GPT-5 Cost Calculator (2026)

Every GPT-5 call uses the same per-token math. No platform fee, no per-call fee, no minimum invoice. You pay for tokens in and tokens out, at the chosen model's per-1M rate:

``` cost = (input_tokens / 1,000,000) × input_price_per_M + (output_tokens / 1,000,000) × output_price_per_M ```

Two adjustments stack on top. Prompt-cache hits — portions of your input prefix that OpenAI cached because you sent them recently — bill at the cached-input rate (10% of the standard input price across every GPT-5 tier). Long stable system prompts and reused tool schemas are the typical winners. The Batch API takes a flat 50% off both input and output for asynchronous jobs delivered within 24 hours. The discounts compose: a cached + batched GPT-5.5 call pays $0.50/1M cached input divided by 2 = $0.25/1M on the cached portion, and $30/1M output divided by 2 = $15/1M on output.

On GPT-5.5 Pro, reasoning tokens generated internally before the visible response bill at the $180/1M output rate, the same as the answer text. A query that triggers 3,000 reasoning tokens to produce a 500-token answer bills 3,500 output tokens. Budget for a 3-8x reasoning multiplier on Pro if the task is non-trivial. Standard GPT-5.5 and GPT-5.4 do not surface chain-of-thought; their output bill matches the response length.

A representative call — a 1,000-token prompt returning a 500-token answer, roughly a 750-word brief in and a 375-word reply out. At standard rates across the GPT-5 family:

GPT-5.5 Pro: (1000 / 1,000,000) × $30.00 + (500 / 1,000,000) × $180.00 = $0.030 + $0.090 = **$0.120 per call**.

GPT-5.5: 0.001 × $5.00 + 0.0005 × $30.00 = $0.005 + $0.015 = **$0.020 per call**.

GPT-5.4: 0.001 × $2.50 + 0.0005 × $15.00 = $0.0025 + $0.0075 = **$0.010 per call**.

GPT-5.4-mini: 0.001 × $0.50 + 0.0005 × $1.50 = $0.0005 + $0.00075 = **$0.00125 per call**.

A 96x spread between GPT-5.4-mini and GPT-5.5 Pro on identical token volume. The right model is rarely the most expensive one in the family — it is the cheapest GPT-5 tier that passes your held-out eval on the actual task. Most teams default to GPT-5.5 out of caution; in our experience 60-70% of that traffic would survive a move to GPT-5.4 or GPT-5.4-mini with no perceptible quality drop.

Multiply the per-call numbers by 100,000 — a realistic mid-size workload (daily classification on 3,000 records, weekly summarization runs, a low-volume internal agent):

GPT-5.5 Pro: $12,000/month. GPT-5.5: $2,000. GPT-5.4: $1,000. GPT-5.4-mini: $125.

Apply the Batch API discount to GPT-5.4 for any portion that does not need synchronous delivery (nightly summarization, weekly digests, eval runs): the GPT-5.4 row drops from $1,000 to $500 on the batched portion. Add prompt caching where 800 of every 1,000 input tokens are a stable system + tool prefix hitting cache 80% of the time: those 640 cached tokens drop from $2.50/1M to $0.25/1M — saving roughly 90% on 64% of input volume, or ~$144 off the monthly input bill.

Stack both — the same workload runs around $400/month on GPT-5.4 at 100k calls, a 60% reduction versus standard rates. The lesson generalizes: on GPT-5, the model choice sets the ceiling, but cache structure and batch eligibility set what you actually pay. Teams that pick GPT-5.5 and ignore caching often pay more than teams that pick GPT-5.5 Pro and structure prompts for cache hits.

Now scale to 1M calls per month — production scale for a SaaS app with 30,000 active users running roughly 33 GPT-5 calls each, or a single-product team running per-record automation at high volume:

GPT-5.5 Pro: **$120,000/month**. GPT-5.5: **$20,000**. GPT-5.4: **$10,000**. GPT-5.4-mini: **$1,250**.

The Batch + cache stack on GPT-5.5 takes that $20,000 to roughly $8,300/month — 58% off — on the same input/output mix. On GPT-5.4-mini, the same stack lands at around $500/month, which is $0.0005 per call at scale, an order of magnitude cheaper than most companies budget for AI features in their first planning round.

The canonical lever order for scaling cost down on GPT-5: (1) run an eval to find the cheapest tier in the family that hits quality, (2) batch every asynchronous workload for 50% off, (3) restructure prompts so the cacheable prefix is stable across calls, (4) cap output length where you control the consumption shape. Most teams reverse the order — they tune output last when output is 6x the input price on every GPT-5 tier.

