OpenAI: GPT-5.6 Terra

Review this 40-page technical specification for our distributed caching layer and identify architectural risks we haven't considered. Focus on edge cases at scale.

In an illustrative response, the model would parse the entire specification within its 1M+ token context, then surface three critical concerns: first, a race condition in the invalidation protocol when partition counts exceed 512 nodes; second, memory pressure scenarios during simultaneous cache warming across regions; third, an unaddressed failure mode where stale reads persist beyond TTL during network partitions. Each concern would include the specific section reference, a concrete failure scenario, and a mitigation path grounded in the spec's existing constraints.

This example showcases the 1.05M token context window — the model can hold an entire technical document in working memory and reason across distant sections. The depth of architectural analysis reflects GPT-5.6's improved reasoning over prior generations, though at $12/Mtok output, lengthy reviews become expensive quickly.

Generate a React component that renders a data table with virtual scrolling, inline editing, and undo/redo. Include TypeScript types and accessibility attributes.

The model would produce a complete, production-ready component spanning roughly 200 lines: a virtualized table using Tanstack Virtual, cell-level edit state managed through useReducer with a command pattern for undo/redo, ARIA labels on interactive elements, and strict TypeScript interfaces for row data and column definitions. The code would include inline comments explaining performance trade-offs (why virtualization matters above 1000 rows) and accessibility decisions (keyboard navigation between editable cells).

Demonstrates code generation quality and architectural taste — the model selects appropriate libraries rather than reinventing virtualization, and balances feature completeness with maintainability. The multimodal file support means you could attach existing component files for context, though that drives up input costs at $2/Mtok.

I'm attaching a photo of a handwritten flowchart from our whiteboard session. Convert this into Mermaid diagram syntax and suggest two improvements to the logic flow.

The model would first describe what it sees in the image — boxes, arrows, decision diamonds, handwritten labels — then output clean Mermaid syntax recreating the flowchart structure. Following that, it would propose two concrete improvements: perhaps collapsing three sequential validation steps into a single guard clause, and adding an explicit error-handling branch that the whiteboard sketch left implicit. Each suggestion would reference specific node labels from the original diagram.

Highlights the image modality: the model can parse visual information and translate it into structured code. This workflow (photo → diagram syntax) is common in documentation tasks. The suggestions demonstrate reasoning about logic flow, not just transcription. However, handwriting recognition quality varies with photo clarity and penmanship.

When 1M+ token context justifies the $12/Mtok output premium

A 4-person corporate dev team closing two acquisitions per quarter needs to cross-reference 80+ contracts, financials, and IP filings in a single session. GPT-5.6 Terra's 1.05M token window lets you load the entire data room—purchase agreements, employment contracts, patent docs—and ask comparative questions without re-uploading or chunking. The $2 input cost is negligible when you're ingesting once; the $12/Mtok output matters only if you're generating 50-page summaries (you're not—you're extracting 2-3 page red-flag reports). If your diligence sets stay under 400K tokens or you're running this weekly instead of quarterly, Claude 3.5 Sonnet's 200K window at $15 output saves money. For infrequent, document-heavy closings where context is everything, Terra's window wins.

Overkill for most teams, right-sized for executive synthesis

A 12-person SaaS startup's finance lead compiles board decks from 90 days of Slack exports, CRM notes, and financial statements. Terra's million-token context means you can feed three months of conversation threads, customer feedback, and P&L detail in one prompt, then ask for a 15-slide narrative with specific call-outs. The $12/Mtok output cost stings—a 10K token deck costs $0.12, versus $0.075 on Claude—but you're generating four decks per year, so the annual delta is under $2. The real question is frequency: if you're drafting weekly investor updates or customer case studies, the output premium compounds fast and you should default to Sonnet. For quarterly synthesis where context breadth beats cost, Terra delivers.

When you need the whole repo in context but can't afford the output tax

A 6-engineer team migrating a 300K-line monorepo to microservices wants to ask architectural questions across 40 modules at once. Terra's 1M+ window handles the entire codebase—controllers, models, tests, configs—so the model sees dependency chains without you manually mapping them. The $2 input rate is cheaper than Sonnet ($3) for this use-case. The problem is output: refactoring generates thousands of lines of diff suggestions, and at $12/Mtok you're paying 60% more than Sonnet's $15 rate per generated token. If you're prototyping and only need 2-3K token answers, Terra works. If you're generating migration scripts or bulk rewrites (20K+ tokens per session), the output cost kills the ROI. Use Terra for analysis and planning; switch to Sonnet or Gemini 2.0 Flash for code generation.