GPT-5.5 ships: native desktop control, 40% fewer tokens, double the price

OpenAI shipped GPT-5.5 on April 23, 2026 — five weeks after GPT-5.4. The release cycles are getting shorter. The capabilities are getting weirder. And the pricing just doubled.

What’s actually new

Native desktop control. This is the headline feature. GPT-5.5 is the first general-purpose model with built-in capability to navigate desktop applications, click buttons, and type text across multi-step workflows. Not Computer Use as a tacked-on agent loop — desktop interaction as a model-native action class. You can hand it “operate this CRM, find the customer, update their address” and it does the clicking and typing itself.

40% fewer tokens for the same Codex tasks. OpenAI’s claim, and one I’d weight seriously because token efficiency is verifiable. On agentic coding loops, GPT-5.5 reaches the correct answer in fewer turns than GPT-5.4. If the claim holds at scale, it changes the per-task cost arithmetic — even with the doubled per-token price below.

Same per-token latency as GPT-5.4 in real-world serving. OpenAI claims they shipped “more intelligence at the same speed.” Per-second throughput stays roughly equivalent.

FrontierMath Tier 4: 39.6% for GPT-5.5 Pro, nearly double Claude Opus 4.7’s 22.9%. FrontierMath Tier 4 is the hardest tier — research-level mathematics. This is a real lead on hard reasoning, not benchmark contamination noise.

Workspace agents. OpenAI also launched workspace agents for ChatGPT Business, Enterprise, and Edu — shared team-level agents that can use Slack, Gmail, request approvals, and improve over time.

MCP auto-detection. GPT-5.5 can figure out how to use an MCP server without you explaining it — it inspects the schema and works it out. For anyone building agent tooling, this materially cuts the prompt-engineering tax.

The pricing reality

Here’s where the news is less fun:

Tier	Input ($/1M tokens)	Output ($/1M tokens)	vs GPT-5.4
GPT-5.5	$5	$30	2× input, 1.5× output
GPT-5.5 Pro	$30	$180	2× input, 1.5× output
GPT-5.4 (still available)	$2.50	$20	baseline

Doubling the input price at the same time as launching a smarter model is a deliberate move. OpenAI is signalling that the model tier ladder going forward will have meaningful price separation — not just feature gates.

For consumer ChatGPT users, the model is rolling out to Plus, Pro, Business, and Enterprise users. GPT-5.5 Pro is Pro, Business, and Enterprise only. Free tier doesn’t get it.

What this changes vs Claude and Gemini

On hard math: GPT-5.5 Pro is now decisively ahead. 39.6% vs 22.9% on FrontierMath Tier 4 is not close.

On agentic coding: Mixed. The 40% token efficiency gain matters, but Claude Opus 4.7’s SWE-bench Verified at 87.6% is still the leader on standard coding benchmarks. The narrative will land where you choose to look.

On general intelligence: GPT-5.5 vs Claude Opus 4.7 vs Gemini 3.1 Pro is now a genuinely close three-way race. None of the three is the “obvious answer” the way GPT-4 was the obvious answer in 2023.

On price: GPT-5.5 just got more expensive than Claude Opus 4.7 ($5 input both, but $30 output for GPT-5.5 vs $25 for Opus 4.7). Gemini 3.1 Pro at $2/M input remains the value choice.

The “super app” framing

OpenAI is openly positioning GPT-5.5 as the engine of a unified ChatGPT super-app — chat + Codex + browser + image gen + voice + Sora’s spiritual successor (whatever that turns out to be) all in one product surface.

Native desktop control is the technical underpinning that makes this credible. If the model can drive your machine, you stop needing separate “agent” products and just have ChatGPT do the thing.

It also explains the price doubling: the unified super-app is a more expensive product to run than discrete chat. OpenAI is recouping that compute through the API price ladder while keeping the consumer subscription tier flat.

The risk classification footnote

Worth flagging: GPT-5.5 meets OpenAI’s “High” risk classification for cybersecurity, though it does not cross the “Critical” threshold. OpenAI’s framing: it could “amplify existing pathways to severe harm.” This is the same kind of language Anthropic used about Claude Mythos Preview — the model the lab acknowledges is too dangerous to ship.

The pattern is now consistent: each frontier-lab launch ships the safe-to-release version while explicitly conceding more dangerous capability tiers exist. We’re no longer in 2023’s “AI safety is theoretical” framing.

If you’re on ChatGPT Plus ($20): You get GPT-5.5 in chat. The desktop control feature alone makes this worth using. No price change at the consumer tier (yet).

If you’re an API user paying for GPT-5.4: Don’t blindly upgrade. The doubled input price plus 40% token efficiency means break-even depends on your task mix. Run a week of production traffic against GPT-5.5 before committing.

If you’re on ChatGPT Pro ($100/$200): GPT-5.5 Pro access at the higher tier. If you’re doing real research-grade math work, the FrontierMath jump is the upgrade reason.

If your stack is Claude Code and Cursor: GPT-5.5 doesn’t change your immediate decision. Claude Opus 4.7 still leads SWE-bench Verified. But Codex is closing fast, and the 40% token efficiency claim is exactly the metric that erodes Claude Code’s “best agent” lead over the next six months.

The bigger picture

In the last 9 days the AI category got: Claude Opus 4.7 (April 16), GPT-5.5 (April 23), and tomorrow’s launch — DeepSeek V4 (April 24). Three frontier-lab launches in a single working week.

The 2024 narrative was “OpenAI is far ahead.” The 2025 narrative was “Anthropic catches up on quality.” The 2026 narrative is shaping up as “everyone ships every six weeks and pricing arbitrage is the actual differentiator.”

That’s good for users. It’s brutal for the labs trying to maintain margin. The doubled GPT-5.5 prices say OpenAI knows it.

Related tool reviews

Questions or corrections? Email Pick Right. Want the full list? See all news.