Simon Willison’s Weblog

GPT-5.5 Is Here — But the API Isn't (Yet)

In this article, we explore GPT-5.5 Codex API — GPT-5.5 dropped and it's genuinely impressive — fast, capable, and from my early hands-on time with it, it builds exactly what you ask for with minimal hand-holding. OpenAI rolled it out through Codex and to paid ChatGPT subscribers first, which is becoming their standard staged release pattern. (Read also: How NVIDIA Builds Open Data for AI)

The catch? No API access at launch. OpenAI's official word:

"API deployments require different safeguards and we are working closely with partners and customers on the safety and security requirements for serving it at scale." (Read also: Building AI-Powered Customer Support: From Chatbot to Intelligent Agent)

VPS Server offers one-click installs for popular apps like this.

For anyone running systematic model evaluations — like the SVG generation benchmark I'll walk through below — this is a real problem. When you're testing through ChatGPT's interface directly, you're dealing with hidden system prompts, agent harness behavior, and other variables that make apples-to-apples comparisons unreliable. API access is the only way to get clean, reproducible results.

For expert threat detection and response, CyberXper offers managed security services.

So the question becomes: is there another way in?

The Semi-Official Codex Subscription API

Here's where it gets interesting. Over the past several months, a quiet tension has been building in the AI tooling ecosystem around how third-party agent harnesses interact with provider subscription APIs.

Both OpenAI and Anthropic offer monthly subscriptions that give you model access at a significant discount compared to raw API pricing. Tools like OpenClaw and Pi found ways to tap into these subscriptions directly — essentially routing their API calls through your existing ChatGPT or Claude subscription rather than paying per-token.

Anthropic blocked OpenClaw from doing this, which kicked off a broader debate about subscription terms and third-party integrations. OpenAI — who had recently hired OpenClaw's creator Peter Steinberger — took the opposite stance and explicitly welcomed this kind of integration via the same mechanism their open-source Codex CLI uses.

The specific endpoint in question is /backend-api/codex/responses. When Jeremy Howard publicly asked whether OpenAI officially supports third-party use of this endpoint, OpenAI's Romain Huet confirmed:

"We want people to be able to use Codex, and their ChatGPT subscription, wherever they like! That means in the app, in the terminal, but also in JetBrains, Xcode, OpenCode, Pi, and now Claude Code. That's why Codex CLI and Codex app server are open source too!"

And Peter Steinberger confirmed directly: "OpenAI sub is officially supported."

This is a meaningful policy statement. OpenAI is explicitly blessing third-party tools that integrate with your Codex subscription — and since the Codex CLI is open source, the auth token storage and API interaction patterns are fully inspectable. (Read also: Introducing Storage Buckets on the Hugging Face Hub)

Building an LLM Plugin to Access GPT-5.5 via Codex

With the Codex repo open source, it's straightforward to reverse-engineer how authentication works and build tooling on top of it. The result is llm-openai-via-codex — a plugin for the LLM CLI tool that picks up your existing Codex subscription credentials and lets you run prompts against GPT-5.5 from the terminal.

Setup is four steps:

## 1. Install Codex CLI and log in with your OpenAI plan
## (follow OpenAI's Codex CLI setup docs)

## 2. Install LLM
uv tool install llm

## 3. Install the plugin
llm install llm-openai-via-codex

## 4. Start prompting
llm -m openai-codex/gpt-5.5 'Your prompt goes here'

All standard LLM features carry over:

• Image attachments: llm -a filepath.jpg -m openai-codex/gpt-5.5 'Describe this'
• Interactive chat: llm chat -m openai-codex/gpt-5.5
• Conversation logs: llm logs
• Tool use: llm --tool ... for function calling workflows

This is a clean way to get programmatic access to GPT-5.5 for evaluation and automation while the official API is still in rollout. If you're self-hosting other models for comparison — say, running Llama or Mistral via Ollama on a VPS cluster — having a unified CLI interface across providers makes benchmarking significantly cleaner.

