OpenAI API Pricing 2026: Every Model, Every Rate, The Real Cost Per Million Tokens

What OpenAI charges in 2026

OpenAI runs eleven billable model families in July 2026, and the price spread between the cheapest and the most expensive is now 1,250x. You can run a fine-tuned gpt-5.4-nano for $0.20 per million input tokens, or you can spend $30 per million input on gpt-5.5-pro with long-context pricing on top. The decision is no longer “should I use OpenAI?” — it’s “which OpenAI model fits this specific workload without burning margin.” Every rate below is from the OpenAI dashboard on July 3, 2026.

If you’re a solo operator running a production agent, a freelancer billing clients for AI work, or a small agency owner trying to keep a 60% gross margin, this page is the rate card I wish I’d had at the start of 2026. I’ve worked three real bills below using these exact numbers so you can stop guessing and start budgeting.

The full OpenAI rate card (July 2026)

All rates per 1M tokens unless noted. Cached input is the rate when you re-use a prompt prefix within the cache TTL window — default 5 to 10 minutes, extendable to 24 hours on dedicated cache slots. “Long context” kicks in past the model’s threshold — 200K for most GPT-5.x models, 128K for the legacy GPT-4o line.

Image, audio, and embeddings

Pattern to notice: cached input is roughly 90% off raw input on every model that offers it. If you’re running a system prompt longer than ~2K tokens (and most production agents are), caching is the single biggest line item you’ll save on. gpt-5.4-mini cached at $0.075/M is now cheaper than Google’s Gemini 2.5 Flash raw input at $0.075/M — the caching frontier has moved.

Three real workloads with actual bills

Theoretical rate cards lie. Here are three production patterns with the math worked at the actual rates, using a six-week month and weekly volume figures you’d see at a 5-person agency or solo SaaS in 2026.

Workload A — 1M-token customer support analysis

You’re classifying 50,000 support tickets per week at ~20K tokens each (10K transcript + 10K system prompt + small KB snippets). That’s 1 billion input tokens per week, with roughly 60M tokens of generated output (1,200 tokens per ticket on average — tag, sentiment, routing, draft reply). The system prompt and KB are identical across calls, so they’re perfect caching candidates.

• Uncached on gpt-5.4-mini: $0.75/M × 1,000 + $4.50 × 60 = $750 + $270 = $1,020/week. That’s $5,100/month before any reply generation.
• With prompt caching on gpt-5.4-mini: roughly 95% of input hits cache at $0.075/M; 5% is fresh at $0.75/M. Cost drops to $133/week, $665/month. A 7.6x reduction.
• Same workload on gpt-5.5 (uncached): $5 × 1,000 + $30 × 60 = $6,800/week, $27,200/month. Same workload, 53x more expensive than cached mini.
• Same workload on gpt-5.5 with caching: $0.50 × 950 + $5 × 50 + $30 × 60 = $475 + $250 + $1,800 = $2,525/week. Still 19x the cached mini bill, but a 2.7x reduction from uncached.

The takeaway is brutal: the model choice matters less than whether you cached. Cached gpt-5.5 at $2,525/week undercuts uncached gpt-5.4 at roughly $5,200/week on the same workload.

Workload B — 500K-token code review

PR review agent reads the full repo context (around 500K input tokens including dependencies) and returns ~3K tokens of inline comments. Five reviews per day, six days a week — a typical boutique dev agency pattern.

• On gpt-5.4 (uncached): $2.50 × 0.5M × 30 reviews + $15 × 3K × 30 = input $37,500 + output $1,350 = $38,850/month. That’s dev-salary-burn.
• With caching (90% of repo is a fixed prefix that changes rarely): input drops to roughly $6,500 (cached at $0.25/M for 450K tokens, fresh at $2.50/M for 50K tokens), output unchanged. Total $7,850/month. Caching pays for itself on week one and converts a mid-five-figure bill into a manageable one.
• On gpt-5-mini with caching: Half the cached bill, around $4,000/month. Quality drop is real but tolerable for lint-style reviews and PR summary generation.
• On gpt-5.5 (uncached) with tool-use tuning: input $5 × 0.5M × 30 = $75,000/month. Don’t do this. Use Sonnet 4.6 instead — see the Anthropic article.

Workload C — 10K image generations

A boutique agency renders 10,000 product photos at 1024×1024 in gpt-image-1.5 medium quality, plus 5K input tokens of caption and brand context per image.

• Medium tier ($5/$10): 10,000 × 5K input = 50M tokens = $0.25; 10,000 images × ~$0.02 (1,290 output tokens each at $10/M) = $200. Total ≈ $200.
• High tier ($8/$32): input $0.40, output ≈ $800. Total ≈ $800.
• Replacing with Imagen 3 via Google at $0.039/image: 10,000 × $0.039 = $390. Worth modeling if quality holds for product photography.
• Cached caption context: if your caption is templated, the cached-input rate ($1.25/M) slashes the input leg to ~$0.06. Negligible per image but meaningful at 100K+ volumes.

Caching, batching, and the 10% data-residency surcharge

Three knobs move your bill more than the model choice does.

Prompt caching. The biggest lever in the toolkit. Cached input is $0.50/M on gpt-5.5 (a 90% discount), $0.25/M on gpt-5.4, $0.02/M on gpt-5.4-nano. Cache writes themselves are billed at the cached-input rate — not free. Cache TTL is 5 to 10 minutes by default; extend it explicitly with cache_control breakpoints if your agent idles between calls. For long system prompts (10K+ tokens) and busy production agents, the math on caching is essentially free money.

