OpenAI: GPT-5.6 Terra (batch)

Review this 40,000-line TypeScript monorepo for architectural debt. Flag patterns that will hurt us at 10x scale and suggest incremental refactors we can ship without a rewrite.

In this illustrative example, the model would produce a structured audit identifying three critical patterns: (1) a shared state manager coupling 18 modules that will bottleneck horizontal scaling, (2) synchronous file I/O in request handlers adding 200ms p99 latency under load, and (3) a circular dependency chain across the auth/billing/notifications boundary. For each, it would propose a phased migration path—introducing an event bus for state, moving I/O to background workers, and extracting a shared interface layer—with effort estimates and backward-compatibility notes for each phase.

The 1.05M token context window allows the model to hold an entire medium-sized codebase in working memory, enabling whole-system analysis that smaller models can't match. The batch pricing ($1 input) makes this economical for nightly CI runs. Trade-off: batch processing means no real-time interaction—you submit the job and wait for results.

Here's a 90-page RFP and our last three winning proposals. Draft a compliance matrix, executive summary, and technical approach section that mirrors our voice while addressing every requirement.

In this illustrative example, the model would generate a three-part response: a compliance matrix mapping all 47 RFP requirements to proposal sections with page references, a two-page executive summary matching the client's priorities (cost predictability, vendor lock-in concerns, timeline risk) in the firm's established tone, and a 12-page technical approach that adapts proven patterns from the reference proposals while addressing this client's AWS-native constraint and emphasizing the team's federal compliance experience.

Long context enables the model to cross-reference the full RFP against multiple past proposals simultaneously, catching requirement nuances that would be lost in chunked processing. The file modality means you can upload PDFs directly. Trade-off: at $6/Mtok output, a 15,000-word draft costs ~$0.40—cheap per document, but adds up across dozens of proposals.

Analyze these 600 customer support tickets from Q4. What are the top friction points, which ones indicate product gaps vs. documentation gaps, and what should we prioritize in the next sprint?

In this illustrative example, the model would surface a prioritized breakdown: (1) 140 tickets cluster around OAuth callback failures on Safari—a product bug affecting 23% of mobile users, (2) 95 tickets ask how to export data in CSV format—a missing feature with clear demand, and (3) 78 tickets misunderstand the difference between 'archive' and 'delete'—a documentation and UX labeling issue. It would recommend fixing the Safari bug immediately (high impact, contained scope), adding CSV export to the roadmap (medium effort, high satisfaction gain), and revising in-app copy before rewriting docs (faster fix for the labeling confusion).

The model can process hundreds of unstructured tickets in a single pass, identifying patterns across the full dataset rather than sampling. Image support means it can also parse screenshots users attached. Trade-off: batch mode means this is a post-mortem tool, not a live triage assistant—you run it after the quarter closes, not during an incident.

When 1M+ token context makes batch legal review viable in-house

A 12-person legal ops team needs to cross-reference 40-page NDAs against master service agreements and flag inconsistencies before quarterly board reviews. GPT-5.6 Terra handles this in batch mode: the 1.05M token window fits 8-12 full contracts plus your internal playbook in a single call, and at $1/Mtok input you're paying ~$1.05 per analysis run. Output is $6/Mtok, so a 2,000-token summary costs another $0.012. Run 200 contract sets per quarter and you're at $210 total, versus $8K for a paralegal week. The catch: batch mode adds 12-24 hour latency, so this only works if your review cycle isn't same-day. If you need answers in under an hour, you're paying real-time premiums elsewhere.

Why batch pricing beats real-time for low-frequency financial analysis

A 4-person investor relations consultancy processes 60 earnings call transcripts per quarter, extracting risk disclosures and sentiment shifts for client decks. Each transcript runs 25K-40K tokens; GPT-5.6 Terra in batch mode costs $0.025-$0.04 per input call, and a 1,500-token summary adds $0.009 output. Total per transcript: ~$0.05. Compare that to real-time models at $3-5/Mtok input where you'd pay $0.075-$0.20 each. Over 60 calls that's a $1.50 spend versus $4.50-$12. The 24-hour batch window is fine because earnings calls cluster in a 3-week period and clients expect decks 48 hours post-call. If you're doing daily news monitoring instead, the latency kills you and you need a streaming model.

How mega-context batch runs handle once-a-year compliance rewrites

A 20-person HR tech startup updates its 150-page employee handbook once per year, cross-checking new state labor laws against 80 existing policies. The handbook plus legal references total 320K tokens; GPT-5.6 Terra's 1.05M window fits the entire corpus and all amendments in one batch call. At $1/Mtok input that's $0.32, and a 5K-token redline output costs $0.03. One annual run is $0.35 versus $18K for outside counsel. Batch latency doesn't matter because the refresh happens over a 2-week sprint in Q4. The model's file and image modalities let you feed scanned state PDFs directly. If you're doing monthly policy tweaks instead of annual overhauls, the per-run savings shrink and you'd want a faster model with lower input costs.