GPT-5.5 Pro Benchmarks: The Numbers OpenAI Published (and the Ones It Didn't)

GPT-5.5 Pro Benchmarks: The Numbers OpenAI Published (and the Ones It Didn't)

OpenAI shipped GPT-5.5 Pro on April 23, 2026, and priced it at $30 per million input tokens and $180 per million output. That's 6× the $5/$30 base GPT-5.5. For the money you get a higher-compute variant that posts a 90.1% on BrowseComp. But here's the part nobody puts in the headline: OpenAI's own eval table leaves the GPT-5.5 Pro coding row blank. Same for computer-use. Same for long-context. So when you search for the model's SWE-Bench Pro score, there isn't one to find. We pulled the published table and graded every figure ourselves.

Quick answer:

• GPT-5.5 Pro is OpenAI's higher-compute variant of GPT-5.5 (released April 23, 2026), not a new model.
• It leads on browsing (BrowseComp 90.1%) and frontier math, but OpenAI left coding, computer-use and long-context rows blank.
• On the coding rows that exist, base GPT-5.5 trails Claude Fable 5 (SWE-Bench Pro 58.6% vs 80.3%).
• Claude Fable 5 is currently suspended (US export-control directive, around June 12, 2026), so its wins come with a "can't buy it now" asterisk.

A quick note on method, because accuracy matters more than speed on a benchmark post: the numbers below are cross-checked against OpenAI's and Anthropic's published tables and the SWE-Bench Pro paper (arXiv 2509.16941). Where only one vendor reports a figure, we say so. Where OpenAI never published a Pro score, we write "not published" instead of quietly swapping in the base number.

GPT-5.5 Pro Benchmarks: What OpenAI Actually Published

GPT-5.5 Pro's published benchmarks cover a narrow slice: browsing, frontier math, professional knowledge work, genetics and broad reasoning. OpenAI ran every eval at reasoning effort xhigh in a research environment, which differs from the production ChatGPT default. The headline number is BrowseComp 90.1%, well clear of base GPT-5.5's 84.4%.

Here's the master table, with a column showing whether OpenAI published a separate Pro score for each test:

Read that bottom block carefully. OpenAI published GPT-5.5 Pro scores for browsing and math, but left the coding, computer-use, cybersecurity and long-context rows blank. The BrowseComp, FrontierMath and GDPval-base figures are confirmed against secondary trackers; the GDPval-Pro 82.3%, GeneBench 33.2% and the Investment Banking number are reported by OpenAI but not independently verified, so we attribute them rather than state them flat.

What "Pro" Actually Means: Parallel Test-Time Compute, Not a New Model

GPT-5.5 Pro is the same GPT-5.5 model running with far more reasoning effort, not a different architecture. Think of it less as a new engine and more as the same engine running longer and in parallel before it answers. OpenAI calls the setting reasoning effort, and Pro pins it at the top tier: xhigh.

Is GPT-5.5 Pro a different model from GPT-5.5?

No. Same weights, same training. When people search gpt 5.5 pro vs gpt 5.5 thinking or vs xhigh, they're really asking about one dial: how hard the model thinks before it replies. "Parallel test-time compute" means the model explores several reasoning paths at once and picks the best, which costs more tokens and more wall-clock time. That extra compute is what you pay 6× for.

The specs are otherwise shared. GPT-5.5 Pro carries a context window of roughly 1.1M tokens input and 128K output. So you're not buying a bigger memory or a smarter base model. You're buying more thinking time on the model you already know.

GPT-5.5 Pro vs GPT-5.5 (Standard): Where the 6× Price Buys You Something

GPT-5.5 Pro beats base GPT-5.5 on the hardest tasks: browsing, frontier math, and genetics. It buys you the least on routine work. The cleanest gain is BrowseComp 90.1% vs 84.4%, a 5.7-point lift on deep web research. On FrontierMath Tier 4 it climbs to 39.6% from 35.4%.

