OpenAI: gpt-oss-120b (free)

Review this Python function for potential bugs. It's supposed to merge two sorted lists but users report incorrect output when one list is empty.

The function has an off-by-one error in the while loop condition. When `list_a` is empty, the code tries to access `list_a[i]` before checking bounds, causing an IndexError. Move the length checks before the comparison. Also, the final concatenation logic assumes both pointers are valid — you need separate conditions for appending remaining elements from each list. Here's the corrected version with explicit boundary handling for both edge cases.

This example shows the model's debugging capability within its 131K token context window — enough to handle substantial codebases. The free pricing makes it accessible for exploratory debugging sessions. However, without public benchmarks, we can't verify its accuracy against established coding evaluation suites like HumanEval.

Summarize the key regulatory changes in this 40-page compliance document, focusing on what our finance team needs to action before Q2.

Three critical changes require immediate attention: (1) New expense categorization rules for software subscriptions take effect April 1st — all SaaS contracts must be reclassified under operating expenses rather than capital expenditure. (2) Monthly reconciliation deadlines move from the 10th to the 5th business day. (3) International wire transfer documentation now requires additional beneficiary verification for amounts exceeding $50K. The document also introduces voluntary best-practice guidelines for ESG reporting, but these carry no compliance deadline for your industry segment.

The 131K context window handles long documents without chunking, making this model suitable for enterprise document analysis. Zero cost per token means teams can process high volumes of internal documents without budget concerns. The trade-off is uncertainty around accuracy on specialized domains without benchmark validation.

I'm teaching a 10-year-old about photosynthesis. Explain it using an analogy they'd understand, avoiding scientific jargon.

Think of a plant like a tiny solar-powered factory. The leaves are like solar panels that catch sunlight. Inside the leaves, there are millions of little workers (called chloroplasts, but let's call them 'green helpers'). These green helpers take three ingredients — sunlight from above, water from the roots, and carbon dioxide from the air we breathe out — and mix them together to make food (sugar) for the plant. The leftover stuff they don't need? That's oxygen, which they release back into the air for us to breathe. So plants are basically running on sunshine and making the air better for us at the same time.

This demonstrates the model's ability to adjust technical depth for different audiences — a key strength for educational content and internal knowledge-sharing. The free tier removes cost barriers for high-volume tutoring or documentation use cases. Without benchmark data on instruction-following tasks, we can't compare its pedagogical effectiveness to models with measured MMLU or HellaSwag scores.

When free inference makes sense for early-stage bot testing

A 4-person startup building a customer support bot needs to iterate on prompts and conversation flows without burning runway on API costs. gpt-oss-120b at $0.00/Mtok lets you run thousands of test conversations during the design phase—swap system prompts, test edge cases, refine tone—without watching a meter. The 131k context window handles multi-turn dialogues and lets you include full conversation history plus documentation chunks. The trade-off: no public benchmarks means you're flying blind on quality until you test it yourself, and 'oss' in the name suggests this may be a community or experimental release with less reliability than production models. Use this for the first 2-3 weeks of prototyping, then migrate to a benchmarked model once you've locked the interaction pattern and need predictable output quality.

Why zero-cost inference breaks down at moderation scale

A 12-person social platform needs to classify 200k user comments per day for policy violations. At $0.00/Mtok, gpt-oss-120b looks like the obvious choice—your monthly inference bill stays at zero even as volume spikes. The problem is accuracy. Content moderation demands low false-positive rates (don't ban good users) and low false-negatives (don't miss real abuse), and without MMLU, HumanEval, or safety benchmarks, you can't validate this model's precision on adversarial text. The 131k window is overkill here since each moderation call is under 500 tokens. If you're moderating under 10k items/day and can manually review flags, this works as a first-pass filter. Above that threshold, switch to a model with published safety evals—the cost difference is negligible compared to the reputation risk of a moderation failure.

When free LLM inference works for low-stakes knowledge retrieval

A 20-person engineering team wants to build a Slack bot that answers questions about internal wikis, runbooks, and API docs. gpt-oss-120b's 131k context window fits 40-50 pages of documentation in a single prompt, and at $0.00/Mtok you can serve unlimited queries without justifying budget to finance. The use-case is low-stakes: if the bot hallucinates a config setting, an engineer will catch it before pushing to prod. The missing benchmarks matter less here because you're not making high-consequence decisions based on the output—you're just surfacing relevant passages faster than Cmd+F. Deploy this for 2-3 months and track answer accuracy in Slack reactions. If accuracy stays above 80% and the team uses it daily, keep it. If you see frequent corrections or engineers stop asking, upgrade to a model with published retrieval benchmarks and tighter instruction-following.