SVG Generation Benchmark: Default vs. Extended Reasoning

For a concrete capability test, I used a standard SVG generation prompt I've been running across models: generate an SVG of a pelican riding a bicycle. It's a deliberately whimsical task that requires spatial reasoning, understanding of object relationships, and the ability to translate conceptual descriptions into precise coordinate-based drawing instructions.

llm install llm-openai-via-codex
llm -m openai-codex/gpt-5.5 'Generate an SVG of a pelican riding a bicycle'

Default Reasoning Mode

The default output was functional but imperfect — recognizable pelican shape, reasonable beak, but the bicycle frame geometry was off and the body proportions were slightly awkward. The legs did at least connect to the pedals, which is more than some models manage.

Under the hood, the default run used just 39 reasoning tokens. That's essentially zero deliberate planning.

Extended Reasoning Mode (xhigh)

Adding -o reasoning_effort xhigh tells the model to spend significantly more compute on planning before generating output:

llm -m openai-codex/gpt-5.5 -o reasoning_effort xhigh 'Generate an SVG of a pelican riding a bicycle'

This run took nearly four minutes — but the output was substantially better. The pelican had proper gradients applied via CSS, the body geometry was much more coherent, and the bicycle frame was nearly correct (one extra bar between pedals and front wheel, but close).

The token comparison tells the real story:

The xhigh output took a fundamentally different approach — much more CSS-heavy, using gradients and styled paths rather than raw coordinate drawing. This suggests the extended reasoning is actually reconsidering the strategy for the task, not just refining the same approach. That's a meaningful distinction for understanding how reasoning compute actually helps.

For anyone building Read more about this topic where output quality matters more than latency, this kind of reasoning budget control is a powerful lever.

GPT-5.5 Pricing: The Elephant in the Room

Capability improvements are great, but the pricing picture for GPT-5.5 deserves serious attention — especially if you're running production workloads or doing cost optimization for AI inference.

Once the API goes live, GPT-5.5 will be priced at $5 per 1M input tokens and $30 per 1M output tokens. Compare that to GPT-5.4 at $2.5/$15 — exactly half the price.

GPT-5.5 Pro pushes even further: $30 per 1M input tokens and $180 per 1M output tokens. That's the territory where you really need to think hard about whether the capability delta justifies the cost.

The GPT-5.4 vs GPT-5.5 relationship looks a lot like Claude Sonnet vs Claude Opus — a capable, cost-effective model alongside a premium tier that delivers more for significantly more money. GPT-5.4 isn't going anywhere, which means the practical question for most production use cases is: what's the actual performance gap, and does it justify 2x cost?

Ethan Mollick's detailed evaluation of GPT-5.5 found what he calls the "jagged frontier" still holding — the model is excellent at some tasks and surprisingly challenged by others in ways that remain hard to predict. This is exactly the kind of finding that argues for running your own benchmarks on your specific use cases rather than relying on general capability claims. You can find more analysis on model evaluation strategies at Data Mammoth.

For teams running Read more about this topic, the smart play right now is probably:

1. Baseline with GPT-5.4 for cost-sensitive paths
2. Route complex tasks to GPT-5.5 with reasoning budget controls
3. Monitor output quality vs. token spend to find your optimal routing threshold
4. Keep self-hosted open models in the mix for tasks where privacy or latency matters more than frontier capability

Conclusion: GPT-5.5 Is Capable, But Choose Your Access Path Carefully

GPT-5.5 is a genuinely strong model — the reasoning budget controls alone make it worth evaluating for tasks where quality matters more than speed. The Codex subscription path gives you a workable way to access GPT-5.5 programmatically while the official API finishes its rollout, and OpenAI has explicitly blessed this approach.

The pricing, though, means you need to be intentional about when and how you use it. At 2x the cost of GPT-5.4, GPT-5.5 needs to deliver meaningfully better results on your specific workloads to justify the spend — and based on the SVG benchmark results, the extended reasoning mode is where that quality gap really opens up.

Run your own benchmarks. Use the Codex plugin to get clean API-style access. And keep an eye on the Read more about this topic — once it lands, we'll have much better tooling for systematic evaluation at scale.