Agent loops are the worst-case cost shape on GPT-5. The model takes multiple turns per user query, replaying the full transcript each turn. A typical 5-turn loop with a 2,000-token system + tools prefix and 800-token context growth per turn:

Turn 1: 2,800 in / 200 out. Turn 2: 3,000 in / 200 out. Turn 3: 3,200 in / 200 out. Turn 4: 3,400 in / 200 out. Turn 5: 3,600 in / 200 out. Total: 16,000 input + 1,000 output. On GPT-5.5: 0.016 × $5 + 0.001 × $30 = $0.080 + $0.030 = **$0.11 per user query** — about 5.5x a single call.

Now apply caching. The 2,000-token system + tools prefix is stable across all 5 turns. If cache hits roughly 80% of those 2,000 tokens × 5 turns = 8,000 cached input tokens, those drop from $5/1M to $0.50/1M: $0.040 → $0.004, saving $0.036 per query (33% off the bill). For 100k queries/month: $11,000 → $7,400.

On GPT-5.5 Pro, the same agent loop hits **$0.66 per query** at standard rates — driven mostly by the $180/1M output rate against ~1,000 visible output tokens plus reasoning. Moving the loop to GPT-5.4 ($2.50 / $15) with cache cuts the per-query cost to roughly $0.035 — a 19x improvement over Pro for most agentic workloads that do not require Pro's reasoning depth. Build cache-anchored GPT-5 agent prompts for free with our code prompt builder.

**GPT-5.5 Pro ($30 / $180)** is for tasks where one wrong answer costs more than 100 right ones. Multi-step financial analysis, legal drafting, complex code synthesis with strict correctness gates, scientific reasoning. Pro generates extensive internal reasoning chains before producing its final answer; you pay $180/1M for those tokens even though they don't appear in the response. Justify Pro only when downstream cost-of-error dominates per-call cost.

**GPT-5.5 ($5 / $30)** is the default for general-purpose GPT-5 work: agentic workflows, content generation that ships to humans, complex chat, anything you would have used GPT-4 or early GPT-5 generations for. Substantially higher capability than 2024-era GPT-4 at a fraction of the price. If you are not sure which tier to start with, start here and downshift after eval.

**GPT-5.4 ($2.50 / $15)** is the sweet spot for high-quality structured tasks at scale: summarization with strict format adherence, multi-step extraction, complex classification, RAG synthesis. Most production teams running between 100k and 1M GPT-5 calls per month live on GPT-5.4 — half the price of GPT-5.5 with very small quality deltas on well-scoped tasks.

**GPT-5.4-mini ($0.50 / $1.50)** is for high-volume embedded tasks: simple classification, intent detection, routing, internal telemetry, simple chat where the user expects something fast and lightweight. The 10x price gap from GPT-5.5 makes it viable for use cases that wouldn't survive a $0.02 per-call cost — autocomplete suggestions, per-keystroke intent routing, real-time moderation.

GPT-5 ships three capabilities that distinguish it from the GPT-4 generation. Each has a real cost shape worth understanding before you wire it into a product.

**1M-token context window** on GPT-5.5 and GPT-5.5 Pro. This unlocks single-call analysis of large documents (full books, codebases, transcripts). The cost: at GPT-5.5's $5/1M input rate, filling the full context costs $5 per call before you get a single output token. A 500-page legal contract at roughly 300k tokens hits $1.50/call on input alone. Worth it when the alternative is a 20-call RAG pipeline; brutal if you don't actually need the full context. Most teams should chunk + retrieve before reaching for 1M context.

**Native vision** on every GPT-5 tier. Images are tokenized at roughly 85 tokens per low-detail tile and 170 per high-detail tile, with a base 85 tokens per image. A typical 1024x1024 image at high detail costs around 1,275 input tokens — about $0.006 on GPT-5.5 or $0.0006 on GPT-5.4-mini. Image-heavy workloads (UI testing, document parsing, visual QA) bill primarily on input.

**Real-time mode** for streaming voice and video. Real-time API calls bill on both audio input tokens and audio output tokens at distinct rates published on the live pricing page. The economics shift dramatically — audio output dominates the bill on conversational use cases. If you're building real-time voice on GPT-5, model 1 minute of conversation as roughly 2,000 audio tokens in + 2,000 audio tokens out per minute, and look up the audio-specific rates separately.

**Structured outputs** (JSON schema guarantee) and **tool calling** are included at standard token rates — there is no per-feature surcharge. Tool definitions bill as input tokens every time they are sent; cache them as part of your stable system prefix and they drop to the 10% cached-input rate.

