The short answer
For writing from supplied material, start by testing GPT-6.1 Terra (gpt-6.1-terra): OpenAI describes it as a fast, efficient model for clear, natural writing. For complex, quality-sensitive work, compare it with GPT-6 Astra (gpt-6-astra), which OpenAI describes as its most capable model for demanding work. For focused, high-volume tasks where cost matters, consider GPT-6 Luna (gpt-6-luna). GPT-6.1 Sol (gpt-6.1-sol) is another candidate for complex work when balancing capability and cost matters. These are API model IDs, not ChatGPT subscription tiers. (developers.openai.com)
That is a task-based shortlist, not a proven ranking of writing quality: the official model guidance does not establish a universal best model for fiction, marketing voice, editing, or factual prose. Test the candidates on your own representative prompts. The catalog lists these models as available through the Responses API and OpenAI client SDKs. (developers.openai.com)
Which model should you try?
| Your writing task | Candidates to test | What the official guidance says |
|---|---|---|
| Drafting or revising clear prose from notes, a brief, or supplied source material | Terra; compare with Sol or Astra if needed | OpenAI describes Terra as suited to clear, natural writing. Its model-selection guidance should not be read as proof it will write best for every brief. (developers.openai.com) |
| Complex brief or demanding deliverable | Astra; compare with Sol | OpenAI calls Astra its most capable model for demanding work. It describes Sol as near-Astra performance for complex work at lower cost. Those are general capability positions, not a guarantee of better prose. (developers.openai.com) |
| Focused, repeated writing tasks at high volume | Luna; compare with Terra if the output misses your quality bar | OpenAI describes Luna as its most efficient model for focused, high-volume tasks. Test quality before routing a large share of work to it. (developers.openai.com) |
| Fiction, marketing voice, or highly original creative prose | Test Terra and Astra; add Sol or Luna if cost matters | The official catalog does not rank these models for fiction or originality. Judge voice, freshness, coherence, and editing effort using your own prompts. (developers.openai.com) |
| Factual articles, reports, or summaries | Choose by task fit, then verify important claims | A model recommendation is not evidence that factual claims are true. Check important statements against source material or other reliable references. |
Practical default: Try Terra first for everyday writing from supplied material. Compare it with Astra when a brief is unusually demanding or Terra does not meet your quality bar. Consider Luna for focused, high-volume work only after it passes your evaluation. These recommendations follow OpenAI’s stated model positioning; they are not the results of an independent writing benchmark. (developers.openai.com)
API prices and the quality trade-off
OpenAI lists the following standard API rates in USD per 1 million tokens. Prices below are for standard processing, not the separately listed Fast or Flex/Batch rates. (developers.openai.com)
| API model ID | Input per 1M tokens | Output per 1M tokens | Officially described positioning |
|---|---|---|---|
gpt-6-luna |
$0.10 | $0.50 | Focused, high-volume tasks |
gpt-6.1-terra |
$1.00 | $5.00 | Clear, natural writing |
gpt-6.1-sol |
$2.00 | $10.00 | Complex work at lower cost than Astra |
gpt-6-astra |
$10.00 | $50.00 | Most demanding work |
These are API token rates, not ChatGPT plan prices. Actual charges depend on input and output token use; other processing modes, caching, or tools may change the total. Check the pricing page for the applicable rate before budgeting. (developers.openai.com)
Worked example: If a request uses 1,000 input tokens and returns 1,000 output tokens, its standard text-token cost would be approximately:
- Luna: $0.0006
- Terra: $0.006
- Sol: $0.012
- Astra: $0.06
This is an arithmetic example, not a measured workload bill. At equal input and output token counts, Astra’s listed standard text-token rates make it ten times the cost of Terra. Whether the difference is worthwhile depends on whether it improves the result enough to reduce editing or meet a higher quality bar. (developers.openai.com)
A quick test with your own writing
Don’t pick a production model from a general ranking alone. Compare candidates on work you actually expect them to handle:
- Choose representative briefs. Use five to ten real or safely anonymized examples, such as a rewrite from notes, a short email, a product description, or a factual summary.
- Keep the instructions consistent. Give each candidate the same source material, audience, tone, length, and requested format. Keep relevant settings the same where applicable.
- Score the outputs. Assess instruction-following, voice, structure, useful detail, factual support where relevant, and length. Track how many minutes of human editing each output needs.
- Set a quality floor first. For example, require at least 4 out of 5 for factual support and instruction-following, then choose the least expensive candidate that meets your threshold. That score is your own decision rule, not an official benchmark.
- Check facts against sources. A fluent answer—or a second model’s approval—does not establish that a claim is true. Verify important statements against the supplied source material or other reliable references.
OpenAI’s model catalog gives model positioning, not a writing-quality leaderboard. Treat your comparison as evidence about your prompts and workload, not proof of a universal ranking.
What to confirm before using a model in production
- Use the API model ID: the current catalog lists IDs including
gpt-6.1-terra,gpt-6-astra,gpt-6.1-sol, andgpt-6-luna. Don’t assume an older ChatGPT label maps directly to an API ID. (developers.openai.com) - Confirm the interface: OpenAI lists the latest models as available through the Responses API and its client SDKs. Check the specific model documentation for the capabilities your application needs. (
developers.openai.com) - Estimate input and output usage separately: the prices above are per million tokens. Include any tools or processing modes your application uses when estimating cost. (
developers.openai.com) - Keep API and ChatGPT access distinct: this guide compares API model IDs and API usage. ChatGPT app labels, access, and plan limits are separate product details.
- Re-test when your workload changes: a comparison on your own prompts helps you check whether a model still meets your quality bar.
Bottom line
For writing from supplied material, put GPT-6.1 Terra (gpt-6.1-terra) on your shortlist first. Compare it with GPT-6 Astra (gpt-6-astra) for demanding work, and consider GPT-6 Luna (gpt-6-luna) for focused, high-volume tasks. GPT-6.1 Sol (gpt-6.1-sol) is positioned for complex work at lower cost than Astra. For fiction, voice-led prose, or factual writing, the official model descriptions do not settle which will work best for you: test representative prompts, track editing effort and cost, and verify important facts independently. (developers.openai.com)