GPT-5 vs GPT-5 Mini: When Each Makes Sense

The two tiers, in one sentence

GPT-5 is the flagship tier for reasoning that has to be right. GPT-5 mini is the cheap-and-fast tier for high-volume low-stakes work that should not eat the budget.

Two tiers, two jobs. The mistake most operators make is treating the lineup like a ladder they must climb. The OpenAI GPT-5 stack is not a ladder. It is two parallel tools with different cost and capability profiles, and the right one depends entirely on the call you are about to make.

The decision rubric

Task	Pick
One-line reply, classification, JSON extraction	mini
High-volume triage, tagging, routing	mini
Single-file code edit with clear intent	mini
Cheap shell-out layer in a multi-agent system	mini
Multi-file refactor across modules	full
Long-document analysis (200k–400k tokens)	full
Multi-step agentic workflows with tool use	full
Vision-heavy prompts (image input + reasoning)	full
Math, proof, multi-hop reasoning	full
Reasoning where a missed step is a real cost	full

If a task fits two rows, pick the cheaper tier first. Promote only after the cheaper tier demonstrably fails on the downstream.

Cost per million tokens

Live prices from the OpenAI model pages, captured 2026-07-15. USD per million tokens.

Model	Input	Cached input	Output	Context
GPT-5	$1.25	$0.125	$10.00	400k
GPT-5 mini	$0.25	$0.025	$2.00	400k
Ratio (full ÷ mini)	5x	5x	5x	1x

The per-token ratio is exactly 5x in every direction. Output costs 8x input on each tier. The “≈25x” figure that circulates is a workload composite: a typical agentic call generates ~5x more output than input (reasoning, function-call payloads), so on a mixed workload the 5x output multiplier compounds to ≈25x end-to-end. Use 5x for per-token forecasting; reach for ≈25x when comparing two production dashboards.

The six capability axes

1. Reasoning depth — Mini handles single-hop and most two-hop problems; GPT-5 pulls ahead on multi-step chains. Gap widens past four hops.
2. Coding — Mini writes code fine for single-file edits; GPT-5 wins on multi-file refactors and code review where missed edge cases are real cost.
3. Agentic tool use — Mini handles ≤ ~5 tool steps; GPT-5 holds long trajectories reliably. Drift on long chains is the single biggest reason to upgrade.
4. Speed — Mini is noticeably faster on short prompts because it spends less compute per token. On long prompts the gap narrows — latency is dominated by output length.
5. Cost — 5x per-token in every direction; ≈25x on output-heavy workloads.
6. Context utilization — Both tiers see 400k tokens; GPT-5 holds detail across the window more reliably. Past 200k tokens the flagship pulls ahead on recall.

When to override the rubric

A “should-be-mini” task failing three times in a row — promote. A “should-be-full” task landing clean on mini first try — demote and bank the savings. The rule: move up when output is unfit for the downstream; move down when output is fine and cost is obvious. The rubric is a starting bias, not a law.

The 3 tasks where mini is wrong

1. Long-document analysis (≥ 200k input tokens)

Both tiers share the 400k window, but context length is not context utilization. When the input is a long PDF, multi-file code dump, or full transcript, mini loses the middle more often — the classic “lost in the middle” failure. A 250k-token contract analysis on mini is a coin flip on whether the clause on page 94 surfaces; the same prompt on full lands it. More input tokens amplify the 5x input multiplier, but a wrong answer usually forces a downstream re-run. Promote when the document drives a real decision.

2. Multi-step agentic workflows (≥ 5 tool calls)

Mini handles a 3-step chain fine. Past 5 steps, mini drifts: the wrong tool gets called, a constraint from step 2 drops by step 7, the loop forgets the goal. Each drift costs a retry, and each retry burns the savings that justified mini. A 7-call task on mini needing 2 retries costs more than a 7-call task on full that completes first try. For agent harnesses: route the orchestration shell (planning, decision) to full, let mini handle cheap leaf nodes (classification, extraction).

