OpenAI: GPT-5.4 Pro (batch)

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 job registry in worker.py lines 203-217, creating a race when multiple workers claim the same task; an unbounded retry loop in scheduler.py that could exhaust memory under sustained Redis failures; and a subtle ordering bug in the priority queue implementation where jobs with identical timestamps get non-deterministic scheduling. It suggests specific fixes—a threading.RLock for the registry, exponential backoff with a ceiling for retries, and adding a sequence counter to the priority tuple. The analysis stays architectural, ignoring formatting concerns.

The 1M+ token context window handles entire codebases in one pass, letting the model reason about cross-file interactions that shorter-context models miss. Output pricing at $90/Mtok makes exhaustive code review expensive—best reserved for pre-merge audits, not iterative debugging.

Summarise these 340 pages of clinical trial data (PDF attached). Extract adverse event rates, efficacy endpoints, and any protocol deviations. Structure as a 2-page executive brief for non-statisticians.

The model produces a structured brief: primary endpoint (progression-free survival) showed 4.2-month improvement vs. placebo (HR 0.68, p<0.001); secondary endpoints mixed—overall survival trended positive but missed significance. Adverse events: grade 3+ neutropenia in 18% of treatment arm vs. 4% control; no unexpected safety signals. Three protocol deviations noted—two sites enrolled patients outside the age window, one site had incomplete follow-up data for 12 subjects. The summary avoids jargon, translating hazard ratios into plain language and flagging which findings matter for regulatory decisions.

File ingestion and the massive context window make this viable for document-heavy workflows—legal discovery, research synthesis, compliance audits. Batch processing keeps costs manageable for high-volume jobs, but the 24-hour turnaround means this isn't for real-time use cases.

Generate 15 variations of this product description for A/B testing. Keep core facts identical but vary tone, sentence structure, and emphasis. Target: e-commerce landing page, mobile-first audience.

The model produces 15 distinct versions: some lead with social proof ('Join 12,000+ users who...'), others with urgency ('Limited stock—ships in 24 hours'), others with feature benefits ('Waterproof to 50m, so your morning swim becomes...'). Sentence lengths vary—short punchy fragments for skimmers, longer narrative arcs for engaged readers. Tone shifts from aspirational to practical to playful across variants. Each version preserves the same technical specs, dimensions, and warranty terms while feeling meaningfully different in voice and structure.

Multimodal input means you can attach the existing product page screenshot for context. The model's training on commercial copy shows—it understands conversion-focused writing patterns. At $15 input per million tokens, running dozens of creative variations is affordable, unlike per-request pricing on smaller models.

When 1M+ token context justifies $90/Mtok output on contract review

A 12-person legal ops team needs to cross-reference clauses across 40+ vendor agreements before a merger closes in three weeks. GPT-5.4 Pro (batch) fits here because the 1.05M token window holds entire contract sets in one pass, eliminating the error-prone chunking that breaks cross-document reasoning. At $15 input / $90 output, a typical 800k-token load with 12k-token summaries costs ~$13.08 per batch run—acceptable when the alternative is 60 billable associate hours at $250/hour. The batch API adds 12-24 hour latency, so this works only when you can queue overnight runs and review results the next morning. If you need same-day turnaround or your contract sets exceed 1M tokens after OCR expansion, switch to a streaming endpoint or a cheaper high-context model and accept the chunking tax.

How a 4-person finance team uses batch mode for investor decks

A Series B startup's finance team compiles board decks from 9 months of Stripe exports, HubSpot pipeline snapshots, and payroll CSVs every quarter. GPT-5.4 Pro (batch) handles this because the 1M+ context window ingests all source files as structured data, and the batch discount (typically 50% off synchronous pricing) matters when output runs 8-10k tokens per deck. The $90/Mtok output rate is steep, but a 900k-token input with 9k-token output costs ~$14.31—cheaper than three hours of fractional CFO time at $200/hour. Batch latency (up to 24 hours) aligns with the quarterly cadence; no one needs the deck in real time. If you're generating these monthly or need sub-hour turnaround for investor calls, pay the synchronous premium or use a model with cheaper output tokens.

When multimodal batch processing beats per-image API calls for alt-text

A 3-person e-commerce agency manages 1,200+ product images across six Shopify stores and needs WCAG-compliant alt-text before a site relaunch. GPT-5.4 Pro (batch) works because it accepts image inputs and the batch queue lets you submit all 1,200 images in one job, avoiding rate-limit throttling and synchronous API overhead. At $15 input (images tokenize variably, assume ~1k tokens each) and $90 output (~80 tokens per alt-text), total cost runs ~$26.64 for the full set—$0.02 per image. The 12-24 hour batch window is fine when the relaunch is five days out. If you need real-time alt-text generation as products upload, or if your image volume exceeds 5k/month, switch to a synchronous multimodal model with lower output pricing to avoid the $90/Mtok penalty on high-frequency workflows.