Codex cost 2026 — actual token-based costs and the break-even point

Codex has two pricing models, and almost every article covers only one of them. The subscription model is a monthly ChatGPT plan, with Codex included and usage windows as the trade-off; the API key model has no monthly fee, charges per token, and has no usage window. Most “Codex pricing” pages online are plan comparison charts — this page covers the other model: what one step actually costs, where the break-even point between the two options lies, and how to finish your work on the day your plan allowance runs out.

As of October 3, 2026, Kunavo’s gpt-5-6-sol costs $2.00 per 1M input tokens / $12.00 per 1M output tokens (official OpenAI pricing: $5.00 / $30.00; currently offered at the promotional price of $4.00 / $20.00, which the official pricing page states will last at least until November 21, 2026). Rates are sourced from the catalog on this site, not entered manually, so this page won’t disagree with what the API actually charges.

What one step actually costs

Plan comparison charts can’t answer “How much will this cost to run for a month?” To answer that, we need a unit. All our pricing pages use the same one: one agentic step — one back-and-forth with the model during a task, about 25,000 input and 1,200 output tokens (system prompt + conversation + file context going in, one edit or explanation coming out).

To calculate it with your own token counts, use this section:

# 一個 Codex 工作階段的實際花費，從它自己的 token 數算出來。
IN_RATE, OUT_RATE = 2.00, 12.00  # 每 1M token USD

def session_cost(inp: int, out: int) -> float:
    # 注意：reasoning token 不會出現在回覆裡，但會用「輸出」費率計費。
    return inp / 1_000_000 * IN_RATE + out / 1_000_000 * OUT_RATE

print(session_cost(25_000, 1_200))      # 一個 agentic step
print(session_cost(25_000, 1_200) * 20) # 一件 20 步的任務

Break-even point — one division

The break-even point depends on how many steps you run, not how many days you work. At $0.064 per step:

• $20 per month ≈ 310 steps ≈ 16 20-step tasks
• Below that volume → pay-as-you-go is cheaper, and months with no work cost $0
• Consistently above that volume → a subscription is a better deal until you hit the plan’s usage limit

Teams almost always use keys, not because of the unit price but because of attribution: keys can be assigned by person and purpose, and hard limits can be set. Seat-based monthly plans can do neither.

How to finish your work on the day your allowance runs out

This is the question you run into a week after reading a plan comparison chart, but no article answers it. When the usage window closes, you don’t have to wait for it to reset — point Codex CLI to a pay-as-you-go endpoint.

# 把 Codex CLI 指向按量計費的端點。
# Codex 只吃 Responses API，chat completions 不會動。
export OPENAI_BASE_URL=https://api.kunavo.com/v1
export OPENAI_API_KEY=sk-kn-...

codex

After setting the authentication environment variables, Codex uses that key instead of your login. The plan’s usage limits no longer apply, and the usage is charged to the key’s owner; remove the variables to return to the subscription.

One hard limitation to know: Codex CLI’s custom providers only support the Responses API, so the endpoint must provide POST /v1/responses. Services that implement only chat completions won’t connect. Kunavo provides both; see Quickstart for details.

Two things that can push the actual cost higher

Reasoning tokens. Reasoning models generate thinking tokens before giving an answer. They don’t appear in the response, but are billed at the output rate. The table above doesn’t include them, so the actual bill for a hard problem will be higher than the table shows, with a larger gap for more difficult problems. The output token count in the usage field is the actual figure.

Retries. Requests that consume tokens before failing downstream still cost money. The difference is most noticeable with agentic tools, because what users see as one action is many back-and-forth calls for billing purposes. (Requests that fail on the Kunavo side and never reach the upstream provider aren’t charged.)

The most effective approach is to choose a model suited to the task: gpt-5-6-terra ($0.70 / $4.20) is enough for tasks such as classification, extraction, and formatting; you don’t need to use the top tier every time.

Codex or Claude Code

A comparison only makes sense when you use the same unit. The cost of the same agentic step across Claude tiers is listed on the Claude Code pricing page, using the same token counts for a direct comparison. Overall Claude pricing (subscriptions, API, and Claude Code) is on the Claude pricing page.

The real difference is usually not the unit price but the number of steps — the number of back-and-forth calls each tool takes to complete the same task varies with the task type, not the rate card. So take the task you do most often, run it once with each, note the usage, then multiply by the rates above.

Paying from Taiwan

Both routes use an international credit card; Taiwan has no local payment channel — JKoPay and LINE Pay are not on the available list. The difference is timing: a subscription charges once per month, incurring an international transaction fee each time; Kunavo's prepaid system charges the card only once when you top up, with a minimum of $10; the balance does not expire, and subsequent charges are deducted according to usage. Failed requests are not charged. Supported options include cards with JCB (Visa, Mastercard, American Express, JCB, UnionPay), Apple Pay, and Google Pay. Billing details are in the billing documentation.

Frequently asked questions