by OpenAI · GPT-5 family · best for superseded legacy mini, migrate to GPT-5.4 nano
GPT-5 mini is the original mini-tier sibling to GPT-5, released 2025-08-07 — the model that established the family's mini pricing tier at $0.25/$2.00. It remains GA, but it occupies the rare position of being beaten on both quality and cost simultaneously: GPT-5.4 nano is cheaper ($0.20/$1.25) and scores higher on SWE-bench Pro (52.4% vs 45.7%) and GPQA Diamond (82.8% vs 81.6%). The one-sentence buyer's take: a competent legacy model with no clear advantage today — migration to GPT-5.4 nano or mini is the only correct strategic action.
| Benchmark | Score | Source |
|---|---|---|
| GPQA Diamond | 81.6% | datacamp.com 2025-08-07T00:00:00.000Z |
Six personas, six verdicts — the same panel that reviews every product on TopReviewed.
“Outperformed and out-priced at once — there's essentially no architectural rationale to start a new build here.”
GPT-5 mini is in an uncomfortable position: it has been outperformed and out-priced simultaneously. GPT-5.4 nano is cheaper ($0.20 vs $0.25 input) and scores higher on SWE-bench Pro and GPQA Diamond. There is essentially no architectural rationale to start a new build on GPT-5 mini today. The only reasons to keep it are in-flight workloads where re-validation costs more than the delta, and fine-tuning (which the 5.4 generation lacks). Migration to GPT-5.4 nano or mini should be on the roadmap for any team still here.
“A model with no wedge left — every competitive axis points to GPT-5.4 nano or mini, leaving only fine-tuning and inertia.”
Strategically, GPT-5 mini has no remaining wedge. On cost, GPT-5.4 nano undercuts it; on quality, both 5.4 nano and 5.4 mini beat it; on tooling, the newer generation has computer use and apply_patch. Its only differentiation is fine-tuning at the mini tier. Positioning is purely transitional, and market timing has passed entirely — it competes in 2026 as a 2025 model that its own successors have lapped on both price and capability.
“Cheap, but GPT-5.4 nano is cheaper and benchmarks better — there's no finance case to remain on it for new spend.”
At $0.25/$2.00 with 90% cached discount and 50% Batch, GPT-5 mini is cheap. But GPT-5.4 nano is cheaper still — $0.20/$1.25 — and benchmarks slightly better on coding and reasoning. There is no finance case to remain for new spend. For sunk-cost in-flight workloads, run the migration math: a predictable prefix-cached pipeline typically saves 20–30% moving to GPT-5.4 nano on equivalent traffic. Plan migration within the quarter unless fine-tuning ties you here.
“Behaves like any family model at the SDK level — but GPT-5.4 nano gives the same role, cheaper, with better behavior.”
At the SDK level GPT-5 mini behaves like any Responses-API model in the family. The missing pieces — computer use, apply_patch, skills — limit it from real agent work, but that was never the mini role. The honest take: GPT-5.4 nano gives the same operational role with better behavior at lower cost, and GPT-5.4 mini gives a real capability jump for ~3x the price. Either is a better target than staying. The one developer reason to remain is an existing fine-tune.
“Invisible to ChatGPT users for nearly a year — fast and adequate where it surfaces, thin the moment depth or current knowledge is needed.”
End users do not pick GPT-5 mini — it surfaces inside backend services in API-backed products. Where it does, the experience is fast and adequate for simple tasks. For anything requiring depth or current knowledge, the gaps versus the GPT-5.4 family show through quickly, and the 2024-05 cutoff is the oldest in the lineup. The model is essentially invisible to ChatGPT users; it has been routed around for almost a year.
“The clearest 'migrate now' case in the lineup — its own cheaper successor beats it on the benchmarks OpenAI chose to publish.”
The adversarial read is also the consensus read: GPT-5 mini is the one model where the skeptic and the vendor agree you should leave. Its own cheaper successor (GPT-5.4 nano) beats it on the two benchmarks anyone published. The 2024-05 cutoff is the lineup's oldest. The only non-inertia reason to stay is fine-tuning, and even that is a question of whether re-training on a newer base is worth it. There is no marketing claim to dispute here because OpenAI isn't really marketing this model anymore — it's a migration-continuity SKU.
The full research notes behind this review — verified against primary sources.
