GPT-5.5 Pro lists at $30/$180 per million tokens, and it drops the cache discount that keeps the standard model cheap.

What you actually pay

Here is the rate card next to the models people put it up against, pulled from each provider's own pricing page. The cached column is the one to watch, because it is where Pro breaks ranks with everything else in the table.

GPT-5.4 Pro sits at the identical $30/$180, so the Pro price has not moved across a generation. The older o3-pro is the one premium reasoning model that undercuts it, at $20/$80, though it caps at a 200K window and a generation behind on quality. The line that does the damage is the cached column: standard GPT-5.5, Opus 4.7, and Gemini all knock most of the cost off reused input, while GPT-5.5 Pro shows a flat "none". That single entry rewrites the comparison the moment your workload starts repeating itself.

What the Pro tier is actually doing

It helps to know what you are buying, because "Pro" sounds like a bigger, smarter model and it is not. GPT-5.5 Pro is the same GPT-5.5 weights running with parallel test-time compute. Instead of one reasoning pass, it spins up several at once, lets them work the problem along different paths, and picks the strongest result. That is the same trick the original GPT-5 Pro used. More compute per question, not more model.

That single fact explains the whole pricing shape. You are renting extra GPU time on every call, which is why the markup is flat 6x on both input and output. It is also why OpenAI steers the model toward the stateful Responses API, the request format parallel reasoning is built around. It also explains the missing cache discount. Caching saves money by skipping recomputation on a stable prefix. A model whose entire pitch is doing more computation per request has the least to gain from a feature that does less of it, so OpenAI simply does not offer it on this tier.

The cache discount you give up

This is the part most pricing pages bury. Standard GPT-5.5 charges $0.50 per million on cached input, a 90 percent cut on the tokens it has seen before. Anything you re-send unchanged, the system prompt, a big spec, the file tree an agent keeps in view, rides that discount. GPT-5.5 Pro charges $30 on those same tokens. Every time. On the cached slice of a workload, that is a 60x gap, not 6x.

Whether that matters comes down to how much you repeat. A chatbot answering fresh one-line questions barely caches anything, so the penalty is small. An agent is the opposite: it resends a fat, mostly-stable context on every turn, and on standard GPT-5.5 the second turn onward is nearly free on input. Take that same agent to Pro and you pay full freight on the whole prefix, forty turns deep. The headline says 6x. Your bill says more.

Two workloads, two very different bills

Numbers make this concrete. Take two jobs at opposite ends. First, a hard one-shot problem: 30K tokens of context in, 40K tokens of reasoning and answer out, nothing reused. Second, a working agent session: 2M input tokens across the turns, 80 percent of them cache hits on a model that caches, plus 200K tokens of output. Same two models, side by side.

On the one-shot job the multiple is exactly the 6x you were promised, because there is nothing to cache and the discount never enters the math. On the agent session it jumps to almost 11x, driven entirely by that 2M of input getting billed at $30 on Pro instead of mostly $0.50 on standard. Run a hundred of those sessions a month and you are choosing between an $880 bill and a $9,600 one. The work is identical.

What the extra compute actually buys

None of this matters if Pro is not measurably better, so here is where the evidence actually stands. OpenAI has not published a clean Pro-versus-standard table across the usual benchmarks, and Artificial Analysis scores GPT-5.5 by reasoning effort rather than listing Pro as a separate entry. The one Pro-specific number that is clearly reported is FrontierMath Tier 4, the hardest research-math set, where GPT-5.5 Pro hit 39.6 percent against 22.9 for Claude Opus 4.7 and 16.7 for Gemini 3.1 Pro. On a problem set that brutal, the gap is real.

That is the shape of where Pro earns its keep: the further out on the difficulty curve you go, the more parallel reasoning pays. On everyday work it matters far less. Standard GPT-5.5 already posts 93.5 on GPQA Diamond, a hair behind Gemini 3.1 Pro at 94.1, and still tops the Artificial Analysis intelligence index overall. The extra passes mostly re-derive an answer the base model would have reached anyway, so you are paying 6x for a margin that shrinks toward zero as the task gets easier. We have not run the two side by side on our own evals, so treat that FrontierMath gap as OpenAI's to defend rather than ours to vouch for.

When to reach for it, when to walk past

Reach for Pro on the rare, high-stakes, single-shot question: a gnarly proof, a subtle financial model, a security review where a missed edge case costs more than the entire month of API spend. There you are not buying tokens, you are buying the best odds of a right answer on the first try, and 6x on a job you run once is noise next to the cost of being wrong.

Walk past it for anything that loops. Coding agents, batch extraction, anything that re-sends context turn after turn, those are exactly the workloads the missing cache discount punishes hardest, and standard GPT-5.5 at high reasoning effort gets you most of the way for a tenth of the money. If you genuinely need more thinking on a repeated task, turning up the reasoning effort on the standard model is almost always the better lever than jumping to Pro.

To put your own token mix against the spread, the calculator runs the math for any input/output split, and the pricing page lists every model side by side. For how the standard model's per-token price shifted, see the GPT-5.5 token-efficiency breakdown, and for the three-way flagship fight, the GPT-5.5 vs Opus 4.7 vs Gemini 3.1 Pro comparison.

Sources

• OpenRouter: GPT-5.5 Pro - $30 input, $180 output per 1M, context and output limits
• OpenAI: API pricing and model guide - GPT-5.5 $5/$30 with $0.50 cached, cache discount not offered on Pro, Responses API only
• Anthropic: Claude pricing - Opus 4.7 $5/$25, $0.50 cached input
• Google: Gemini API pricing - Gemini 3.1 Pro $2/$12 per 1M
• Artificial Analysis - GPT-5.5 intelligence index standing, GPQA Diamond, reasoning-effort scoring
• The Decoder - FrontierMath Tier 4 results and GPT-5.5 launch coverage