Full OpenAI API pricing for every GPT-5.5, GPT-5.4, and multimodal model in 2026. Per-token costs, cached input discounts, Batch API savings, and how to cut your bill further.
OpenAI API pricing changes fast. New models launch, old ones deprecate, and the cost difference between choosing the right model versus the wrong one can be thousands of dollars per month.
This is the complete breakdown for June 2026. Every flagship model, every multimodal option, cached input discounts, and the processing tiers that actually move the needle on your bill.
These are the core text generation models most teams use in production. All prices reflect standard processing rates for context lengths under 270K tokens.
| Model | Input (per 1M tokens) | Cached Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|---|
| GPT-5.5 | $5.00 | $0.50 | $30.00 |
| GPT-5.4 | $2.50 | $0.25 | $15.00 |
| GPT-5.4 mini | $0.75 | $0.075 | $4.50 |
A few things jump out here.
GPT-5.5 is OpenAI's frontier model, described as "a new class of intelligence for coding and professional work." At $5.00 per million input tokens and $30.00 per million output tokens, it is the most expensive flagship option. Use it when quality matters more than cost.
GPT-5.4 halves the price at $2.50/$15.00 for input/output. OpenAI positions it as "a more affordable model for coding and professional work." For most production workloads, this is the sweet spot.
GPT-5.4 mini is the budget workhorse. At $0.75 input and $4.50 output per million tokens, it handles coding, computer use, and subagent tasks at a fraction of the flagship cost. If your use case doesn't need frontier-level reasoning, start here.
Every flagship model offers a 90% discount on cached inputs. GPT-5.5 drops from $5.00 to $0.50 per million cached tokens. GPT-5.4 mini drops from $0.75 to $0.075.
OpenAI's built-in prompt caching works by reusing the static prefix of your prompt (system instructions, few-shot examples) across calls. If you structure your prompts so the unchanging parts come first, the model only computes the new portion.
This is significant for agent workflows where the system prompt stays constant across hundreds of turns. A 10,000-token system prompt that hits cache on GPT-5.4 costs $0.025 per thousand calls instead of $0.25. That is a 10x difference on every single request.
OpenAI's multimodal lineup covers voice, translation, transcription, and image generation.
The most capable model for realtime voice interactions. Pricing splits across three modalities:
| Modality | Input (per 1M tokens) | Cached Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|---|
| Audio | $32.00 | $0.40 | $64.00 |
| Text | $4.00 | $0.40 | $24.00 |
| Image | $5.00 | $0.50 | N/A |
Audio output at $64.00 per million tokens makes GPT-Realtime-2 the priciest model in the lineup. But cached audio input drops to $0.40, a 98.75% discount. If you are building voice applications with repeated context, caching is not optional. It is survival.
Live translation that keeps pace with the speaker: $0.034 per minute ($0.00057 per second). For a 30-minute translated call, that is roughly $1.02.
Streaming speech-to-text at $0.017 per minute ($0.00028 per second). Half the price of live translation.
State-of-the-art image generation with token-based pricing:
| Modality | Input (per 1M tokens) | Cached Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|---|
| Image | $8.00 | $2.00 | $30.00 |
| Text | $5.00 | $1.25 | N/A |
OpenAI now charges separately for built-in tools that extend model capabilities.
Web search costs $10.00 per 1,000 calls. The search content tokens themselves are free, so you only pay for the call, not the results. For applications that need grounding in current data, this is straightforward to budget: 10,000 search-augmented queries per day runs $100/day.
Containers (secure code execution environments) cost $0.03 for 1 GB up to $1.92 for 64 GB per container. Starting March 31, 2026, this shifts to per 20-minute session pricing at the same rates.
OpenAI offers three processing tiers beyond standard pricing:
50% off inputs and outputs. Tasks run asynchronously over a 24-hour window. The single biggest native discount OpenAI offers.
Reliable, high-speed performance with pay-as-you-go flexibility. Built for production workloads where latency matters.
Slower response times and occasional resource unavailability in exchange for lower costs. Ideal for development and testing.
If your workload tolerates latency (data processing, bulk classification, content generation queues), the Batch API is the most impactful optimization available. A GPT-5.4 mini batch job drops to $0.375 input and $2.25 output per million tokens.
Suppose you are running a customer support agent that processes 10,000 requests per day, averaging 500 input tokens and 300 output tokens per request.
Monthly volume: 300,000 requests. That is 150M input tokens and 90M output tokens.
That is a 6.7x difference between GPT-5.4 standard and GPT-5.4 mini batch, for the same 300,000 monthly requests. Model selection and processing tier selection matter more than any other optimization you can make.
OpenAI's cached input pricing and Batch API are powerful. But they still assume every request needs to hit the API.
The reality? A significant portion of production API calls are redundant. Research across production applications shows that 35-45% of prompts are either identical or semantically equivalent to previous requests. You are paying OpenAI to answer the same questions repeatedly.
This is where a semantic cache can change the math. It intercepts requests before they reach OpenAI, compares the semantic meaning of incoming prompts against cached responses, and serves matches instantly.
The difference between OpenAI's native prompt caching and semantic caching is important. Native caching reduces the cost of processing the static prefix of your prompt: you still pay for the completion. Semantic caching eliminates the API call entirely when the full response already exists in your cache. Those are fundamentally different cost profiles.
SHIM is a tenant-isolated gateway for native OpenAI Chat Completions and Responses APIs. It can apply policy, accounting, and PII placeholder processing while preserving documented OpenAI request payloads and success response formats; it does not add a response or semantic cache.
For the customer support agent example above, if 40% of those 300,000 monthly requests are cache hits, the math shifts dramatically. On GPT-5.4 mini standard pricing, you go from $517.50/month to roughly $310/month, and every cache hit responds in milliseconds instead of seconds.
OpenAI's own guidance: use large GPT models (GPT-5.5, GPT-5.4) for everyday tasks where quality matters, and mini models (GPT-5.4 mini) for simpler, speed-sensitive tasks. Reasoning models are best for complex, multi-step STEM problems.
The practical rule: start with GPT-5.4 mini. Measure quality. Only move up to GPT-5.4 or GPT-5.5 if the mini model's output doesn't meet your threshold. Most teams find GPT-5.4 mini handles 70-80% of their workloads just fine.
Those are direct OpenAI pricing examples. When using SHIM, query /v1/models and choose from the returned catalog; GPT-5.5 is not currently admitted by SHIM.
Stack your optimizations. Use the Batch API where latency tolerance allows. Structure prompts for cached input discounts. And put a semantic caching layer in front of the whole thing to eliminate calls that never needed to happen.
SHIM preserves prompt-caching fields on its supported Responses and Chat Completions routes without adding a semantic response cache. OpenAI's Batch API is outside SHIM's endpoint contract and must be called directly.
const openai = new OpenAI({
baseURL: "https://api.getshim.tech/v1",
apiKey: process.env.SHIM_API_KEY
})