GPT-5 API pricing

By the aitokensai editors · Verified 2026-11-14 Sources: OpenAI official pricing

OpenAI's GPT-5 tier. Available at listed at platform rate (no published official USD price).

USD, standard tier. Rate read from the live rate config on 2026-11-14. OpenAI publishes no USD list price for this model, so nothing here is cross-checked against one — this is a gateway rate, not a discount.

Specifications

What GPT-5 costs at your volume

Calculated at $0.625 input / $5 output per 1M tokens, 8.0× apart on this model. The 5:1 input-to-output assumption in this table is generous, responses longer than that cost more per response than the row suggests, so treat these as floors rather than estimates. Put your own split through the cost calculator before committing a budget.

Reading GPT-5's numbers correctly

$0.625 per million input tokens looks like the headline, and it is the one people compare. It is also the smaller half of the bill. Output here runs at 8.0× input, so on any interactive workload the generation side decides what you actually pay. Comparing this model against another on input alone will point you the wrong way.

This is a reasonable rate for agent loops and interactive editing, where the turn count is high but nobody in particular resends a huge context. It is a poor fit for anything that resends a large context every turn, because at $0.625 the input side stops being negligible.

Where GPT-5 sits in the index

GPT-5 is 119 of 231 on input rate and 165 of 231 on output rate. A gap of 46 places, because on this model the output rate is 8.0× the input rate, steeper than the index median of 4.0× An output-heavy workload therefore pushes it 46 places down the table, which is why the cheapest model per input token is routinely not the cheapest per task. Percentile is the share of the 231 models in this index that cost more than GPT-5.

On an agent mix the lowest blended rate in the same index is Gemma2 9B It at $0.01 per 1M tokens, about 390.6× less than this model at the same ratio. Across the whole index, sorting by input price alone moves a model by as much as 77 places; on GPT-5 specifically the shift is 46 — which is what the re-ranked index exists to show.

GPT-5 is 119 of 231 across every model in the index; the comparison that usually settles a choice is OpenAI's own lineup further down.

GPT-5 among OpenAI's 12 models

GPT-5 is the 6th cheapest of OpenAI's 12 models in this index at $0.625 per 1M input tokens. The range inside that group runs from GPT-5 nano at $0.025 to GPT-6 Astra at $5 (a spread of 200×, and this one sits at 25).0× the group's cheapest tier.

On input rate alone, stepping up to GPT-5.6 Terra costs 1.6× the input rate and dropping to GPT-6.1 Terra cuts it by 20%. Both measured at 2026-11-14, the same check as every other rate on this page.

Full OpenAI range, including models outside this shortlist: OpenAI rate index.

Similarly priced alternatives

Blended comparison (input + output per 1M). Price is one axis, these are the three closest to GPT-5 in blended rate, not interchangeable substitutes for it.

Calling GPT-5

The identifier you send is gpt-5. Copy it exactly, because a mismatch returns a model error rather than a connection error, which sends most people to debug the wrong thing. Everything else is a base URL argument, GPT-5's own SDKs work unchanged.

from openai import OpenAI

client = OpenAI(base_url="https://aicomp.ai/v1", api_key="sk-...")

resp = client.chat.completions.create(
    model="gpt-5",
    messages=[
        {"role": "system", "content": "Answer in one short paragraph."},
        {"role": "user", "content": "Explain what changed in this release."},
    ],
    max_tokens=300,
)

# prompt_tokens / completion_tokens 是真实账单的唯一依据 —— 记下来，别猜
print(resp.usage.prompt_tokens, resp.usage.completion_tokens)

Log the usage object on every call, not just while you are optimising. On GPT-5 output costs 8.0× what input costs, so the same monthly token count bills very differently depending on how much of it comes back as output — measure your own split before extrapolating from the headline input rate, because rates move with vendor promotions.

Setting it up

Pointing an existing integration at GPT-5 is a base URL change, not a rewrite. Pick your tool:

• Claude Code — set the base URL via environment variables
• Cursor — override the API endpoint in settings
• OpenAI / Anthropic SDK — one base_url argument
• Python (sync), streaming and async examples
• Cutting the bill — eight levers ranked by effect

Where GPT-5 bills from: the output half

Quoting a monthly token figure tells you very little about GPT-5. Output here costs 8.0× what input does, above the index median of 4.0× across 231 models, so two deployments moving the same number of tokens can differ by several times depending purely on how much the model writes back. Generation is the bill; everything else is rounding. The five shapes below are the ones GPT-5 actually gets used for, priced out at each one's real ratio.

The cheapest comparison is GPT-5 nano at $0.025 / $0.2 per 1M — 25.0× cheaper than GPT-5 on a blended rate. Treat it as a rate floor, not a recommendation: it says nothing about whether GPT-5 nano can do the work you need GPT-5 for.

The spread between the first and last row above is not caused by doing more work — it is caused by asking GPT-5 to write more. At 8.0×, the cheapest intervention available to you is shortening responses: a tighter system prompt, a lower max_tokens, a shorter answer format. Optimising the input side barely registers against your $5 per 1M output rate.

FAQ

Get GPT-5 access