3. Vision-heavy prompts

Both tiers accept image input. The capability gap on image-grounded reasoning is wider than on text. Mini reads the image; GPT-5 reasons over it. “Summarize this screenshot” — mini is fine. “Given the diagram, derive the next step the user is stuck on” — full. Vision tokens bill as input on both tiers, but vision calls generate more output reasoning, pushing them toward ≈25x before any agentic loop.

The 3 tasks where mini is right

1. Chatbot one-liners and short replies

The bulk of chatbot traffic is a few hundred input tokens, a paragraph of output, and a downstream that tolerates a slightly worse answer because the user can ask a follow-up — exactly the shape mini is tuned for. A 200-token support reply on full multiplies the bill 5x per-token to ≈25x workload-composite for a quality delta the user cannot detect. Exception: chatbots with regulated outputs (medical, legal, financial), where the answer has to be defensible.

2. High-volume classification and routing

A ticket classifier, an email router, a JSON extractor from a structured receipt — all high-volume, short-input, clear ground truth. Mini handles them at frontier quality; per-call cost dominates everything. A support pipeline classifying 50,000 tickets per day on mini costs ≈$0.25 per 1k tickets input; the same on full costs ≈$1.25. Across a year, the difference is a hire.

3. First-draft generation for human or downstream editing

Email drafts, blog outlines, code stubs, product descriptions, meeting recaps — anything a human or second pass will edit. Mini’s quality is “good enough to start from”; the downstream editor saves more time than the mini-vs-full quality gap. Promote only when the draft is the final deliverable (no editor) and the audience cannot tolerate a rewrite. The instinct to “use the best model for important emails” is wrong: pick the model that produces a draft needing the least editing, not the model with the highest benchmark.

Common failure modes

Mistake	What it costs
Defaulting every call to GPT-5	5x input, ≈25x workload budget on trivially easy tasks; mini would ship faster and cheaper
Routing flagship tasks to mini to save money	Retried or quietly shipped wrong — savings turn into rework
Locking one tier forever	New OpenAI release shifts the curve; worst task becomes most expensive line
Benchmarking on one prompt	One call is one data point; pick tier on five diverse tasks
Confusing model aliases	gpt-5, gpt-5-mini, gpt-5-nano are three models at $1.25 / $0.25 / $0.05 per MTok input
Quoting input pricing ignoring output	Output is 8x input on every tier; output-heavy full calls dominate the bill

What not to do

• Don’t lock your app to one tier. Compose a routing layer: cheap first, expensive on failure.
• Don’t pick GPT-5 by reflex. Operators fresh out of a benchmark high are the most expensive users OpenAI has.
• Don’t benchmark on one prompt. Test on five diverse tasks before committing.
• Don’t confuse aliases. gpt-5, gpt-5-mini, gpt-5-nano are different models at $1.25 / $0.25 / $0.05 per MTok input with different ceilings.
• Don’t quote input pricing while ignoring output — output is 8x input on every tier.
• Don’t run long-context, multi-step agentic, or vision-heavy prompts on mini; per-call savings evaporate across retries.

Verification checklist

1. Which model is loaded? Inspect the request payload or the Hermes config layer: cat ~/.hermes/config.yaml | grep -A2 'models:'. The default: and alias on the active provider must match the tier you intended.
2. How many tokens moved? Capture the response usage block (input_tokens, output_tokens, cache_read_input_tokens). On the OpenAI Responses API this is in every response; on Hermes it surfaces in the run log.
3. What did that call cost? Multiply: (input × tier.input_price) + (cached_input × tier.cached_price) + (output × tier.output_price). Pre-compute it so cost is visible before the next call, not after the bill.

Sources

• OpenAI GPT-5 model page
• OpenAI GPT-5 mini model page
• OpenAI business pricing
• ABS companion: OpenAI / Codex auth provider on Hermes

Last verified: 2026-07-15 against the OpenAI GPT-5 and GPT-5 mini model pages.