How to Use GPT-5.5 API: Pricing and Setup Guide

GPT-5.5 launched on the OpenAI API April 23, 2026. If you’re upgrading from GPT-5.4 or building new agents, this guide covers everything: endpoints, pricing, rate limits, computer-use setup, and the one-line migration path.

Last verified: April 24, 2026

1. The one-line upgrade

If you’re already using the OpenAI Python SDK:

from openai import OpenAI
client = OpenAI()

response = client.chat.completions.create(
    model="gpt-5.5",  # was "gpt-5.4"
    messages=[{"role": "user", "content": "Fix this bug: ..."}]
)

Or with the Responses API (recommended for agents):

response = client.responses.create(
    model="gpt-5.5",
    input="Fix this bug: ...",
    tools=[{"type": "computer_use_preview"}],  # native computer use
    max_output_tokens=10000,
)

If your account doesn’t have GPT-5.5 access yet, OpenAI’s official guidance is: “If gpt-5.5 isn’t available in your account yet, use gpt-5.4.” Access is rolling out to all paying developers through early May 2026.

2. Pricing in April 2026

Token type	Price per 1M
Input (fresh)	$1.50
Input (cached)	$0.15
Output	$12.00
Reasoning	$12.00 (billed as output)

GPT-5.5 prices match GPT-5.4 flagship. A GPT-5.5 mini tier is expected in 4–6 weeks at ~$0.15 / $1.20 per million — target date based on OpenAI’s past release cadence.

Realistic task costs:

Task	Cost
50K in / 5K out (typical chat)	$0.13
50K in / 10K out (agent task)	$0.20
200K in / 30K out (long agent session)	$0.66
400K in / 50K out (max context + heavy output)	$1.20

For comparison: Claude Opus 4.7 on the same workload costs 10x more.

3. Rate limits

As of launch day (April 23, 2026):

Tier	RPM	TPM (input)
Tier 1 ($5 paid)	500	30,000
Tier 2 ($50 paid)	5,000	450,000
Tier 3 ($100 paid)	5,000	800,000
Tier 4 ($250 paid)	10,000	2,000,000
Tier 5 ($1,000 paid)	10,000	20,000,000

Rate limits match GPT-5.4 tier-for-tier. If you’re already a Tier 3+ customer, you have plenty of headroom for production use.

4. Native computer use

The biggest API-level change in GPT-5.5 is native computer use. In GPT-5.4, you enabled it like this:

# GPT-5.4 style — verbose
tools = [{
    "type": "computer_use_preview",
    "model": "computer-use-preview-2025-10"  # separate model
}]

In GPT-5.5, it’s just a flag:

# GPT-5.5 style — native
tools = [{"type": "computer_use_preview"}]

The response stream includes action items you execute in your environment:

for chunk in response:
    if chunk.type == "action":
        if chunk.action == "click":
            # click at chunk.x, chunk.y
        elif chunk.action == "type":
            # type chunk.text
        elif chunk.action == "screenshot":
            # take screenshot, return as next input

Latency is ~40% faster than GPT-5.4 with the old computer-use model, because it’s a single model call per step instead of two.

5. Building agents with the Agents SDK

OpenAI’s Agents SDK (Python and TypeScript) defaults to gpt-5.5 as of April 23:

from openai.agents import Agent, Runner

agent = Agent(
    name="CodeReviewer",
    model="gpt-5.5",  # default in v1.5+
    instructions="You review pull requests and flag issues.",
    tools=["computer_use", "file_read", "bash"],
)

result = Runner.run_sync(agent, "Review PR #1234")

Dynamic Reasoning Time (DRT) lets agents run for 7+ hours on a single task. Enable with:

agent = Agent(
    model="gpt-5.5",
    max_reasoning_hours=7,  # was capped at 2 in GPT-5.4
    instructions="...",
)

6. Codex CLI / IDE extension

If you use Codex:

• Codex CLI: gpt-5.5 is the default as of April 23. No action needed.
• Codex IDE extension (VS Code): Auto-updated. Verify in Settings → Codex → Model.
• Codex Cloud: Default in all new environments.

Existing Codex sessions keep their current model until restarted.

7. Migration checklist from GPT-5.4

• Change model parameter to "gpt-5.5"
• Remove explicit computer-use model version strings
• Test on 10–20 representative tasks
• Check token budget — GPT-5.5 uses ~10% more output tokens on average
• Verify tool-calling behavior in your agent loop
• If you use structured outputs, re-run your Pydantic tests (schema adherence is tighter)
• Update evals and A/B test in production for 48 hours
• Swap defaults if metrics improve

8. What not to do

• Don’t upgrade batch classification jobs if you use GPT-5.4 mini — wait for GPT-5.5 mini (expected May/June 2026).
• Don’t assume longer context. Still 400K. If you need more, route to Claude Opus 4.7 or Gemini 3.1.
• Don’t re-tune your prompts immediately. GPT-5.5 is more instruction-following by default; over-prompting can actually degrade quality.
• Don’t skip evals. OpenAI’s own blog acknowledges “there are enough model releases that it’s probably getting hard to distinguish one from another.” Prove the upgrade is worth it on your workload.

9. Gotchas

• Reasoning tokens are billed as output tokens. If you enable high reasoning effort, expect 2–5x more output tokens than GPT-5.4 for the same task.
• Computer use requires a safe execution environment. Don’t run GPT-5.5 browser actions on your own machine without sandboxing.
• Zero-data-retention (ZDR) is available for enterprise customers. NVIDIA uses ZDR for its production agents running on GPT-5.5.
• Structured outputs work but use strict JSON schema. Loose schemas occasionally cause higher refusal rates than GPT-5.4.

10. When to use GPT-5.5 vs alternatives

• Use GPT-5.5 for: Agents, computer use, Codex, multi-step reasoning, cost-sensitive production
• Use GPT-5.4 mini for: High-volume classification until 5.5 mini ships
• Use Claude Opus 4.7 for: Hard coding tasks, long-context (1M), refactors
• Use Gemini 3.1 Pro for: Very long documents (2M), multimodal, Workspace integration
• Use Kimi K2.6 for: Budget-optimized agents at $0.10/$0.30 per million

---

Last verified: April 24, 2026. Sources: OpenAI introducing GPT-5.5 (openai.com/index/introducing-gpt-5-5), OpenAI API docs (platform.openai.com/docs), Codex docs (developers.openai.com/codex), NVIDIA Blog.