GPT-Realtime-2 Pricing: Full Cost Breakdown for a 5-Minute Call

The Announcement, Plain English

On May 7, 2026, OpenAI published the pricing for GPT-Realtime-2:

• Audio input: $32 per 1M tokens
• Audio output: $64 per 1M tokens
• Cached input: $0.40 per 1M tokens
• Context window: 128K tokens
• Max output: 32K tokens

The cached-input number is the line that quietly moves the math. This post walks through what it actually costs to run a typical 5-minute voice call, where the dollars go, and exactly when caching pays for itself.

The 5-Minute Call, Token By Token

A representative healthcare-style voice call. Numbers are rounded to make the math legible:

• System prompt: 6,000 tokens (one time, cacheable)
• Tool schemas: 3,000 tokens (one time, cacheable)
• Per-turn user audio in: avg 150 tokens/turn x 30 turns = 4,500 tokens
• Per-turn assistant audio out: avg 200 tokens/turn x 30 turns = 6,000 tokens
• Function call I/O: ~2,000 tokens across the call

Total non-cached audio in: ~6,500 tokens Total audio out: ~6,000 tokens Total cacheable prefix: ~9,000 tokens (system + tool schemas)

The Dollar Math

Without prompt caching:

• Cacheable prefix as input: 9,000 x $32/1M = $0.288
• Other input: 6,500 x $32/1M = $0.208
• Output: 6,000 x $64/1M = $0.384
• Per-call cost: ~$0.88

With prompt caching (after the first call):

• Cached prefix: 9,000 x $0.40/1M = $0.0036
• Other input: 6,500 x $32/1M = $0.208
• Output: 6,000 x $64/1M = $0.384
• Per-call cost: ~$0.60

Caching saves roughly $0.28 per call on this profile. At 50,000 calls per month that is $14,000/month — significantly more than most teams' entire engineering budget for the agent.

When Caching Pays Back

Cached input is 80x cheaper than non-cached input on GPT-Realtime-2. The break-even is almost immediate: as soon as the same prompt prefix is reused on a second call, you save more than the engineering cost of wiring up caching.

The real engineering question is not "should we cache" but "what is our cache hit rate." Three things drive it down:

1. Per-call dynamic content at the top of the prompt. Move all dynamic content (customer name, account state) to the bottom.
2. Prompt edits. Every time you change the system prompt, the cache invalidates. Version your prompts and roll out carefully.
3. Multi-tenant prompt fragmentation. If each tenant has a custom system prompt, your cache hit rate is bounded by per-tenant traffic.

Monthly Volume Scenarios

At our 5-minute-call profile with caching enabled:

• 2,000 calls/mo (~CallSphere Starter volume): 2,000 x $0.60 = $1,200/mo raw model
• 10,000 calls/mo (~Growth volume): 10,000 x $0.60 = $6,000/mo raw model
• 50,000 calls/mo (~Scale volume): 50,000 x $0.60 = $30,000/mo raw model

These numbers are raw model spend only. They exclude STT (if you split the stack), telephony minutes (Twilio is usually $0.0085–$0.014/min inbound), CRM integration, hosting, observability, on-call, and the engineers maintaining all of it.

What Teams Get Wrong

Three pricing mistakes we see weekly:

• Forgetting output dominates. Output at $64/1M is 2x input. Verbose assistants are 2x more expensive than terse ones. Prompt your agent to be concise.
• Treating audio tokens like text tokens. Audio tokenization is denser. A "500-word" reply in audio is not 500 audio tokens.
• Ignoring the silent tail. Some implementations keep streaming after the user has stopped talking. Cut the listening window aggressively.

Where CallSphere Fits

CallSphere is a managed voice and chat agent platform. Customers pay per-interaction tiers — Starter $149/mo (2,000 interactions), Growth $499/mo (10,000), Scale $1,499/mo (50,000) — instead of metering raw audio tokens. The platform handles prompt caching, multi-tenant prompt management, STT/TTS routing, and the ops layer. For teams who would rather not run the cache hit-rate dashboard themselves, that is the trade.

Try the math against your volume: callsphere.ai/pricing.

What To Do This Week

1. Compute your current cache hit rate. If you do not measure it, assume it is 0.
2. Refactor your prompt so dynamic content is at the bottom, not the top.
3. Set up a per-call cost dashboard. The unit economics of voice agents move fast at scale; flying blind is expensive.

FAQ

Q: Is cached input really 80x cheaper? A: Yes — $0.40 vs $32 per 1M is 80x. It is the largest single cost lever in the entire pricing sheet.

Q: How long does cache persistence last? A: OpenAI does not publish a hard TTL; in practice expect minutes-to-hours, not days. Plan accordingly.

Q: Can I cache tool call results? A: No — the cache is for input tokens. Tool results are computed each call.