
• Vision tiers — the self-service Developer tier runs on a credit-based system with rate limits and standard rates; the Enterprise tier removes rate limits, adds bulk discounts, flexible storage with recurring costs, dedicated support and monthly invoicing billed at end of month — contact sales, no public floor.

What makes this different: Reka exposes a per-million-token meter that is genuinely multimodal — text, image, video and audio each carry their own line — and layers two non-token meters (per-request Research, per-minute Vision) on top, so the same lab prices three different value units instead of forcing everything through tokens.

Pricing by product

Reka Chat API — per-1M-token + multimodal units (USD)

Three models on the public card as of 2026-07-22 — Edge, Flash and Core. Reka Core is the flagship and carries a 3x output-over-input premium ($6.00 vs $2.00 per 1M); Reka Edge is the only tier with no audio rate. Multimodal units are billed in addition to text tokens, and there is no upfront cost — “you only pay for what you use.”

Reka Research — per-request web research (USD)

All Research prices are quoted per 1k requests — a usage unit that maps to a completed research task rather than raw inference. The pinned build is reka-flash-research-20250708.

Reka Vision — Developer (self-service) tier pricing (USD)

Reka Vision — Developer vs Enterprise tiers

Sales motions across products: PLG / self-serve for the per-token Reka Chat API, Reka Research and the Reka Vision Free/Developer tiers; sales-led (“Request a Demo”, [email protected]) for the Vision Enterprise tier and the unpriced Labs, Infer and Claru lines.

Hidden costs : What Reka AI users actually pay

Reka’s headline token rates are public and low, but the real bill is shaped by three things the per-token line doesn’t show: the multimodal surcharges (images, video and audio are billed on top of tokens), the output-token premium on Core and Flash, and the fact that Research and Vision are separate meters entirely. Two archetypes show how the total assembles.

Archetype 1 — a video-understanding agent on Reka Core + Vision. A team indexes 20,000 minutes of video a month through Reka Vision, runs 50,000 semantic searches, and answers questions with Reka Core (assume ~30M input + ~8M output tokens/mo plus the video minutes the model reasons over).

The lesson: for video workloads the per-minute Vision meter dominates the token lines, and on Core the $6/M output rate is 3x the input rate, so verbose generation costs more than ingestion. The video minutes the model reasons over (at $0.08/min on Core) stack on top of the Vision indexing fee — the same footage can be metered twice across two products.

Archetype 2 — a research desk running Reka Research. A small analyst team fires 8,000 multi-step web-research requests a month, half on standard and half on parallel-thinking-high for the hard questions.

Here the surprise is the 2.4x spread between standard and parallel-thinking-high ($25 vs $60 per 1k): the same product costs very differently depending on how much parallel reasoning each query invokes, so request mix — not request count — drives the bill. There’s no token line to reconcile, which makes Research unusually easy to forecast once the mode mix is known.

Want to estimate your own Reka AI bill? Use the Reka AI pricing calculator to model your costs based on token volume, video minutes and research requests.

Pricing evolution : Reka AI pricing history and changes

Reka’s pricing has widened from a single per-token chat API into three meters. The text API has billed per million tokens since the model family launched; Reka then added a per-request Research meter and a per-minute Vision meter, while keeping enterprise on-prem deployment sales-led. The dated milestones below are reconstructed from primary docs and contemporaneous press; quarter-level cadence will be tightened with archived snapshots on a later pass.

Cadence

Tracked range: 2023 Q3–2026 Q2. Quarters not listed had no publicly announced price or SKU change. Dated milestones below cite primary/secondary sources; per-snapshot price reconstruction is a later pass.

