GPT-5.6 Sol vs Terra vs Luna: The Real Cost of an AI Video Project

A token price table cannot tell you what an AI video costs. It can tell you the price of a model response. The project cost depends on how often the response must be revised, whether it prevents failed renders, how much context is repeatedly sent, and which tasks actually require advanced reasoning.

As of August 28, 2026, OpenAI lists GPT-5.6 Sol at $4 per million input tokens and $20 per million output tokens, Terra at $2 and $12, and Luna at $0.20 and $1.20. Cached input costs $0.40, $0.20, and $0.02 respectively. Sol's current price is promotional through at least November 21, according to OpenAI's rate card.

This article models a realistic one-minute anime-video preproduction workflow. GPT-5.6 is used for text and image analysis, planning, prompts, and review—not direct video rendering.

The Project We Are Costing

Assume a solo creator is producing a 60-second anime short containing eight shots. The workflow includes:

premise development;
a 900-word script;
a character specification;
an eight-shot storyboard brief;
eight generation prompts plus revisions;
voice and sound direction;
continuity review using selected frames;
titles, description, subtitles, and platform metadata.

Video, image, voice, and editing charges vary by platform and are excluded from the token calculations. That exclusion is deliberate: we want to understand the planning layer, then examine how it affects the expensive rendering layer.

You can execute this full pipeline in Elser AI, where scripts, characters, storyboards, scenes, audio, and final editing remain connected. The calculations below help decide which language model should support each decision.

Current Price Inputs

| Model | Uncached input / 1M | Cached input / 1M | Output / 1M | | Sol | $4.00 | $0.40 | $20.00 | | Terra | $2.00 | $0.20 | $12.00 | | Luna | $0.20 | $0.02 | $1.20 |

The basic formula is:

cost = (uncached input tokens × input rate / 1,000,000) + (cached input tokens × cached rate / 1,000,000) + (output tokens × output rate / 1,000,000)

Always use actual API usage data for billing. Word-to-token estimates are planning tools, not invoices.

Scenario A: Use One Model for Everything

Assume the project consumes 160,000 input tokens and 55,000 output tokens across research notes, iterations, and reviews, with no caching.

Sol

Input: 0.16 × $4.00 = $0.64 Output: 0.055 × $20.00 = $1.10 Total: $1.74

Terra

Input: 0.16 × $2.00 = $0.32 Output: 0.055 × $12.00 = $0.66 Total: $0.98

Luna

Input: 0.16 × $0.20 = $0.032 Output: 0.055 × $1.20 = $0.066 Total: $0.098

Luna is about one-tenth the cost of Terra in this simplified case. Yet the absolute difference between Terra and Sol is only $0.76. If Sol prevents one unnecessary video rerender, it may be cheaper at the project level. Conversely, using Sol to generate file names saves almost nothing and consumes premium capacity where it adds little value.

Scenario B: Route Work by Difficulty

A more realistic production routes tasks:

Luna: 80,000 Input, 25,000 Output

Use Luna for idea expansion, metadata, formatting, subtitle cleanup, shot-log normalization, and prompt-template population.

0.08 × $0.20 + 0.025 × $1.20 = $0.046

Terra: 65,000 Input, 24,000 Output

Use Terra for script development, shot planning, dialogue, scene prompts, and revision summaries.

0.065 × $2.00 + 0.024 × $12.00 = $0.418

Sol: 25,000 Input, 6,000 Output

Use Sol for final narrative review, contradiction detection, and diagnosing the most expensive failed scene.

0.025 × $4.00 + 0.006 × $20.00 = $0.22

Routed Total

The routed workflow costs approximately $0.684 in text-model tokens. It is cheaper than Terra-only in our example and places stronger reasoning at the two moments where it can prevent expensive mistakes.

This is not a universal benchmark. It demonstrates a method: estimate task volumes, measure actual usage, and pay for capability where it changes acceptance rates.

Scenario C: Add Prompt Caching

Suppose the project repeatedly sends a 12,000-token stable production bible containing the character sheet, style guide, safety constraints, terminology, and output schema. It appears in 12 requests.

Without caching, that block creates 144,000 input tokens. With one uncached write and 11 cached reads, Terra's approximate input cost is:

First 12,000 tokens uncached: 0.012 × $2.00 = $0.024 Next 132,000 cached: 0.132 × $0.20 = $0.0264 Total stable-context input: $0.0504

Without caching, the same context would cost:

0.144 × $2.00 = $0.288

That is an 82.5% reduction for this context block. It is not the full-project savings because variable prompts and output remain normally priced.

