OpenAI: GPT-5.4

Review this 847-line Python codebase for a distributed task queue. Flag architectural issues, not style nitpicks. Focus on concurrency bugs and data-race conditions.

The model identifies three critical issues: a missing lock around the shared dequeue operation in worker.py lines 203-217 that creates a race condition under high concurrency; an unbounded memory growth pattern in the retry logic where failed tasks accumulate without expiration; and a subtle deadlock risk in the shutdown sequence where workers wait on a condition variable that the coordinator may never signal if exceptions occur during cleanup. It suggests specific fixes with code snippets, referencing the exact line ranges and variable names from the uploaded files.

The 1M+ token context window enables whole-codebase analysis that smaller models can't attempt. GPT-5.4 maintains coherence across hundreds of files without summarization loss. However, at $15/Mtok output, a thorough review of this scale costs $12-18 per run—viable for critical pre-deployment audits, prohibitive for iterative debugging.

I'm attaching 340 pages of legal discovery documents as PDFs. Extract every mention of 'Project Falcon' with surrounding context, then build a timeline of decisions from January 2023 to present.

The model processes all 340 pages, extracting 47 references to 'Project Falcon' with 2-3 sentences of surrounding context for each. It constructs a chronological timeline spanning 14 months, identifying five key decision points: initial budget approval in January 2023, scope expansion in April, vendor selection in July, the November compliance review that flagged IP concerns, and the February 2024 pivot to an in-house solution. Each timeline entry cites specific page numbers and document titles from the uploaded set.

Multi-modal file ingestion handles native PDFs without preprocessing, and the extended context keeps all 340 pages active simultaneously—no chunking artifacts or lost cross-references. The $2.50/Mtok input cost makes this a $2.60 query for the document set, reasonable for legal workflows. Output quality depends on OCR accuracy in the source PDFs.

Generate a 12-week onboarding curriculum for a senior ML engineer joining our recommender systems team. Include reading lists, hands-on projects, and knowledge-check milestones. Assume they know PyTorch but not our internal tooling.

The model produces a structured 12-week plan: weeks 1-2 cover the company's feature store architecture and data pipelines with links to internal documentation; weeks 3-5 introduce the production recommendation stack through progressively complex hands-on tasks (replicate a historical A/B test, then propose a new ranking signal); weeks 6-8 focus on model monitoring and debugging with real incident post-mortems as case studies; weeks 9-12 involve a capstone project—shipping a small ranking improvement to 5% traffic. Each week includes 3-4 curated readings, a concrete deliverable, and a 30-minute knowledge-check conversation topic.

GPT-5.4's training data extends into 2024, so it references current ML tooling (Feast, Metaflow, Ray) without anachronisms. The curriculum structure is immediately usable, not a generic template. The model doesn't hallucinate internal tool names, staying appropriately vague where it lacks specifics. At this output length (~800 tokens), cost is $0.01—negligible for planning tasks.

When 1M+ token context justifies the $15/Mtok output premium

A 12-person litigation support team needs to cross-reference depositions, contracts, and email threads spanning 800+ pages per case. GPT-5.4's 1.05M token window lets you load entire discovery sets in one context—no chunking, no retrieval lag, no missed cross-references. At $2.50 input / $15 output per Mtok, a typical 600k-token load costs $1.50 to ingest; if the model generates 20k tokens of analysis ($0.30), you're at $1.80 per case brief. That beats the hourly cost of a paralegal re-reading files or the engineering overhead of a RAG pipeline. The trade-off: if your cases average under 200k tokens, Claude 3.5 Sonnet's 200k window at $3/$15 is cheaper and faster. Above 400k tokens, GPT-5.4's context advantage pays for itself in the first query.

Why the output cost kills this model for chat-heavy workflows

A 20-seat SaaS support team handles 2,000 tickets daily, each requiring 300-500 tokens of response (routing notes, canned replies, escalation summaries). At $15/Mtok output, 400 tokens per ticket × 2,000 tickets = 800k output tokens/day = $12/day in generation costs alone. Over a month, that's $360 in output fees before you count input tokens or API overhead. Compare Gemini 1.5 Flash at $0.30/Mtok output: the same workload costs $9/month. GPT-5.4's image and file handling might justify the premium if you're triaging screenshot-heavy tickets, but for text-only support queues under 5,000 tickets/day, the output pricing makes this a non-starter. Switch to a sub-$5/Mtok model unless you're also using the 1M+ context for knowledge-base grounding in every call.

When infrequent, high-stakes analysis justifies the context-output combo

A 4-person finance team at a PE firm runs quarterly portfolio reviews: 10-Qs, earnings transcripts, and analyst decks from 15 portfolio companies, totaling 1.2M tokens per quarter. They need a single synthesis memo comparing revenue trends, risk disclosures, and management commentary across all holdings. GPT-5.4's 1.05M window fits the entire corpus; the team prompts once, gets a 15k-token memo ($0.225 output cost), and the total run is under $4 including input. For a task that happens 4× per year and replaces 8 hours of analyst time, the $16 annual model cost is invisible. The caveat: if you're running this monthly or need to iterate on the memo (3-4 revisions), the output cost stacks up fast. At that frequency, build a RAG system with a cheaper model or cap output tokens aggressively.