Different tasks have different cost shapes on GPT-5. Knowing the shape changes the model you should pick.

**Writing (blog posts, marketing copy, drafts)**: typical 500-1,500 token input, 1,500-3,000 token output. Output-dominant. On GPT-5.5 a 1,000-in / 2,500-out draft costs $0.005 + $0.075 = $0.080/call. On GPT-5.4 the same draft costs $0.0025 + $0.0375 = $0.040/call. Output cap matters most here — running without a `max_tokens` ceiling on writing tasks burns 30-50% extra on tokens you'll trim in editing.

**Coding (refactor, generate, review)**: typical 2,000-10,000 token input (file contents, context), 500-2,000 token output. Input-dominant for context-heavy refactors. On GPT-5.5 a 5,000-in / 1,500-out code task costs $0.025 + $0.045 = $0.070/call. Caching the project's stable file headers and the system prompt cuts that to roughly $0.030/call on a warm cache. GPT-5.5 Pro is only justified when the task requires architectural reasoning across many files.

**Reasoning (math, multi-step analysis, complex planning)**: this is GPT-5.5 Pro's home. Typical 1,000-token prompt with 500-token visible answer but 2,000-5,000 reasoning tokens hidden in output. On GPT-5.5 Pro: $0.030 input + $180/1M × 5,500 = $0.030 + $0.99 = roughly $1.02/call. For comparison, the same task on GPT-5.5 might cost $0.020 and produce a worse answer; the question is whether the answer quality difference is worth 50x the cost. Often it is for one-off high-stakes work, rarely for scaled inference. For comparison against o-series reasoning models that price reasoning tokens at lower output rates, see o1 reasoning cost.

**Agent loops (multi-turn tool use)**: examined in worked example 4 above. 4-8x a single call on a 5-turn loop, dropping to 2-3x with aggressive caching. The biggest win is keeping the system prompt + tool definitions stable across the loop so cache holds; the second biggest win is summarizing turns past turn 5 into a compact recap rather than replaying the full transcript.

The Batch API takes 50% off both GPT-5 input and output for jobs delivered within 24 hours. It accepts a JSONL file of requests, returns a job ID, and webhooks or polls to completion. No quality difference, no behavior difference — same models, same outputs, half the price.

Workloads that are textbook Batch wins on GPT-5: nightly content generation, bulk summarization (newsletters, weekly digests), classification of yesterday's records, eval and regression test runs, embedding precompute (use embedding endpoints), training-set generation, scheduled report drafting. Any output that lands in a dashboard, email, or CSV consumed asynchronously is a Batch candidate.

Workloads that cannot use Batch: synchronous chat (user waiting), real-time agent loops, anything inside a request handler that returns to the user, anything with sub-minute SLA. About 30-60% of typical production GPT-5 traffic can move to Batch with no UX change.

The compounding effect: a $20,000/month GPT-5.5 bill with 50% of the workload Batch-eligible drops to $15,000/month — $5,000/month saved on a single configuration change with zero quality impact. For most teams this is the single highest-EV optimization they can make on their GPT-5 spend.

Cached input on GPT-5 bills at exactly 10% of the standard input rate: $0.50/1M on GPT-5.5 (vs $5), $0.25/1M on GPT-5.4 (vs $2.50), $3.00/1M on GPT-5.5 Pro (vs $30), $0.05/1M on GPT-5.4-mini (vs $0.50). The cache is opportunistic — OpenAI computes a fingerprint of your input prefix and caches it server-side. Subsequent calls with the same prefix read from cache.

The hard rule: caching is a **prefix match**, not a substring match. Anything you want cached must come at the start of your message array. Stable system prompt, tool definitions, and reusable few-shot examples go first. User-specific content and dynamic context go last. A 1,500-token cached prefix on GPT-5.5 drops from $5/1M to $0.50/1M — saving $0.0068 per call. At 1M calls per month, that is $6,800 saved with one structural change.

Most LLM SDKs do not require code changes to opt in to caching on GPT-5; the cache activates automatically once you structure your prompts prefix-first. The single biggest mistake we see in audits: teams interpolate dynamic data (current date, user ID, session state, retrieved RAG chunks) into the system prompt, which breaks every cache hit. Move that to a user message and the cache holds across calls.

Cache TTL on GPT-5 is typically minutes (not hours), so traffic patterns matter. A workload with sustained calls every few seconds caches reliably; a workload with one call every 20 minutes mostly cache-misses. If your traffic is bursty, look at warming the cache with a synthetic call at the start of each session — the marginal $0.005 to warm a 1,500-token prefix saves multiples of that across the next 50 user calls.