OpenAI discloses no parameter counts or structure for GPT-5 mini — all null. It is a unified reasoning model with text + image input and text-only output, o200k_base tokenizer, and a 400K context window. No further architectural detail is public.
GPT-5 mini is a competent 2025 mini-tier model, justifying its 7.3 reasoning score — GPQA Diamond 81.6% and SWE-bench Pro 45.7% were strong for its tier at launch but now trail GPT-5.4 nano. It is solid at high-volume, low-latency work — classification, extraction, summarization, simple chat, RAG. Long-context (7.6) reflects the 400K window. Multilingual (7.8), instruction-following (8.0), and function-calling (8.0) are reliable. Coding (7.2) and agentic (6.8) are capped by the missing native computer use, apply_patch, and skills, and by the generational gap. Vision (7.4) and document/OCR (7.2) handle image input. Real-time data (5.0) is the weakest axis: the 2024-05 cutoff is ~24 months stale, the oldest in the in-scope lineup, and the practical reason most teams should migrate.
| Benchmark | Score | vs Successor (GPT-5.4 nano) | vs Successor (GPT-5.4 mini) | Source |
|---|---|---|---|---|
| SWE-bench Pro | 45.7% | -6.7pp vs nano (52.4%) | -8.7pp vs mini (54.4%) | DataCamp |
| GPQA Diamond | 81.6% | -1.2pp vs nano (82.8%) | below mini (high 80s) | DataCamp |
SWE-bench Pro is not the same key as SWE-bench Verified, so that front-matter key stays null. Other standard evals (MMLU, AIME, LiveCodeBench, Tau-bench, etc.) have no published GPT-5-mini figure.
GPT-5 mini serves at roughly 150 tokens/sec at low/default reasoning with sub-second time-to-first-token — genuinely interactive. latency_tier is fast. It is faster than GPT-5 base but notably slower than GPT-5.4 mini, which OpenAI reports runs more than 2x faster. For bulk work, Batch removes latency from the equation.
| Surface | Cost | Notes |
|---|---|---|
| API input | $0.25 / 1M tok | |
| API output | $2.00 / 1M tok | |
| Cached input | $0.025 / 1M tok | 90% discount |
| Batch (in/out) | $0.125 / $1.00 | 50% off, 24h SLA |
| Direct UI | n/a | not user-selectable |
| Free tier | none (API) | |
| Fine-tuning | available | GPT-5 mini supports fine-tuning |
Reasoning-token note: hidden reasoning tokens bill as output at higher effort tiers. Keep mini at low/default effort.
API-only via the Responses API; not open-weights, license Proprietary, not self-hostable. Cloud-managed via Azure OpenAI and Azure AI Foundry; OpenRouter proxies it. Data residency covers US and EU. It supports fine-tuning, which GPT-5.4 mini does not — the one feature reason to remain. The tool surface predates computer use, apply_patch, and skills.
Governed by OpenAI's Preparedness Framework. No training on API inputs by default; opt-out and zero-retention available for enterprise. Compliance covers SOC2, GDPR, CCPA, and HIPAA (BAA). Content moderation is built in. Refusal calibration is in line with the original GPT-5 generation.
First-party SDKs in Python, TypeScript, Java, Go, and .NET, plus the OpenAI Agents SDK. Framework integrations span LangChain, LlamaIndex, Vercel AI SDK, and Pydantic AI. It powered a share of ChatGPT free-tier traffic in 2025 and remains embedded in unmigrated API backends. Popularity tier: growing (declining as migration proceeds).
No — GPT-5.4 nano is cheaper and benchmarks better; GPT-5.4 mini is a real capability jump. Pick one of them.
The base mini model is not on the deprecation list as of 2026-05-28, but it is fully superseded; treat it as migration-only.
Fine-tuning at the mini tier (GPT-5.4 mini does not offer it) and sunk-cost in-flight workloads.
The 2024-05 cutoff is ~24 months — the oldest in the in-scope lineup. Time-sensitive work needs web search.
Typically 20–30% on equivalent prefix-cached traffic moving to GPT-5.4 nano, plus better quality.
No, not by API default; enterprise opt-out and zero-retention exist.
cheaper successor at $0.20/$1.25 with better SWE-bench Pro and GPQA; the default migration target.
higher-capability successor at $0.75/$4.50 with computer use and a real quality jump.
generational peer from the 2025 cohort; comparable role.
Primary references used to verify this review.
Does not train on API inputs by default
Last verified 2026-05-27