Batch API. Submit jobs asynchronously and you get ~50% off input and output. SLA is 24 hours (or 6 hours for a smaller discount of ~25%). Use it for any backfill, re-embed, weekly report generation, or offline dataset processing that doesn’t need to be live. If your agent has a nightly summarisation step, that step belongs on batch.

Regional surcharges. If you require EU data residency on gpt-5.5, the long-context input rate climbs from $10/M to roughly $11/M — a 10% uplift baked into the EU region endpoint. This is rarely visible in marketing pages and is buried in the dashboard. Build it into your margin if you sell into EU regulated verticals (health, finance, public sector). US-only workloads pay no surcharge.

Fine-tuning savings. Fine-tuned gpt-5.4-mini models still bill at base rates, but you can usually drop gpt-5 -> gpt-5.4-mini if you fine-tune for your task. A 4x rate cut is common.

When OpenAI wins, when it loses

Use OpenAI when:

• You need multimodal in one model — vision, tool use, code, JSON, plus image gen via gpt-image-1.5 in a single workflow.
• You’re running structured agent frameworks (LangChain, CrewAI, the OpenAI Agents SDK successor) — the tooling around the API is still the deepest in the industry.
• Real-time voice matters and you want gpt-realtime-2 at $0.034/min translation, $0.017/min whisper.
• Your workload fits inside 200K context and you actually cache — the rate-card math is unbeatable.
• You need fine-tuning on a small/medium tier model and care about predictable latency.

Avoid OpenAI when:

• You’re running very high-volume classification. Google Flash at $0.075/M is 10x cheaper than gpt-5.4-mini cached input on raw volume.
• You need guaranteed long-context quality past 200K. Claude Opus 4.7 still holds the quality crown for huge document analysis (see the Anthropic page).
• You’re locked into a strict EU data-residency contract — the 10% uplift and selector variability start to hurt at scale.
• You’re burning tokens on tiny tasks. Use gpt-5.4-nano at $0.20/M for classification, JSON reformats, router calls, anything under 500 output tokens.
• You’re a solo founder with low volume and need a free tier — Google AI Studio is still free; OpenAI is paid-only since November 2025.

FAQ

How much does GPT-5.5 cost per million tokens in 2026?

$5.00 per 1M input tokens and $30.00 per 1M output tokens. Cached input is $0.50/M (90% off). Long-context (above 200K) is $10/M in and $45/M out. gpt-5.5-pro jumps to $30/$180 with a $60/$270 long-context rate — that’s the top of the OpenAI stack in 2026.

Is GPT-5 API still available in 2026?

Yes — gpt-5 is kept as the legacy stable model at $1.25/$10, which is roughly 75% cheaper than gpt-5.5. Many production teams stay on it because pricing is predictable, quality is sufficient for extraction and classification, and deprecation is unlikely through 2027.

How much does OpenAI image generation cost?

gpt-image-1.5 at medium quality works out to roughly $0.02–$0.06 per 1024×1024 image (billed as $5 input + $10 output per million tokens; one image costs ~1,290 tokens of output). High-quality renders land at $0.08–$0.19 each. Low quality is roughly half the medium price.

How much does gpt-realtime-2 cost?

$32/M input tokens, $0.40/M cached input, $64/M output. Audio minutes work out to $0.034/min for translation and $0.017/min for the whisper endpoint. A 30-minute customer call at heavy token usage costs roughly $0.40–$1.20 of API.

What is the cheapest OpenAI embedding model in 2026?

text-embedding-3-small at $0.02 per 1M tokens. For most retrieval pipelines this is the right pick. text-embedding-3-large at $0.13/M is worth the upgrade only if recall on tiny chunks is the bottleneck. The legacy ada-002 is $0.10/M and rarely the right choice — 3-small beats it on quality and price.

Does OpenAI still offer a free tier?

The playground credits remain for evaluation, but for production workloads OpenAI is paid-only since the November 2025 cutover. Google is still the only vendor with a free production tier in July 2026.

When should I pick GPT-5 over GPT-5.5?

When the workload is extraction, classification, or structured JSON and you don’t need the 8–12% quality bump on nuanced reasoning that 5.5 brings. The 4x price gap ($1.25 vs $5 input) compounds fast. Use 5.5 only for tasks where the quality gap translates to real revenue or fewer human reviews.

Verdict

OpenAI is the best default for multimodal agentic workloads in 2026, but only if you turn on prompt caching. On gpt-5.4-mini, caching cut a $1,020/week support classification bill to $133/week — that’s the move. For pure-text high-volume work, gpt-5.4-nano at $0.20/M is the new floor and it undercuts most of the open-source-hosting math once you account for engineering time. Spend the savings on a smaller moat: better evals, a longer context window, or a redundancy path to Claude for the workloads where OpenAI’s quality is genuinely worth 6x the bill.

If you have to pick one OpenAI model for an unknown new workload in July 2026: start on gpt-5.4 with caching, route 80% of low-stakes calls down to gpt-5.4-mini, reserve gpt-5.5 for the reasoning-heavy 20%, and never touch gpt-5.5-pro unless you have measured it paying for itself.

Updated July 3, 2026. Reviewed against the OpenAI pricing dashboard. Next review: August 1, 2026.

More from