But more compute doesn't always mean a higher score. On GDPval, OpenAI's table actually lists Pro lower than base GPT-5.5 (82.3% claimed vs 84.9% confirmed). That's counterintuitive, and it's reported as Pro trading some raw wins for calibration. The point stands: paying more isn't a guarantee.

Where Pro pulls ahead:

• BrowseComp: 90.1% vs 84.4% (+5.7)
• FrontierMath T4: 39.6% vs 35.4% (+4.2)
• GeneBench: 33.2% vs 25.0% (claimed by OpenAI)
• HLE with tools: 57.2% (claimed) vs 52.2% (confirmed base)

Where it's flat or worse:

• GDPval: 82.3% (claimed) vs 84.9% base
• HLE no-tools: 43.1% vs 41.4%, barely a move

If your work is mostly everyday drafting, code review and summaries, the 6× premium is hard to justify. The math only flips when the task is genuinely hard. While we're on cost: if your bill is the problem, our guide to cut your LLM API costs covers routing cheaper models for the easy 80% and reserving Pro for the hard 20%.

GPT-5.5 Pro vs Claude Fable 5

On the coding benchmarks both vendors report, Claude Fable 5 beats base GPT-5.5 decisively. But Fable 5 is currently suspended, so the win comes with an asterisk. And the comparison is messier than it looks, because OpenAI did not publish GPT-5.5 Pro coding scores at all. So the honest matchup is really Fable 5 vs base GPT-5.5 on coding, with Pro absent.

Here's the three-way table. The Fable 5 figures are attributed to Anthropic's published table; blank Pro cells are exactly that:

Two caveats live in that table. First, Terminal-Bench: base GPT-5.5 posts 82.7% on v2.0 and 83.4% on v2.1, while Fable's 88.0% is on v2.1, so make sure you're reading the same version before declaring a gap. Second, GDPval-AA isn't a percentage. It's an Elo-style score (1932 for Fable vs 1769 for base GPT-5.5), a different scale entirely, so don't mentally line it up against the percentage rows. The HLE no-tools Fable figure also wobbles by source: CodingFleet lists 56.8% while other summaries say 59.0%, so we report the range.

SWE-Bench Pro is the benchmark to anchor on for agentic coding. It runs 1,865 problems across 41 active repositories (per arXiv 2509.16941), with tasks that take a senior engineer hours or days and require multi-file patches. On that test, Fable 5's 80.3% against base GPT-5.5's 58.6% is a wide margin.

Now the asterisk. Claude Fable 5 was suspended in June 2026 under a US export-control directive (around June 12). So "Fable wins coding" is true on the numbers and currently moot at the checkout. If you want the deeper Anthropic-side picture, here's our full Claude Fable 5 breakdown. For the model GPT-5.5's table benchmarks against on math, see Claude Opus 4.8.

The Benchmarks OpenAI Did NOT Publish for Pro

OpenAI never released GPT-5.5 Pro scores for coding (SWE-Bench Pro), computer-use (OSWorld-Verified), cybersecurity, long-context recall, or abstract reasoning. Those rows are blank in the official table. Blank does not mean the model is bad at them. It means there is no published number, and anyone quoting one is either guessing or quoting the base score by mistake.

This matters because it's the most-searched gap. If you searched for GPT-5.5 Pro's SWE-Bench Pro score, there isn't one. OpenAI never published it. The only figure that exists for that benchmark on the GPT-5.5 family is the base model's 58.6%, and that's a base figure, not a Pro figure. We're flagging it that way on purpose.

What we can responsibly say:

• Coding (SWE-Bench Pro): Pro not published. Base GPT-5.5 = 58.6% (base, not Pro).
• Computer-use (OSWorld-Verified): Pro not published. Base = 78.7% (base, not Pro).
• Cybersecurity: Pro not published, no usable base proxy in the table.
• Long-context: Pro not published, no usable base proxy.
• Abstract reasoning: Pro not published.

Could Pro's parallel compute lift those base numbers? Probably, given the pattern on browsing and math. But "probably" is not a benchmark, and we won't print a number OpenAI didn't. That restraint is the whole reason this section exists.

