Platform decision · 16 July 2026

Two endpoints. One clean stack.

One endpoint generates every image and video. One endpoint reasons, orchestrates, and drives the agent fleet. The picks below optimize for Piexels, not benchmark theatre.

Endpoint 1 · Generation

Luma Agents API

UNI-1 images and Ray3.2 video through one REST surface, with the strongest keyframe control and the only convincing agent board in this field.

Runner-up: Runway API for model breadth and mature production recipes.

Endpoint 2 · Reasoning

Kimi K2.7 HighSpeed

Keep the existing flat-rate, Anthropic-compatible worker lane. It is the best fleet endpoint even though Opus 4.8 remains the raw quality winner.

Runner-up: Claude Opus 4.8 for premium judgment and high-stakes escalation.

Planning basis: 120 five-second 720p clips, 500 2K images, and the existing Kimi subscription. Public price floor: $120.45 per month. Practical budget with a 25% media reroll reserve: about $141 per month.

Generation endpoint

Quality matters. Control, API maturity, agent workflow, and the ability to keep a property or product coherent across motion matter more.

Platform	Quality	Unified suite and agents	Motion control	API maturity	Public cost	Fit
Luma Agents APIPick	Ray3.2 production video. UNI-1 leads Luma's cited human-preference categories for overall, editing, and references.	One async REST envelope for images, edits, video, video edits, and reframing. The app adds a shared agent board.	Start frame, end frame, up to 16 keyframes, camera and motion transfer, product swap, and reframe.	Python, TypeScript, Go, CLI, Files API, request IDs, limit headers, pay-as-you-go, and provisioned throughput.	$0.30 per 5s 720p clip $0.0909 per 2K image	Best direct fit for property, product, and reusable brand systems.
Runway APIRunner-up	Gen-4.5 plus Seedance 2, Veo 3.1, Gemini Omni, Seedream 5, and GPT Image 2.	Broadest catalog and production recipes. No persistent Luma-style creative agent board.	First and last frames through Seedance and Veo. Strong product ad, swap, UGC, and multi-shot recipes.	No request-per-minute cap inside daily allowance. Server-side concurrency queue and spend controls.	Gen-4.5 5s: $0.60 Gen-4 Turbo 5s: $0.25 Images: $0.02 to $0.08	Best escape hatch when model breadth wins.
Google Vertex AI	Veo 3.1 and Imagen 4 are top-tier, with native audio on Veo.	One cloud platform, but no media-specific collaborative agent board.	Reference images and supported first and last frame variants. Veo 2 exposes explicit interpolation and camera controls.	Enterprise-grade. Earlier Veo 3 docs list 10 requests per minute per project. Provisioned throughput exists.	Veo 3.1 5s video-only: $1.00 Imagen 4: $0.04	Final-quality specialist, not the cleanest creative endpoint.
Seedance 2.0	Second on the Artificial Analysis text-to-video snapshot and first on Arena's image-to-video snapshot.	Unified text, image, audio, and video inputs inside one model. Global platform access is strongest through partners.	First, last, and reference frames through supported partners, with broad aspect ratios.	Official model and client exist. Partner endpoints remain the practical global path.	Runway 720p 5s: $1.80	Top model, not yet the best primary platform endpoint.
Kling 3.0 Omni	Strong realism, multi-shot direction, audio, and subject consistency.	Video and image share Kling's platform. Agent workflow and billing are less transparent.	Multi-image video, frame control, and multi-shot storyboarding.	Official Open Platform is live, with improving docs and complex mode-based billing.	Some multi-image modes start near $0.056/s, then rise by quality and audio.	Strong specialist, weaker endpoint simplicity.
Higgsfield	Good model access, presets, camera moves, and campaign tooling.	Canvas, Marketing Studio, Supercomputer, MCP, and CLI make it the strongest app-level Luma rival.	Excellent camera vocabulary. Guarantees vary by underlying model.	Public API pricing and service levels were not available in the accessible pricing page.	Not publicly verifiable	Keep for experiments. Do not standardize.
OpenAI Sora 2 API	Good physics, synced audio, editing, characters, and remix.	Video and image APIs share a vendor, without a unified media board.	Reference image, edit, extend, remix, and character tools. No comparable multi-keyframe surface.	Tier 1 starts at 25 requests per minute. Consumer Sora closed, API remains available.	Sora 2 5s: $0.50 Pro 5s: $1.50	General fallback.
MiniMax Hailuo 2.3	Competitive low-cost text and image-to-video output.	Text, image, video, audio, and agent products exist, but the creative workflow is less cohesive.	Useful image-to-video and camera control. Keyframe documentation is weaker.	Official API and prepaid packages. Package tiers list 20 to 50 requests per minute.	768p 6s: $0.19 Fast or $0.28 standard	Budget specialist.
Pika 2.5	Strong effects, swaps, additions, scenes, and frames. Less consistent for premium property motion.	Consumer edit suite, not an agentic image and video production system.	Pikaframes supports longer frame-driven sequences.	API link exists. Consumer pricing is clearer than production API detail.	Standard annual plan: $28/month	Effects specialist.

