OpenAI shipped GPT-5.5 yesterday — April 23, 2026 — codename "Spud." It's the first fully retrained base model since GPT-4.5, and the first one OpenAI is explicitly marketing as an agentic model rather than a chatbot.
I've been running it in Codex for about a day. Here's what's actually new, what the benchmarks show, and why ClawBox owners don't need to care about any of it — until they want to.
What GPT-5.5 Actually Is
Three variants ship:
- GPT-5.5 — the default frontier model
- GPT-5.5 Thinking — extended reasoning mode
- GPT-5.5 Pro — the higher-accuracy variant for harder problems
Availability today (April 23–24):
- ChatGPT — Plus, Pro, Business, Enterprise
- Codex — all paid plans, 400K context window
- API — "very soon," pending additional safeguards
Multimodal is text + vision (same input stack as GPT-5). No native image, audio, or video output.
The Benchmark Reality
Every model launch is a benchmark parade. Here's the honest tally against the two models that matter for agentic work — Anthropic's Claude Opus 4.7 and Google's Gemini 3.1 Pro:
| Benchmark | GPT-5.5 | Claude Opus 4.7 | Gemini 3.1 Pro |
|---|---|---|---|
| Terminal-Bench 2.0 (agentic CLI) | 82.7% | 69.4% | 68.5% |
| GDPval (knowledge work) | 84.9% matches/beats pros | — | — |
| Expert-SWE (long-horizon coding) | 73.1% (up from 5.4's 68.5%) | — | — |
| SWE-Bench Pro (real-repo PRs) | 58.6% | 64.3% | — |
OpenAI claims state-of-the-art on 14 benchmarks vs. 4 for Opus 4.7 and 2 for Gemini 3.1 Pro.
The honest read:
- Unattended terminal / tool-coordination workflows — GPT-5.5 wins clearly.
- Multi-file repo work with real PRs — Opus 4.7 still leads on SWE-Bench Pro.
- Everyday chat and writing — you won't notice the difference day-to-day.
This is not "GPT-5.5 replaces everything." It's "there's now a model that's measurably better at specific agentic tasks — and a different model that's still better at others."
The Price
This is the part OpenAI's landing page doesn't lead with:
- GPT-5.5 (standard) — $5 per 1M input tokens / $30 per 1M output tokens
- GPT-5.5 Pro — $30 per 1M input / $180 per 1M output
That's roughly double GPT-5.4's API pricing. OpenAI's counter-argument is that 5.5 uses fewer tokens per task in Codex, so total cost on real workloads is closer to flat than to 2×. Early reports from teams using it (Canva, among others) support that — but "closer to flat" still means you pay per token, and per token, this is the most expensive frontier model OpenAI has ever shipped.
What This Means for ClawBox
Short answer: nothing has to change, but you get a new option.
ClawBox is provider-agnostic by design. During setup, you pick from:
- Anthropic Claude
- OpenAI GPT
- Google Gemini
- ClawAI — free, no setup
The "OpenAI GPT" option supports both API keys and ChatGPT Plus/Pro OAuth — so if you already pay for ChatGPT, you're not paying twice. When GPT-5.5 lands in the API (OpenAI says "very soon"), it slots straight into that same provider slot. No ClawBox update required for most setups.
The more interesting question is which model you route what to.
Why Local Memory + Cloud Routing Gets More Valuable, Not Less
As frontier models climb toward $180 per million output tokens, the cost of routing every question to the best model goes up fast. The architecture ClawBox has been shipping since day one — local memory on the device, cloud APIs on demand — becomes more defensible the pricier the frontier gets.
A sensible routing pattern on a ClawBox in April 2026 looks something like:
- ClawAI (free, local-ish): daily chat, quick lookups, the 80% of requests that don't need frontier intelligence
- GPT-5.5 or Opus 4.7: agentic coding, long-horizon research, genuinely hard problems
- GPT-5.5 Pro: the rare task where the accuracy delta is worth $180 per million output tokens
Your memory, your conversation history, your automations — all stay on the device. You only pay the frontier price for the queries that actually need it.
Compare that to paying a flat monthly subscription that routes everything through a frontier model whether you need it or not. The math gets worse every time OpenAI raises prices, not better.
The Real Takeaway
GPT-5.5 is a genuinely strong agentic model. Terminal-Bench 2.0 at 82.7% is not a marginal jump — it's the current state of the art for command-line workflows, and anyone doing serious Codex work should try it.
It's also twice the price of GPT-5.4. And Opus 4.7 still beats it on real-repo software engineering.
Which is exactly why ClawBox doesn't pick a winner. The whole point of owning the hardware is that you get to choose — and that choice gets cheaper as the model landscape gets more expensive and more fragmented.
ClawAI for everyday. GPT-5.5 on tap when you need the agentic firepower. Opus 4.7 when the task is multi-file surgery. Your data stays yours either way.
That's the only architecture that makes economic sense when the frontier is this crowded.
Sources: