OpenAI: GPT-5.1 (batch)

Review this Python function for potential race conditions. It's meant to update a shared counter across multiple threads, but we're seeing inconsistent totals in production.

The function uses a simple increment without synchronization. Between reading `self.counter` and writing back the new value, another thread can modify it — classic read-modify-write race. You'll lose increments under load. Wrap the operation in a `threading.Lock()` or switch to `threading.local()` if each thread needs its own counter. If you need atomic cross-thread totals, consider `multiprocessing.Value` with a lock or move to a queue-based architecture where one thread owns the counter.

This example highlights GPT-5.1's ability to diagnose concurrency bugs from code context alone. The 400k token window means you can paste entire modules for holistic review. Trade-off: batch mode adds latency (minutes to hours), so this isn't suitable for interactive debugging sessions.

Summarise the key regulatory changes in this 180-page EU AI Act PDF, focusing on obligations for general-purpose AI providers. Flag any ambiguities that might affect our compliance roadmap.

General-purpose AI providers must conduct systemic risk assessments if their model exceeds 10²⁵ FLOPs during training (Article 51). You're required to document training data governance, implement adversarial testing, and report serious incidents to national authorities within 72 hours. Ambiguity: 'serious incident' lacks a quantitative threshold — it's defined as causing 'significant harm', which remains subjective. Article 53's transparency obligations also don't specify whether synthetic data counts toward the data-source disclosure requirement. Recommend legal review on both points before your Q3 audit.

Demonstrates file ingestion and long-document synthesis — the model can process the full Act in one pass rather than chunking. The compliance-focused framing shows domain reasoning. Trade-off: batch pricing is lower than real-time, but you can't iterate on follow-up questions immediately.

Generate a product photo concept for a minimalist ceramic mug. The setting should feel Scandinavian — natural light, muted tones, wood textures. Describe the composition and mood in detail.

The mug sits off-center on a pale oak surface, angled 30° to reveal its hand-thrown rim texture. Soft north-facing window light creates a gradient from cool white (top-left) to warm beige (bottom-right). A linen napkin in oatmeal folds loosely beneath the handle. Background: defocused birch branches, barely visible. The glaze is matte sage, with a subtle finger-drag mark near the base — intentional imperfection. Mood: quiet, tactile, unhurried. The composition uses negative space heavily; the mug occupies only the lower third of the frame.

Shows multimodal prompt handling — the model interprets visual direction without generating an actual image (GPT-5.1 is text-output only, but accepts image inputs for analysis). Useful for briefing designers or generating alt-text. Trade-off: you'll still need a separate image model to render the concept.

When batch pricing makes 400K-token contract review profitable

A 12-person legal ops team processes 80-120 vendor contracts nightly, each 30-80 pages with exhibits. GPT-5.1 batch handles the full contract in one 400K-token context at $0.63/Mtok input—roughly $0.25 per contract for extraction and risk flagging. The 50% batch discount over real-time API makes this viable where streaming models at $1.26/Mtok would blow the budget. Output at $5.00/Mtok stays reasonable because summaries run 2-3K tokens per contract. The trade-off: batch jobs complete in 12-24 hours, so this only works for overnight workflows where morning review is acceptable. If you need sub-minute turnaround on any contract, route those to real-time GPT-5.1 and keep batch for the nightly queue.

How 400K context and image support handle legacy Rails migrations

A 5-engineer SaaS team is migrating a 7-year-old Rails app from Heroku to AWS, touching 40+ migration files, schema diagrams, and deployment scripts. GPT-5.1 batch ingests the entire migration history, ERD screenshots, and current schema.rb in one 400K-token job, then generates a sequenced refactor plan with rollback steps. The image modality reads the hand-drawn architecture sketches from the 2018 Notion doc that no one documented in code. At $0.63 input, a 300K-token job costs $0.19; the 8K-token output plan costs $0.04. Batch latency (12-24 hours) is fine because the team reviews the plan in standup the next morning. If you're doing this interactively in a pairing session, real-time GPT-5.1 is the call despite the 2x cost.

When batch latency beats real-time cost for 90-day data aggregation

A 20-person finance team at a Series B startup generates board decks every 90 days, pulling from Stripe exports, QuickBooks CSVs, and 60-page investor update drafts. GPT-5.1 batch processes 250K tokens of structured financial data plus file attachments, outputting a 15-slide narrative with variance analysis and cash runway projections. The job costs $0.16 input + $0.08 output ($0.24 total) and completes overnight before the CFO's 9am review. Real-time GPT-5.1 would cost $0.48 for the same job—double the price for latency the team doesn't need. The 400K context window means no chunking or multi-pass summarization, which historically introduced calculation errors. If your board deck is due same-day and you're starting at 2pm, pay the real-time premium.