Why Luma wins

Luma is the only candidate where one API surface, the right property and product control system, and an agent-native shared workspace meet cleanly. Runway wins breadth. Google and Seedance can win individual quality battles. None matches Luma's complete fit.

One generation envelope
Up to 16 keyframes
Shared agent board
$0.30 exploration clips

Reasoning endpoint

The quality winner is Opus. The fleet winner is Kimi. This decision is about total throughput, integration, and cost across every worker.

Platform	Benchmark evidence	Cost	Speed and context	Tools and compatibility	Reliability	Decision
Kimi K2.7 Code HighSpeedPick	Vendor: Kimi Code Bench V2 62.0, MCP Mark Verified 81.1, and 30% fewer reasoning tokens. No comparable independent SWE or Terminal submission yet.	Subscription: $39 to $199/month API: $1.90/M in, $8/M out	About 5 to 6 times standard output speed. 262K public model context. HighSpeed uses about 3 times quota.	OpenAI and Anthropic-compatible. Official Claude Code support. Existing Piexels route is live.	Weekly, five-hour, and monthly quotas. Metered Extra Usage can bypass quota limits.	Best whole-fleet economics and integration.
Claude Opus 4.8Runner-up	Raw quality winner. Only model to complete every case in an external Super-Agent benchmark. Z.ai comparison: Terminal-Bench 2.1 at 85.0.	$5/M in, $25/M out Fast: $10/M in, $50/M out	1M-class context, effort controls, compaction, and 2.5-times fast mode.	Native Claude Code, MCP, computer use, and dynamic workflows.	Best documented judgment, self-checking, and end-to-end completion.	Quality escalation, not the fleet default.
Z.ai GLM-5.2	Vendor: Terminal-Bench 2.1 at 81.0 and SWE-bench Pro at 62.1. Strongest open-weight long-horizon contender.	Reported public rate: $1.40/M in, $4.40/M out. Recheck before purchase because the accessible official pricing index lagged the launch.	1M context, effort control, and strong long-context throughput.	Claude Code, ZCode, OpenCode, and OpenAI-compatible runtimes.	Very new. Vendor flags more reward-hacking tendency than GLM-5.1.	Best open-weight challenger.
OpenAI GPT-5.5 and Codex	GPT-5.3-Codex: Terminal-Bench 2.0 at 77.3, SWE-Bench Pro at 56.8, OSWorld-Verified at 64.7.	GPT-5.5: $5/M in, $30/M out Cached input: $0.50/M	1.05M context, 128K max output.	MCP, skills, hosted shell, patching, computer use, web, and tool search.	Mature API and an existing Piexels fallback subscription.	Strong fallback, not the default.
Google Gemini 3.1 Pro	Google: SWE-bench Verified 80.6 and Terminal-Bench 2.0 68.5. Independent Terminal-Bench: 70.3.	Under 200K: $2/M in, $12/M out Above 200K: $4/M in, $18/M out	1M input, 64K output, strong multimodal input.	Function calling, custom-tools endpoint, search, URL context, code execution, and remote MCP.	Still preview. Google notes quality fluctuation on the custom-tools variant outside tool-heavy work.	Best multimodal API alternative.

Evidence discount: Kimi K2.7 does not yet have the same independent benchmark coverage as Opus, Gemini, or established Codex models. The recommendation is based on Piexels deployment economics and proven compatibility, not a claim of benchmark supremacy.

Why Kimi wins the fleet

A single high-stakes reasoning call belongs on Opus. A whole worker fleet needs predictable spend, high throughput, tool compatibility, and enough intelligence to complete the loop. The existing Kimi subscription already provides that shape.

Already integrated
Flat-rate capacity
Anthropic-compatible
Fast output

Cost model

A transparent planning workload, not a fabricated current-volume claim.

Combined public floor

$120.45

120 Luma clips at $0.30, 500 UNI-1 images at $0.0909, and Kimi Allegretto at $39.

Practical monthly budget

≈ $141

Add a 25% reserve to media generation for rerolls. Replace the $39 Kimi line with the actual current tier if higher.

Reasoning route	20M input	4M output	Estimated monthly API cost
Kimi K2.7 HighSpeed	$38	$32	$70
Gemini 3.1 Pro, requests under 200K	$40	$48	$88
GLM-5.2, reported public rate	$28	$17.60	$45.60
Claude Opus 4.8	$100	$100	$200
OpenAI GPT-5.5	$100	$120	$220

Migration

Keep the working brain. Replace the fragmented media path.

Keep

Kimi HighSpeed as the default brain.
Anthropic-compatible worker contracts and MCP loops.
Codex as runtime fallback.
Approval gates, visual QA, and deterministic post.

Swap

Make Luma the canonical generation dependency.
Map images and edits to UNI-1.
Map video, edits, keyframes, and reframe to Ray3.2.
Upload campaign assets once and chain from IDs.

Two means two: Higgsfield can remain available during migration, but it leaves the production path. Runway stays documented as the runner-up, not a hidden third endpoint.

Sources

Official product, pricing, API, and model documentation, plus current independent benchmark surfaces.