OpenAI runs two completely separate billing tracks. The **GPT-5 API** (priced per token, accessed at platform.openai.com) is for developers building applications. The **ChatGPT consumer subscription** (Free, Plus $20/mo, Pro $200/mo, Team, Enterprise) gives end-users access to GPT-5 in the ChatGPT UI. Same models underneath, distinct billing.

A $20/month ChatGPT Plus subscription does **not** include any GPT-5 API credit. If you are building on the API, set up API billing independently at platform.openai.com and add a payment method to your API account. The two billing relationships use your OpenAI identity but track usage, payment methods, billing limits, and tier promotions independently.

A $200/month ChatGPT Pro subscription includes GPT-5.5 Pro in the ChatGPT UI with effectively unlimited use, but it gives you zero API access to GPT-5.5 Pro. If your team needs programmatic GPT-5.5 Pro access, you pay $30/$180 per 1M tokens on the API regardless of any ChatGPT Pro subs you hold.

What this means: budget for two separate line items if your team uses both. A 5-person team with ChatGPT Plus seats ($100/month) plus a GPT-5 API bill is not double-paying — the seats fund interactive use and the API funds production traffic. See our ChatGPT cost guide for the consumer tier breakdown.

**Mistake 1: defaulting every workload to GPT-5.5.** Most production traffic is classification, summarization, or short-form generation that GPT-5.4 or GPT-5.4-mini handles at 1/5th or 1/40th the price with quality indistinguishable on a real eval. The fix: build a 50-example held-out eval per task, run all four GPT-5 tiers, pick the cheapest tier that passes. This single exercise typically cuts GPT-5 bills by 40-70%.

**Mistake 2: huge system prompts that never cache.** If your system prompt interpolates anything that changes between calls (timestamps, user names, context summaries, retrieved chunks), the cache never hits and you pay full input rate every time. The fix: rewrite so the system prompt is static across users and sessions; move all dynamic content to user messages. A static 2,000-token system prompt that hits cache 90% of the time saves ~$8 per 1,000 calls on GPT-5.5.

**Mistake 3: no `max_tokens` cap.** A 300-token answer that returns 1,500 tokens because you forgot a ceiling costs 5x on output. On GPT-5.5 Pro that is $0.27 vs $0.054 per call. The fix: cap output everywhere you control the consumption shape; let it run uncapped only where genuinely needed.

**Mistake 4: replaying full chat history every turn.** Past turn 5, summarize the prior conversation into a compact 200-token recap and replay that instead of the full transcript. Saves 50-80% on input across long sessions with no perceptible quality loss.

**Mistake 5: using GPT-5.5 Pro for everything that 'feels important'.** Pro's 6x premium over GPT-5.5 is only justified when downstream cost-of-error exceeds the per-call premium. For most generative work, GPT-5.5 plus a sanity-check pass on GPT-5.4 is cheaper and more reliable than Pro alone. Build evals; let cost follow data, not vibes.

**Mistake 6: ignoring Batch eligibility.** 30-60% of typical production GPT-5 traffic can move to Batch with no UX change for 50% off. Audit your traffic for any output that lands in a dashboard, email, or report consumed asynchronously — that is a Batch candidate.

Every price on this page comes from OpenAI's live API pricing page at developers.openai.com/api/docs/pricing, fetched 2026-06-20 and verified against three independent corroborating sources (recent integration commits in popular open-source projects, community pricing aggregators, the public OpenAI cookbook). Where a number could not be verified against the official page it was omitted — we'd rather ship a guide missing a row than ship a guide with a fabricated number.

OpenAI does not version their pricing page with explicit changelog entries. They push changes silently. Since the GPT-5 family launched, we have seen two price moves: a 20% input-rate reduction on GPT-5.4-mini at launch maturity, and a tightening of the cached-input ratio to a clean 10%. Expect quarterly motion on at least one tier in the family.

**How to verify before you budget**: open developers.openai.com/api/docs/pricing in an incognito window (no logged-in session interfering with rendering), copy the four GPT-5 rows into a spreadsheet, compare against the table above. If they match, this guide is current for your purposes. If they don't, trust the live page. Re-verify quarterly if your GPT-5 bill is over $1,000/month — at that volume a single price move shifts the budget materially.

**Reproducible methodology**: every dollar in the table and every worked example traces to the four published prices above. No row was synthesized from 'plausible' rates. If you find a discrepancy with the live page, treat the live page as canonical and tell us — we re-fetch and update. Sibling drill-downs: Claude API cost for Anthropic-side comparison, DeepSeek cost for open-source alternatives at fractional GPT-5 prices.