Notable changes

• 2023 — Reka founded by ex-DeepMind, Google Brain and Meta FAIR researchers to build natively multimodal models.
• 2024-02 — Reka Flash, Edge and Core launch with a public per-token API plus per-image/video/audio units.
• 2024-05 — Snowflake’s reported $1 billion-plus acquisition talks collapse with no deal; Reka continues independently (Bloomberg).
• 2025-07 — $110M round triples valuation past $1B; Nvidia and Snowflake invest (SiliconANGLE).
• 2025 H2 — Reka Research launches with per-request pricing ($25–$60 per 1,000 requests).
• 2025 H2 — Reka Vision launches with per-video-minute pricing and a 180-minute free tier.
• 2026 Q1 — Reka Spark added as the new cheapest model at $0.05/$0.05 per 1M tokens.

The Snowflake near-acquisition in detail

The most important pricing event in Reka’s history is a deal that didn’t happen. In May 2024, Snowflake was reportedly close to acquiring Reka for more than $1 billion. Had it closed, Reka’s pricing would likely have been absorbed into Snowflake’s data-cloud consumption model — credits, not a standalone public token API. When the talks collapsed, Reka stayed an independent inference provider, which is why it still publishes raw per-million-token rates and sells on-prem weight deployment directly. The July 2025 round — with Snowflake itself returning as an investor — funded that independent path. The lifecycle lesson is that staying independent preserved Reka’s developer-first, publicly-priced posture instead of converting it into an enterprise-only line item inside a larger platform.

What’s unique : Reka AI’s distinctive pricing mechanics

1. A genuinely multimodal token meter. Most “multimodal” APIs still bill everything as tokens. Reka’s chat API charges text per 1M tokens and images per image ($0.005–$0.02), video per minute ($0.01–$0.08), and audio per minute ($0.005–$0.02) as distinct lines. The bill mirrors the modality mix of the workload, so a video-heavy app and a text-heavy app on the same model pay very differently — a more honest usage-based shape for multimodal AI.

2. Three meters, one lab. Reka prices three different value units: per-token (Reka Chat), per-request (Reka Research, $25–$60/1k), and per-video-minute (Reka Vision, $0.05/min). Rather than forcing research and video through a token meter, it matches the unit to the job — a research task is a request, a video is a stream of minutes. Few labs run more than one billing primitive this deliberately.

3. On-prem weight deployment as a priced sovereignty upgrade. Because Reka ships deployable model weights, enterprises can run the models entirely on their own infrastructure (on-prem, private cloud, or edge) for defense, security and media. That data-residency option is a quoted enterprise dimension a closed-weight API cannot offer — sovereignty becomes a sellable upgrade rather than a compliance checkbox.

Strengths & weaknesses

Billing UX : usage tracking and overage controls

• Pay as you go — the docs headline the API as “Get started with no upfront costs. You only pay for what you use,” so Chat spend tracks usage directly with no seat or minimum commitment.
• Credit-based Developer tier (Reka Vision) — the self-service Vision tier is explicitly a credit-based system with rate limits, so balance is pre-funded and consumption draws down against it rather than accruing an open-ended invoice.
• Reka Vision free allowance — 3 free hours of indexed video (180 minutes) covering indexing, search and Q&A; video storage auto-deletes after 30 days and is free, so idle footage doesn’t accrue cost.
• Storage auto-delete as a cost control — the 30-day auto-delete applies to both video (free) and images ($50 per 1M images per month while retained), capping retention spend by default instead of requiring manual cleanup.
• Per-mode Research controls — choosing standard vs parallel-thinking (low/high) is an explicit cost lever ($25.00 / $35.00 / $60.00 per 1k requests) the caller sets per query.
• Rate limits as the tier boundary — Developer has rate limits, Enterprise has none; the Vision docs publish a dedicated Rate Limits page, so throughput ceilings are the visible gate between self-serve and sales-led.
• Enterprise monthly invoicing — the Enterprise tier shifts to monthly invoicing billed at end of month, with bulk discounts, flexible storage options carrying recurring costs, and dedicated support.
• Multimodal line-item visibility — because each modality and endpoint is a separate charge (per token, per image, per video minute, per search, per clip minute), usage can be attributed per surface rather than buried in a blended token figure.