To improve cache reuse:

place stable content first;
keep its wording and order unchanged;
separate variable scene instructions after a breakpoint;
avoid injecting timestamps or request IDs into the stable prefix;
measure cache hits rather than assuming they occur.

The Cost That Matters: Accepted Output

Run the same cost ledger against a real short in Elser AI. Register, produce one representative scene, and measure approved deliverables—not raw generations.

Imagine two prompt-generation approaches:

Luna produces a usable shot prompt after four revisions.
Terra produces a usable shot prompt after two revisions.

Luna may still be cheaper in token cost, but the extra human review time can dominate. Now suppose one weak prompt causes two failed video renders at $1 each. The text-token savings disappear.

Track cost per accepted shot, including:

text model cost

image/reference generation
video generations and extensions
voice and audio
upscale/export
human review time
cost of discarded outputs

The correct model is the one that minimizes this total while meeting quality and turnaround requirements.

A Stage-by-Stage Recommendation

Premise Exploration: Luna

Ask for 30 constrained premises in a table. Human selection is cheap, so generating many options with Luna makes sense.

Script and Character Logic: Terra

Terra is the default for drafting and revision. It offers more room for nuanced dialogue and continuity without Sol's full price.

Storyboard and Prompt Formatting: Terra, Then Luna

Use Terra to define shot intent and camera logic. Once the pattern is approved, Luna can populate repeated structured fields.

High-Risk Scene Diagnosis: Sol

Send the prompt, reference frames, failed outputs, and a concise test history. Ask Sol to rank likely causes and propose one-variable experiments.

Metadata and Subtitle Cleanup: Luna

These are high-volume, testable tasks. Use strict schemas and automated checks.

Final Continuity Audit: Sol or Terra

Use Sol for a complex narrative with many references. Terra is enough for a simple short with clear rules.

After approval, move the assets into Elser AI to generate and assemble scenes, voice, music, and effects while keeping revisions connected to the original plan.

Do Not Pay for Unnecessary Output

Output tokens cost five to six times more than input tokens across the three models. Ask for the smallest artifact that supports the next decision.

Instead of:

Write a comprehensive analysis of every possible issue with this shot.

Use:

Return the five most likely failure causes, evidence for each, and one controlled test. Maximum 500 words.

Request tables, JSON, or fixed fields when a human or tool will consume the answer. Long motivational introductions do not improve production quality.

Budgeting Image Inputs and Tools

The simple calculations above cover text tokens only. Image inputs are tokenized, and tool calls may carry separate charges. Fast processing can also have different pricing. Use OpenAI's usage response and billing dashboard to record:

uncached input;
cached input;
output and reasoning tokens;
image input;
web search or other tool use;
processing tier.

Do not multiply a text-only estimate across a vision-heavy workflow.

Build a Production Cost Ledger

When your budget assumptions are ready, start creating with Elser AI and validate the complete path from script through final scene before committing monthly spend.

Create one row per accepted deliverable:

| Deliverable | Model | Attempts | Input | Cached | Output | Tool cost | Render cost | Review minutes | Accepted? | |---|---|---:|---:|---:|---:|---:|---:|---:|---|

After five projects, patterns become visible. You may learn that Sol reduces revisions for scripts but adds no value to caption writing, or that a better reference sheet saves more money than any model change.

If Elser is your production environment, include Elser generations and final-cut revisions in the same ledger. The goal is to optimize the complete pipeline, not one API line item.

FAQ

How much does GPT-5.6 cost for one AI video?

There is no fixed amount. In our illustrative one-minute workflow, routed GPT-5.6 text usage cost less than $1, excluding image, video, audio, tools, and human work. Your real usage may differ substantially.

Is Luna always the cheapest choice?

It has the lowest token rates. It may not have the lowest cost per accepted result when a task is ambiguous or high risk.

When should I use Sol?

Use it for complex planning, cross-document reasoning, difficult failure diagnosis, or final review where one better decision can prevent expensive production mistakes.

What is the best default for creators?

Terra is the balanced default. Route repetitive structured tasks to Luna and escalate the hardest decisions to Sol.

Does caching reduce output cost?

No. It reduces eligible repeated input cost. Keep outputs concise and structured to control output spending.

Conclusion

The cheapest AI-video workflow is not the one with the cheapest language model. It is the one that reaches an accepted final video with the fewest wasted cycles. Route repetitive work to Luna, development to Terra, and expensive decisions to Sol. Cache stable context, constrain output, and measure the total cost per accepted shot. That turns pricing from a headline into an actual production advantage.