Reka Flash
Language Available Full comparison ↗ reka/reka-flash · by Reka AI
· decoder-only-transformer
Pricing — 1 offering(s)
Input tokens
- $0.80 / 1M tokens (input) Current 2026-08-04 → present
Output tokens
- $2.00 / 1M tokens (output)
Current
2026-08-04 → present
Same sourcing basis as the input tier.
Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.
See how Reka Flash fits into a cost-aware routing setup
See how →Capability profile
How Reka Flash rates across core capability dimensions, with the task-level evidence behind each rating.
The 2024 report covered many languages; the 2026 update narrows speech I/O to English but text remains multilingual.
No benchmark published for the current API version. The 2024 report's figures (MMLU 75.9, GSM8K 85.8) describe an earlier model this version has diverged from, so they are not carried as ratings here.
2024 report: HumanEval 72.0 — same divergence caveat; not treated as current.
Function calling and structured output added in the 2026-08-01 update (Reka's own post); no agentic-benchmark score.
8,000-token default, extendable to ~128K per Reka's docs — the default is short and long context is opt-in.
Reka's stated positioning: the 'turbo-class' / fast tier of the Reka lineup, ~21B dense.
$0.80 / 1M input, $2.00 / 1M output (Reka API docs) — mid-priced; the value case is the native multimodal (incl. audio/video) input at that price.
Ratings are estimated — limited independent data is available for this model.
Operator guidance
Consider Reka Flash specifically when native audio + video + image input in one model matters and budget is mid-tier; few models in this registry take audio and multi-minute video input directly. For text-only quality at this price, general flagships and cheap MoE models are stronger and better benchmarked. Reka Edge (cheaper) and Reka Core (stronger) are the other tiers — only Flash and Edge are named by Reka as standing public baselines.
Use cases
- Multimodal input tasks spanning text, image, video (3-5 min), and audio in a single model at a mid-tier price
- OCR / document-understanding workloads (a stated focus of the 2026-08-01 update)
- Applications wanting one 'always-available public baseline' Reka model rather than the pricier Reka Core
Limitations
- The only published benchmarks (arXiv:2404.12387) describe the 2024 model, which the current API version has diverged from — capability ratings here are largely `unknown` as a result
- 8K default context is short; long context is opt-in and may carry different cost/latency
- Speech I/O narrowed to English in the 2026 update
- No first-party pricing-change history published — effective_from on the registry entry is the verification date