Pricing: $5/$30 vs $30/$180 vs $10/$50 — Is Pro Worth 6×?

GPT-5.5 Pro costs roughly 6× base GPT-5.5: $30 input and $180 output per million tokens, against base at $5/$30. Claude Fable 5 sits between them at $10/$50. Worth it for hard research and frontier math, rarely worth it for everyday work.

So who should actually pay the premium? A rough verdict by job:

• Hard research, frontier math, deep multi-step browsing: GPT-5.5 Pro earns it. The benchmark gains are real and the task value is high.
• Agentic coding on large repos: Fable 5 if you can get it, otherwise base GPT-5.5 wrapped in a coding scaffold. Paying Pro prices for an unpublished coding score is a bad bet.
• Daily drafting, summaries, code review, support: base GPT-5.5. The 6× spend buys almost nothing here.

The trap is defaulting everything to Pro "to be safe." On output-heavy workloads that $180 figure compounds fast. Route by difficulty and you keep most of the quality at a fraction of the bill (more on that in our LLM API cost guide).

How We Checked These Numbers (and Where the Sources Disagree)

Before any of these figures went into this post, we did the unglamorous part: we pulled OpenAI's published GPT-5.5 eval table and checked every Pro row against independent trackers instead of trusting one screenshot. The sources we cross-referenced were llm-stats, BenchLM, DataCamp's Fable 5 write-up, CodingFleet, and the arXiv SWE-Bench Pro paper (2509.16941). Here is what held up and what didn't.

Most of the headline Pro numbers reconcile cleanly. BrowseComp 90.1%, FrontierMath Tier 1–3 52.4% and Tier 4 39.6%, and the $30/$180 pricing all matched across every source we checked. SWE-Bench Pro at 80.3% for Fable 5 versus 58.6% for base GPT-5.5, and FrontierCode Diamond at 29.3% versus 5.7%, also held up — and the SWE-Bench Pro task count (1,865 problems across 41 repositories) comes straight from the paper, not a vendor blog.

Three things did not reconcile, and you should know about them:

• Humanity's Last Exam, with tools, for Fable 5 shows up as anywhere from 56.8% to 59.0% depending on the source. We cite the range rather than pick a favorite.
• Terminal-Bench numbers float between version 2.0 and 2.1 across OpenAI's and Anthropic's tables, so the GPT-5.5 (83.4%) vs Fable (88.0%) gap isn't a clean apples-to-apples read.
• The Investment Banking Modeling score (88.6% Pro) is labeled "internal" by OpenAI, meaning no independent tracker can confirm it. We flag it as a vendor claim every time it appears.

That's the honest version. The numbers we present as fact reconciled across at least two independent sources; everything else is attributed to whoever reported it. We did not run GPT-5.5 Pro ourselves — at $30/$180 per million tokens, a serious head-to-head run isn't something we'll fake, so we're not going to pretend we have lab results we don't.

Real-World Results Vendors Are Reporting

These are vendor claims, not independent lab results, so read them as marketing-adjacent evidence. OpenAI and Anthropic each published case studies that paint a flattering picture. We're listing them attributed, with the caveat that none were verified by us.

On the OpenAI side, the company says a finance team used Codex with GPT-5.5 to review 24,771 K-1 tax forms (71,637 pages), finishing roughly two weeks faster than the prior year's process. OpenAI also reports a working algebraic-geometry app built from a single prompt in 11 minutes, and a GPT-5.5 Pro analysis of a gene-expression dataset spanning 62 samples and about 28,000 genes.

On the Anthropic side, Stripe (reported via Anthropic) used Fable 5 to run a codebase-wide migration on a 50-million-line Ruby codebase in a single day, against an estimated 2-plus months by hand. Anthropic also says Fable 5 completed Pokémon FireRed from raw screenshots, where earlier Claude models needed extra scaffolding tools, and that with persistent file-based memory it played Slay the Spire about 3× better than Opus 4.8.

Impressive demos, all of them. Just remember they're chosen by the vendors and not reproducible from the press release alone.

Which Model Should You Actually Use?

Match the model to the job, not the hype. For coding agents and large repos, reach for Claude Fable 5 if it's available to you, and base GPT-5.5 wrapped in a coding scaffold if it isn't. For research, frontier math and deep browsing, GPT-5.5 Pro is the pick. For daily work and cost efficiency, base GPT-5.5 wins on value almost every time.

A few pointers if you're choosing a stack rather than a single call. If you actually want an autonomous coding system, you want a tool, not a raw model, so start with the best AI coding agents in 2026 and the background coding agents compared roundup for the Codex and Devin context. Curious where the family goes next? Here's what's leaking about GPT-5.6.

At Techsy we route per task across our agent pipelines, top-tier reasoning models for the hard calls and cheaper ones for the rest, instead of paying premium rates on every request. If you want a second opinion on your own model mix, get a free consultation.

Frequently Asked Questions

Is GPT-5.5 Pro worth it?

It depends on the job. For hard research, frontier math and deep browsing, GPT-5.5 Pro's gains over base (BrowseComp 90.1% vs 84.4%) justify the roughly 6× price of $30/$180 per million tokens. For everyday drafting, code review and summaries, base GPT-5.5 gives you almost the same quality for a sixth of the cost.

Is GPT-5.5 Pro better than Claude Fable 5?

On published coding benchmarks, no: Claude Fable 5 beats base GPT-5.5 on SWE-Bench Pro (80.3% vs 58.6%), and OpenAI never published a Pro coding score to compare. GPT-5.5 Pro wins on browsing and frontier math. One catch: Fable 5 is currently suspended under a US export-control directive, so availability matters as much as the benchmark.

How good is GPT-5.5 Pro at coding?

Honestly, we can't say from OpenAI's numbers, because OpenAI did not publish a GPT-5.5 Pro coding score. The only coding figure for the family is base GPT-5.5's SWE-Bench Pro result of 58.6%, which trails Claude Fable 5's 80.3%. Anyone quoting a "Pro coding benchmark" is likely quoting the base number by mistake.

GPT-5.5 Pro vs GPT-5.5 standard — which should I use?

Use GPT-5.5 Pro for hard research, frontier math and deep multi-step browsing, where its extra parallel compute earns the 6× price. Use base GPT-5.5 for daily work, code review and summaries, where the premium buys little. On some tests, like GDPval, OpenAI's table even lists Pro slightly below base.

How much does GPT-5.5 Pro cost?

GPT-5.5 Pro costs $30 per million input tokens and $180 per million output tokens, roughly 6× base GPT-5.5's $5/$30. For comparison, Claude Fable 5 sits at $10/$50. On output-heavy workloads the Pro premium compounds fast, so routing by task difficulty rather than defaulting to Pro saves real money.

What is reasoning effort "xhigh"?

Reasoning effort is the dial that controls how hard GPT-5.5 thinks before answering. "xhigh" is the top tier, where the model uses parallel test-time compute to explore several reasoning paths and pick the best. OpenAI ran its published GPT-5.5 Pro evals at xhigh in a research environment, which differs from the production ChatGPT default.

Is GPT-5.5 Pro available in Codex?

Yes. You can call GPT-5.5 Pro in Codex CLI using the model string gpt-5.5-pro, and set reasoning effort to xhigh for the highest-compute behavior. Expect higher latency and noticeably higher token cost than base GPT-5.5, so reserve it for the genuinely hard tasks rather than running it as your default coding model.

Is GPT-5.5 Pro a new model?

No. GPT-5.5 Pro is the same GPT-5.5 model running with more reasoning effort and parallel test-time compute at the xhigh setting, not a separate architecture. It shares the same weights and the same context window of roughly 1.1M input and 128K output tokens. You're paying for more thinking time, not a smarter base model.

What is GPT-5.5 Pro's context window?

GPT-5.5 Pro carries a context window of about 1.1M tokens of input and 128K tokens of output, the same as base GPT-5.5. That's enough to load large codebases or long documents in a single call. The difference between Pro and base isn't memory size; it's how much reasoning compute the model spends before it answers.