Strategic wins : Why Reka AI’s pricing decisions worked

1. Matching the meter to the modality

By charging images per image and video and audio per minute — on top of text tokens — Reka built a price sheet that actually reflects multimodal cost drivers. A video workload pays for video; a text workload pays for text. This is a more defensible usage metric choice than flattening everything into tokens, and it lets Reka lean into its multimodal differentiation at the billing layer, not just the model layer.

2. Three meters instead of one forced primitive

Reka resisted the temptation to price Research and Vision as token products. A research task is billed as a request ($25–$60/1k) and a video as minutes ($0.05/min), because those are the units buyers think in. Picking the right unit per product makes each line legible and forecastable — the discipline behind durable usage-based pricing.

3. Independence preserved the public price sheet

Walking away from the Snowflake acquisition (and later taking Snowflake as an investor) kept Reka a standalone, developer-first provider with raw public token rates rather than a consumption line inside a data cloud. That independence is itself a pricing decision: it protects the transparent, self-serve API that seeds developer adoption, echoing the shift away from opaque, sales-only AI pricing.

Areas to improve : Gaps in Reka AI’s pricing approach

1. Surface the multimodal total, not just the token rate

The headline per-token rate hides that images, video and audio are billed separately — and that Vision can meter the same video twice. A combined “estimated cost per multimodal request” view would prevent the bill-shock and unpredictability that usage pricing is meant to remove, especially for video-reasoning apps.

2. Publish an enterprise / on-prem starting point

On-prem and Enterprise are fully sales-gated with no public anchor. A published starting price or a worked deployment example would shorten evaluation for mid-market buyers who need data residency but can’t justify a sales call — the same transparency gap many peers are now closing.

3. Clarify the Spark / Edge model ladder

With Reka Spark ($0.05/$0.05) added beneath Reka Edge ($0.10/$0.10), the cheap end of the ladder now has two overlapping small models. Clearer guidance on when to pick Spark vs Edge — and whether Edge 2 supersedes the original — would make the entry tier easier to reason about and reduce accidental over-spend on the pricier option.

Key takeaways

1. Match the meter to the modality. Reka bills text per token, images per image, video and audio per minute — a price sheet that mirrors real multimodal cost drivers instead of flattening everything into tokens. The unit should follow the value, not convention.
2. One lab can run several meters. Per-token chat, per-request research, and per-video-minute vision coexist because each maps to how buyers think about that product. Don’t force a second product through the first product’s meter.
3. Independence is a pricing posture. Declining the Snowflake acquisition kept Reka’s public, self-serve token API alive instead of converting it into consumption credits inside a platform. Who owns you shapes how you’re allowed to price.
4. Multimodal surcharges are the hidden line. The low headline token rate understates cost because image/video/audio stack on top — and video can be metered twice. Transparency about the assembled total is the missing piece.
5. On-prem weights monetize sovereignty. Shipping deployable weights lets Reka sell data residency as a quoted enterprise upgrade — a lever closed-weight rivals structurally can’t pull.

UBP implications

1. Multimodal AI needs multimodal meters. Reka shows that as models ingest more modalities, a single token meter stops reflecting cost. UBP practitioners building multimodal products should expect to bill several units (tokens, images, minutes) and make each visible rather than blending them.
2. The right unit is the unit the buyer thinks in. Pricing research per request and video per minute — not per token — keeps each product legible and forecastable. The lesson is to choose the value metric per product surface, even within one company.
3. Outcome-shaped units are emerging at the agent layer. Reka Research’s per-request meter, priced by reasoning depth ($25–$60/1k), is an early move from raw inference toward outcome-shaped pricing — billing for a completed task rather than the tokens it consumed.

