GPT-5 mini: Full Spec Sheet (June 2026)

gpt-5-mini is a smaller, faster, cheaper variant of GPT-5 built on the same architecture and training pipeline. OpenAI does not publish parameter counts for either model, but in practice gpt-5-mini is roughly the GPT-5 family's equivalent of gpt-4o-mini in the GPT-4 family — same instruction-following discipline, weaker on hard reasoning, dramatically cheaper.

Crucially, the feature surface is identical: gpt-5-mini supports the same function calling, parallel tool calls, structured outputs (JSON Schema-validated), prompt caching, the Responses API endpoint, vision input, and reasoning-effort control. Anything you write for GPT-5 runs on gpt-5-mini with a single model-ID change. The difference shows up on tasks that need multi-step reasoning, complex code synthesis, or strict factual accuracy.

OpenAI's positioning: gpt-5-mini is the 'default for production' tier. Most chat assistants, structured-data extractors, classification pipelines, content scaffolders, and routing agents should live here. Reserve GPT-5 for the small fraction of traffic where cost-of-error genuinely dominates per-call cost.

Standard rates: `cost = (input_tokens / 1M) × $0.25 + (output_tokens / 1M) × $2`. The representative 1,000-in / 500-out call: `0.001 × $0.25 + 0.0005 × $2 = $0.00025 + $0.001 = $0.00125`. About 0.125¢ per call — one-fifth the GPT-5 cost on identical tokens.

At 1,000,000 calls/month with that profile, the standard rate runs $1,250 vs GPT-5's $6,250. Apply 80% prompt cache hit on a 1,500-token system prefix and the per-call input drops further; apply Batch API for asynchronous workloads (summarization, classification, evaluation) and another 50% off both streams.

Real-world example: a customer-support classification pipeline at 2M calls/month with 800-in / 100-out averages on gpt-5-mini runs `0.0008 × $0.25 + 0.0001 × $2 = $0.0004 per call = $800/month`. With prompt caching activating on the stable system prefix, ~$500/month. On GPT-5, the same workload is $4,000/month. The model choice is the largest cost lever in the GPT-5 family by a wide margin. Worked $ at scale: OpenAI API cost calculator.

gpt-5-mini ships with the same 400,000-token context window as GPT-5 and the same 128,000-token max output ceiling. There is no smaller-context variant; OpenAI standardized the window across the GPT-5 family so prompt-engineering work is portable across tiers.

Practical implication: prompts that work on GPT-5 fit on gpt-5-mini without context-related rewrites. The price gap is purely a quality gap, not a context gap. If you're moving traffic from GPT-5 to gpt-5-mini for cost reasons, the migration is a model-ID swap plus an eval pass — not a prompt rewrite.

As always, cap output. Set `max_output_tokens` to the realistic ceiling for your task (1,500 for chat, 4,000 for code, 8,000 for long-form). The default ceiling (128K) is a defensive guard, not a target.

**Wins (use gpt-5-mini)**: classification, named-entity extraction, summarization of structured input, format conversions (JSON ↔ YAML ↔ Markdown), simple chat assistance, routing/dispatch agents, content scaffolding (outlines, first drafts), structured-output pipelines, internal telemetry classification, document tagging, sentiment analysis. Quality on these is indistinguishable from GPT-5 on a held-out eval for most teams.

**Loses (use GPT-5)**: complex code synthesis with multi-file context, math proofs, multi-step planning that requires backtracking, legal/financial analysis where correctness is non-negotiable, fine-grained vision reasoning (counting objects in dense scenes, reading small text), competitive coding-benchmark-level tasks. The 5× price premium for GPT-5 is justified when these are the bottleneck.

Run an eval before you commit. Take 200 representative inputs from your production traffic, run both models, blind-score the outputs. Most teams discover gpt-5-mini covers 70-90% of traffic with no measurable quality drop and reserve GPT-5 for the rest — typically routed via a simple complexity classifier.

gpt-5-mini at $0.25 / $2 sits between Claude Sonnet 4.6 ($3 / $15 — Anthropic's mid-tier, slightly higher quality, much higher price) and Gemini 2.5 Flash ($0.30 / $2.50 — Google's mid-tier, comparable price, 1M context).

vs **Claude Sonnet 4.6**: 12× cheaper on input, 7.5× cheaper on output. Sonnet wins on long-form writing voice, instruction-following discipline on complex prompts, and prompt caching savings (90% off cached reads via Anthropic). gpt-5-mini wins on raw throughput and price. See Claude Sonnet vs GPT-5 mini for the side-by-side.

vs **Gemini 2.5 Flash**: comparable price ($0.25 vs $0.30 input), comparable output ($2 vs $2.50), but Flash has a 1M context window vs gpt-5-mini's 400K. Flash also supports native audio input. gpt-5-mini wins on structured outputs (OpenAI's JSON Schema enforcement is more mature) and the broader OpenAI tooling ecosystem.

Cross-tier: gpt-5-mini at 5× cheaper than GPT-5 covers the same feature set; most teams should default here and escalate only on demonstrated need.

gpt-5-mini exposes the same `reasoning_effort` parameter as GPT-5 (`minimal`, `low`, `medium`, `high`). The same rules apply: reasoning tokens bill at the output rate, are not returned to you, and can dramatically inflate cost if defaulted to `high`.

On gpt-5-mini specifically, `minimal` is the right default for high-volume mechanical tasks (classification, extraction). The cost-per-call stays under $0.001 for typical inputs. `low` is the right default for chat and content generation. `medium` and `high` are usually wasted on gpt-5-mini — if the task needs serious reasoning, the better move is escalating to GPT-5 with `medium`, not running gpt-5-mini at `high`.

Common mistake: leaving `reasoning_effort` at the default (`medium`) on a classification workload at 1M+ calls/month. The reasoning tokens silently 2-3× the bill. Always explicitly set the effort level on production prompts.

Every number on this page was verified against OpenAI's live documentation on 2026-06-20. Sources: platform.openai.com/docs/models/gpt-5-mini, openai.com/api/pricing, platform.openai.com/docs/api-reference/responses.

Prices on the GPT-5 family have moved once since launch (a downward adjustment on cached input in late 2025). OpenAI does not version their pricing page with changelog entries. Re-verify the live page quarterly if your monthly bill exceeds $500.

If you find a discrepancy with the live page, treat the live page as canonical. We re-fetch and update this card monthly.