OpenAI: GPT-6 Luna (batch)

Review this 40-page technical spec for our new API gateway. Flag any security gaps, inconsistent naming conventions, and sections that contradict each other.

In a typical response, the model would produce a structured analysis across three sections: (1) Security — identifying missing rate-limit definitions in section 7.2 and an unvalidated redirect parameter in the OAuth flow; (2) Naming — noting that endpoint paths switch between camelCase and snake_case after page 18; (3) Contradictions — pointing out that the retry policy in section 3 allows five attempts while the error-handling appendix caps at three. The output would reference specific page numbers and quote conflicting passages side-by-side.

The 1.05M token context window lets GPT-6 Luna ingest entire technical documents without chunking, preserving cross-references that smaller models miss. Batch mode cuts costs to $0.05/Mtok input — viable for bulk document review. Trade-off: batch processing means no real-time interaction; you submit the job and retrieve results later.

Here's a CSV of 80,000 customer support tickets from Q4. Cluster them by root cause, then draft a one-page executive summary with the top five issues and recommended fixes.

A representative output would open with a summary table: authentication failures (18,200 tickets), payment gateway timeouts (14,600), mobile app crashes on iOS 17 (9,800), missing invoice emails (7,400), and unclear onboarding flow (6,100). Each issue would get a two-sentence explanation of the pattern and a concrete fix — e.g. 'Auth failures spike between 2–4am UTC when session tokens expire during overnight batch jobs. Extend token TTL to 8 hours or migrate batch processes to a service account.' The model would cite ticket IDs as evidence.

Batch pricing makes large-scale analysis economical — processing 80k tickets at $0.05/Mtok is cheaper than hiring analysts to read them. The model's file-handling modality means you upload the CSV directly rather than pasting excerpts. Limitation: clustering quality depends on ticket structure; inconsistent tagging in your data will produce noisier groupings.

I'm attaching a screenshot of a competitor's pricing page and our internal cost breakdown spreadsheet. Propose three pricing tiers that undercut them on entry-level plans but capture more revenue from enterprise customers.

In this scenario, the model would analyse both the screenshot (extracting the competitor's $29/$99/$299 structure) and the spreadsheet (identifying your $12 per-user cost and 40% target margin). It would propose: Starter at $19/month (5 users, limited integrations), Professional at $79/month (25 users, API access), and Enterprise at custom pricing starting at $500/month (unlimited users, dedicated support, SLA). Each tier would include a margin calculation and a one-line positioning statement — e.g. 'Starter undercuts their entry tier by 34%, converting price-sensitive solopreneurs.'

Image input lets the model parse visual layouts (pricing tables, charts) without manual transcription, saving setup time. The 1.05M token budget means you can include the full spreadsheet with historical data, not just a summary. Caveat: the model won't know your market's willingness-to-pay; you still need to validate pricing assumptions with customer research.

When GPT-6 Luna batch wins on legal review at scale

A 12-person legal ops team needs to extract clauses from 200+ vendor agreements every quarter and flag non-standard terms before renewal. GPT-6 Luna batch is the right call here because the 1.05M token context window fits most multi-document bundles without chunking, and the $0.05/Mtok input price makes bulk processing economical even at high volume. The batch API trades latency for cost—jobs complete in hours, not seconds—but that's fine when you're running overnight extractions. If you need same-day turnaround or interactive Q&A during review, switch to the real-time endpoint. For scheduled, high-volume document work where you can wait 6-12 hours, this model delivers the best unit economics in the GPT-6 family.

How batch processing cuts research costs by 70% for analysts

A 4-person market research consultancy ingests 50-80 industry reports per client engagement and needs executive summaries by morning. GPT-6 Luna batch handles this because the team can queue all PDFs at 5pm, pay $0.05/Mtok on input (versus $0.30+ on real-time endpoints), and retrieve structured summaries by 8am. The 1M+ context window means each report—even 200-page decks with appendices—processes in a single call without pagination logic. Output cost is higher at $0.25/Mtok, but summaries are 5-10% of input length, so total spend stays under $2 per engagement. If clients demand same-hour delivery or you're summarizing fewer than 10 documents per job, the batch discount doesn't justify the wait. For regular, high-volume overnight work, this is the most cost-effective GPT-6 option.

When to use GPT-6 Luna batch for daily moderation queues

A 20-person community platform reviews 5,000+ user-submitted posts daily for policy violations and needs results before the next business day. GPT-6 Luna batch works when moderation isn't real-time: the team exports flagged content at midnight, runs batch classification overnight, and surfaces results to human reviewers by 9am. At $0.05/Mtok input, processing 5K posts (average 300 tokens each) costs under $0.10 per run—70% cheaper than real-time endpoints. The image modality also handles mixed-media posts without separate pipelines. The threshold: if you need sub-minute moderation (live chat, high-risk verticals), batch latency breaks your workflow. For daily review queues where 6-hour SLA is acceptable, this model cuts moderation costs to near-zero while maintaining GPT-6 accuracy.