Sources

• Reka API pricing — per-token & multimodal rates (accessed 2026-06-11)
• Reka Vision pricing (accessed 2026-06-11)
• Reka AI marketing home (accessed 2026-06-11)
• Bloomberg — Snowflake talks to acquire Reka AI fizzle with no deal (accessed 2026-06-11)
• SiliconANGLE — Reka AI raises $110M at $1B valuation (accessed 2026-06-11)
• Browse the pricing blueprint corpus

Bottom line

Reka AI prices a natively multimodal stack through three public meters: a per-1M-token chat API (Spark $0.05/$0.05 to Core $2/$6, plus per-image/video/audio units), Reka Research at $25–$60 per 1,000 requests, and Reka Vision at $0.05 per indexed video minute. Enterprise on-prem weight deployment is sales-quoted. The defining move is matching the meter to the modality — and staying independent after Snowflake’s $1B-plus acquisition talks collapsed, which preserved the transparent, developer-first price sheet that closed-weight rivals don’t expose.

Want to compare Reka AI against other foundation-model labs? See Mistral AI and OpenAI, or browse the full pricing blueprint.

Pricing timeline : Major events on a vertical axis

Each milestone below corresponds to a public pricing change, product launch, or material adjustment. Major events use a filled marker; minor adjustments use a faded one.

Live snapshot: per-token API + Research + Vision, Enterprise by quote

Jul 2026

Captured live USD pricing from docs.reka.ai: Reka Chat per-1M-token Edge $0.10/$0.10, Flash $0.80/$2, Core $2/$6 plus per-image/video/audio units; Reka Research $25/$35/$60 per 1k requests; Reka Vision $0.05/input-min with 180 free minutes, plus an image meter ($10/1M uploads, $50/1M/month stored, $0.003/search) and a credit-based Developer tier; Vision Enterprise sales-quoted. Reka's marketing home now positions the company around 'the physical AI era' with four lines — Labs, Infer, Vision and Claru.

Edge 2 refresh on the compact line

Jan 2026

Reka refreshes its compact on-device line (Reka Edge 2) while the published API card continues to price Edge at $0.10 in / $0.10 out per 1M tokens, with Flash ($0.80/$2) and Core ($2/$6) anchoring the premium end.

Reka Vision adds per-video-minute pricing

Nov 2025

Reka Vision exposes video intelligence with a Free / Developer / Enterprise ladder: 180 free indexed minutes, then $0.05 per input video minute to index, $0.005/search, $2/1M QA output tokens and $0.04/min clip generation — a third, video-native meter.

Reka Research launches with per-request pricing

Sep 2025

Reka adds Reka Research (reka-flash-research) for complex multi-step web research, metered per 1,000 requests — $25 standard, $35 parallel-thinking low, $60 parallel-thinking high — a per-query meter distinct from the per-token chat API.

$110M round triples valuation past $1B; Nvidia & Snowflake invest

Jul 2025

Reka raises 110M USD, lifting its valuation above 1B USD (new unicorn), with Nvidia and Snowflake among backers — the same Snowflake that walked from the acquisition now joining as a strategic investor. Funds the independent multimodal roadmap and enterprise/on-prem push. (Bloomberg, SiliconANGLE, 2025-07-22.)

Snowflake's reported $1B+ acquisition talks collapse

May 2024

Bloomberg reports Snowflake's talks to acquire Reka for more than 1B USD broke down with no deal (2024-05-22). Reka continues independently rather than being absorbed — a lifecycle inflection that kept its public per-token API and on-prem deployment model intact instead of folding into a data-cloud platform.

Reka Flash, Edge and Core announced

Feb 2024

Reka unveils its model family — Reka Flash and the compact Reka Edge, with the larger Reka Core following — exposing a per-token API where multimodal inputs (image, video, audio) carry their own per-unit charges alongside text tokens.

Reka founded; first multimodal models in development

Jul 2023

Ex-DeepMind, Google Brain and Meta FAIR researchers (Dani Yogatama, Yi Tay, Cyprien de Masson d'Autume, Qi Liu, Mikel Artetxe) found Reka to build natively multimodal models spanning text, image, video and audio — establishing the multimodal-token meter as the eventual pricing primitive.