OpenAI: gpt-oss-20b

Review this pull request for security issues: a new user registration endpoint that accepts email, password, and an optional 'role' field from the request body.

Three concerns here. First, the 'role' field should never come from user input—that's a privilege escalation vector. Move role assignment to server-side logic with proper authorization checks. Second, I don't see rate limiting on this endpoint; credential stuffing attacks will succeed. Add per-IP throttling. Third, the password appears to be hashed after validation—swap that order so plaintext never hits logs. The email regex looks fine, but consider a verification flow to prevent disposable addresses if that matters for your use case.

This example shows the model's ability to spot multi-layered security flaws in a realistic code review scenario. The 131k token context window means it can handle full PR diffs with surrounding codebase context. However, without published benchmarks, we can't verify how its security-pattern recognition compares to specialist models.

Explain how PostgreSQL's MVCC works to a backend engineer who's used MySQL's InnoDB but never touched Postgres.

MVCC in Postgres means every row update creates a new version rather than overwriting in place—think append-only. When you UPDATE, the old tuple stays visible to transactions that started before your change, and a new tuple gets written with your data. InnoDB does MVCC too, but stores old versions in the undo log; Postgres keeps them right in the table heap. This is why Postgres needs VACUUM—to reclaim space from dead tuples. The upside: readers never block writers. The downside: bloat if you don't tune autovacuum. Your mental model from InnoDB mostly transfers, but watch for table size growth patterns you didn't see in MySQL.

Demonstrates technical explanation with audience awareness—bridging from known (InnoDB) to unknown (Postgres). The model's 131k context window is overkill for this task, but the low $0.02/Mtok input cost makes it economical for documentation generation workflows. The explanation is accurate but doesn't cite Postgres version differences, which a specialist model might.

Generate a SQL query to find customers who made purchases in Q1 2024 but not Q2 2024, assuming tables 'customers' (id, name) and 'orders' (id, customer_id, order_date).

```sql SELECT DISTINCT c.id, c.name FROM customers c INNER JOIN orders o1 ON c.id = o1.customer_id LEFT JOIN orders o2 ON c.id = o2.customer_id AND o2.order_date BETWEEN '2024-04-01' AND '2024-06-30' WHERE o1.order_date BETWEEN '2024-01-01' AND '2024-03-31' AND o2.id IS NULL; ``` This uses a LEFT JOIN to find Q1 customers with no matching Q2 orders. The NULL check in WHERE filters out anyone who ordered in Q2. More readable than a NOT EXISTS subquery for this case.

Shows competence at translating business logic into SQL with a practical join pattern. The model provides working code and explains the approach, suitable for mid-level engineering tasks. At $0.09/Mtok output, longer code generation runs cost more than some competitors, though the 131k context window handles large schema definitions without truncation.

When gpt-oss-20b makes sense for 24/7 ticket classification

A 12-person SaaS company routing 800+ support tickets daily needs fast, cheap classification before human handoff. gpt-oss-20b hits the price floor at $0.02/$0.09 per Mtok—roughly 40% cheaper than GPT-4o-mini on input, 10% cheaper on output. The 131k context window handles full ticket histories plus knowledge base snippets in a single call, so you're not chaining requests or losing thread. No public benchmarks means you'll want a 2-week A/B test against your current classifier, but if accuracy lands above 92% on your labels, the cost savings fund an extra support hire by Q3. Buy this if your ticket volume makes every tenth of a cent matter and you can afford the validation sprint upfront.

Why gpt-oss-20b works for overnight research digest jobs

A 4-person investment research team summarizes 200+ earnings transcripts and analyst reports each quarter into a single digest. They run the job overnight, so latency doesn't matter—cost and context do. gpt-oss-20b's 131k window fits most transcripts whole, eliminating the chunking logic that breaks narrative flow in longer models with smaller windows. At $0.02 input per Mtok, processing 50M tokens of source material costs $1 versus $2.50 on GPT-4o-mini. The lack of public benchmarks is less risky here because the team reviews every summary before publication anyway. If your use case is batch, human-reviewed, and input-heavy, this model pays for itself in the first run.

When gpt-oss-20b is the right call for employee self-service search

A 30-person logistics startup wants employees to query their Notion wiki, Slack archives, and SOPs through a single chat interface. Average query volume is 150/day, mostly straightforward lookups—tracking a shipment status, finding a vendor contact, pulling a policy doc. gpt-oss-20b's 131k context lets you stuff the top 20 search results into the prompt without retrieval re-ranking, and at $0.02/$0.09 per Mtok, monthly costs stay under $15 even with generous output. The trade-off: no benchmarks means you can't predict accuracy on edge-case questions, so plan to surface a 'not confident' flag and route 10-15% of queries to Slack for human backup. If your team tolerates occasional misses and you're optimizing for cost over perfection, this is the model.