The cost center of an AI media product is moving. Reasoning models can now plan a campaign, select tools, inspect metadata, and coordinate jobs for fractions of a cent in many short workflows. The image, video, or audio output that the agent requests can cost far more—especially after failed generations, retries, storage, and review.

This is not because reasoning is free. Long contexts, high effort, and multi-agent runs can still be expensive. The change is that teams need separate ledgers for intelligence and media. One is usually priced in input and output tokens; the other may be priced per image, second of video, minute of audio, resolution, model, or transformation.

The right question is no longer “What does the model cost?” It is “What does one accepted, delivered asset cost under realistic failure rates?”

Reasoning prices now span capability tiers

OpenAI made GPT-5.6 generally available on July 9, 2026 in three API tiers. Sol is priced at $5 per million input tokens and $30 per million output tokens. Terra is $2.50 and $15. Luna is $1 and $6.

That range allows an orchestration system to route routine work to Luna or Terra and reserve Sol for difficult planning or review. A million input tokens is a large amount of text for a single media job, so a lightweight agent decision can be cheap compared with even one premium video render.

OpenAI also introduced Programmatic Tool Calling, which lets the model write and run lightweight programs that filter tool results and coordinate steps without sending every intermediate item back through the model. Fewer round trips and less repeated context can lower token cost.

Higher effort changes the equation. GPT-5.6 max spends more time exploring, and ultra coordinates four agents by default. Parallel reasoning may improve a difficult result or reduce wall-clock time, but it consumes more tokens. Use it when a better plan prevents expensive downstream mistakes, not as the default for a trivial file conversion.

One image can outweigh the planning tokens

Google lists Nano Banana 2 Lite at $0.034 per 1K image and reports approximately four-second text-to-image output. A marketing agent can reason about five concepts for little token cost, then spend $0.17 generating one image for each.

That remains inexpensive until volume and acceptance rate enter the model. Ten audiences, six placements, five concepts, and four candidates produce 1,200 images—a nominal $40.80 before review. If only 60 are usable, the generation cost is $0.68 per accepted image before editing, storage, and labor.

Lite supports 1K output and prioritizes speed. A final campaign may route selected concepts to a more capable or higher-resolution model, adding another generation stage. The cheapest draft can still be economical if it eliminates bad directions before premium rendering.

Measure the conversion funnel: generated, technically valid, brand-safe, legally approved, selected, and published. The unit price at the top of the funnel is not the production cost at the bottom.

Video exposes duration and retry costs

Gemini Omni Flash is available to developers in public preview. Google prices it at $0.10 per second of output video and currently limits generations to ten seconds. One full ten-second attempt therefore costs about $1 at the announced rate.

Five storyboard variants with three attempts each become $15. A 60-second campaign assembled from six shots can easily include dozens of rejected seconds. Changes to character consistency, text, product shape, or camera motion may require full regeneration rather than a local edit.

Video also creates larger files, longer queues, and more human review. A reviewer must watch time, not glance at a thumbnail. HDR, upscaling, frame interpolation, audio, captions, and delivery encodes add more meters.

Preview status matters. Omni Flash’s API behavior, limits, and price can change. Google states that audio-reference uploads and scene extension are not supported at launch, short video references have a processing limitation, and character consistency can weaken across scene changes or pans. Budget for the current product, not the roadmap.

Real-time voice has a different denominator

xAI announced Voice Agent Builder in beta on July 1. It lists Grok Voice agent usage at $0.05 per audio minute, with voices included, and an additional $0.01 per minute for telephony on a free provisioned number.

At 100,000 call minutes, those two meters total $6,000 before external phone numbers, transferred calls, tool APIs, storage, analytics, compliance, or human escalation. A voice agent that issues refunds or schedules appointments can also create business costs far larger than inference.

Call duration is only one quality measure. Track successful resolution, handoff rate, repeat calls, incorrect actions, abandonment, and complaints. A longer call may be worthwhile if it resolves a complex issue; a short call that forces the customer to call again is expensive.

The builder records and transcribes calls, which introduces retention, privacy, and review requirements. Compliance work belongs in the economic model even when it does not appear on the API invoice.

Media prices are not directly comparable

$0.034 per image, $0.10 per video second, and $0.05 per audio minute describe different deliverables. A still may be a draft, a video second may include generated motion, and a voice minute may include an interactive reasoning path.

Normalize around the business output. For advertising, use an approved placement or tested creative. For localization, use an approved target minute. For support, use a resolved case. For product design, use an asset that passes the specification.

Then decompose the cost:

This structure makes vendor comparisons meaningful and prevents a low headline rate from hiding a poor usable-output rate.

Caching and context have their own economics

Repeated brand instructions, schemas, and tool definitions can dominate agent input. GPT-5.6 offers prompt caching with explicit cache breakpoints; cache writes are billed above uncached input while cache reads receive a 90 percent discount. A stable, reused prefix can therefore save money, but frequently changing content may not justify the write.

Media references are different. Uploading the same large image or video to several services can create transfer, storage, and privacy costs even if text context is cached. Use controlled object storage, signed URLs with limited life, and a reference manifest rather than copying binary data through every agent turn.

Do not cache secrets, customer media, or instructions beyond their authorized life. Financial efficiency cannot override retention rules.

Failed jobs and retries are a hidden multiplier

Media APIs often run asynchronously. A timeout from the client does not prove that the provider canceled the generation. Blindly resubmitting can create two paid outputs.

Persist a local job record before submission. Store the request hash, provider, model, cost estimate, external task ID, status, attempt number, and output location. Poll or receive a webhook until the first task reaches a terminal state. Retry only the operation that is known to have failed.

Separate creative retries from technical retries. “The product looks wrong” is a new creative attempt and should count against the variation budget. A transient download failure should not trigger a new generation.

Track duplicate prevention as savings. It is one of the easiest ways for orchestration engineering to pay for itself.

Use reasoning to avoid expensive generation

Cheap intelligence is most valuable when it reduces media waste. An agent can validate a prompt against the product facts, verify that reference files exist, calculate aspect ratios, estimate cost, produce a storyboard, and ask for approval before rendering.

It can choose a lower-cost model for drafts, detect that an exact edit is better than regeneration, or stop a batch after enough diverse candidates exist. It can inspect metadata and run automated checks before a human watches the clip.

Do not ask the agent to generate until something looks good. Give it a hypothesis, acceptance criteria, candidate limit, and budget. Make cost a tool parameter, not an afterthought.

Separating Medux processing from orchestration

Medux provides external media-processing tasks that Codex can call after independent MCP configuration. It is not bundled with OpenAI, Google, or xAI, and its processing credits are not the same as agent tokens.

The Codex video-editing tutorial demonstrates an asynchronous video task, while the Codex TTS tutorial covers speech generation. A production ledger should record the model tokens used to plan and call a task, the Medux credits consumed by the task, storage and transfer, polling, retries, and human approval separately.

Use one local job ID to connect these records without pretending they are one provider invoice. Preserve the returned Medux task ID, poll it instead of resubmitting, and verify the result before the next paid stage. For TTS, count rejected pronunciations and script changes; for video, count merge or transcode revisions.

The same rule applies to privacy. A file generated locally or by another API becomes an external upload when sent to Medux. Check current retention and authorization; do not describe a hybrid workflow as zero-data-retention unless every component contract supports that claim.

Cheap reasoning does not make AI media cheap by itself. It gives teams a new lever: spend a small amount on planning, validation, and routing to avoid spending a larger amount on pixels and audio nobody can use.