I spent the better part of a week putting Gemini 3.5 Flash, GPT-5.5, and Claude (Opus 4.8) through 10 tasks I actually do — writing, research, coding, summarizing, image analysis, and more. No cherry-picked benchmark screenshots. No sponsorships. Just the raw results.
Here's what I found.
First, Who Are These Contenders?
Before we get into the tests, let's lay the groundwork because these three models launched within five weeks of each other — and the pace of this race is genuinely dizzying.
GPT-5.5 (OpenAI) — Launched April 23, 2026. OpenAI's first fully retrained base model since GPT-4.5. Built specifically for agentic and coding tasks. Priced at $5/million input tokens and $30/million output tokens.
Gemini 3.5 Flash (Google) — Dropped May 19, 2026 at Google I/O. The wildcard. A Flash-tier model that somehow beats Pro-tier models on several benchmarks. Priced at $1.50/$9 per million tokens — making it dramatically cheaper than both rivals.
Claude Opus 4.8 (Anthropic) — Released May 28, 2026. Anthropic's best publicly available model. Same price as before: $5/$25 per million tokens. Still the coding benchmark king.
Now let's get into it.
The Speed Numbers First (Because They're Insane)
Let me just put this out there before the task tests, because it changes how you think about everything else.
Tokens per second:
- 🟡 Gemini 3.5 Flash: ~284 tokens/second
- 🟠 GPT-5.5: ~90 tokens/second
- 🔵 Claude Opus 4.8: ~60–70 tokens/second
Gemini 3.5 Flash is generating output roughly 3x faster than GPT-5.5 and nearly 5x faster than Claude. In a real session, that's the difference between "instant" and "watching a cursor blink."
For interactive use — quick drafts, rapid back-and-forth — this speed gap is genuinely felt. For deep work where you submit a prompt and walk away for coffee? It matters a lot less.
Keep that in mind as we go through the tasks.
The 10-Task Showdown
Task 1: Write a Marketing Email for a Small Business 🔵 Claude Wins
I gave all three the same brief: write a promotional email for a Denver-based coffee shop's summer loyalty program. Conversational tone, ~200 words.
Claude's result felt like it was written by an actual human copywriter. Warm, specific, not trying too hard.
GPT-5.5's version was clean and competent — but had that slight "AI wrote this" polish that makes you want to edit the second paragraph.
Gemini's came back blazing fast but read a bit generic. It hit all the points, but it lacked personality.
Winner: Claude — by a clear margin for writing quality. This tracks with what multiple reviewers have found: Claude's output reads more naturally and requires less editing.
Task 2: Debug a Python Function 🟠 GPT-5.5 Wins (Barely)
I fed all three a buggy Python script with a logic error in a nested loop. The kind of thing that makes you stare at your screen for 20 minutes.
GPT-5.5 found the bug immediately, explained exactly why it was broken, and offered three approaches to fix it — with clear tradeoffs for each.
Claude also caught it, but the explanation was longer and the fix choices were presented more like a lecture than a toolkit.
Gemini identified the issue correctly but produced a hallucinated comment in the fix — a small thing, but notable.
Winner: GPT-5.5 — for coding precision and practical explanations. Claude is close, but GPT-5.5 is built for this.
Task 3: Summarize a 50-Page PDF Report 🔵 Claude Wins
I uploaded a dense climate policy white paper and asked for a three-paragraph executive summary with the key action items pulled out.
Claude nailed this. Crisp, accurate, correctly identified the three main policy proposals. When I asked follow-up questions, it answered from the document, not from training data.
GPT-5.5 did well but subtly shifted one statistic — off by a small margin, but that matters in this type of document.
Gemini returned the summary faster, but missed one of the four key sections entirely.
Winner: Claude — its reliability with long documents is a genuine competitive advantage. And it's also the least likely to quietly make something up.
Task 4: Analyze an Image with Data (Chart Understanding) 🟡 Gemini Wins
I gave all three the same bar chart screenshot from a quarterly business report and asked for a written analysis.
Gemini 3.5 Flash dominated here — and this is exactly where Google's multimodal depth shows. It scored 84% on MMMU-Pro (the most comprehensive multimodal reasoning benchmark), the highest ever recorded on that test. You can feel that in practice.
GPT-5.5 read the chart correctly but offered less contextual interpretation.
Claude handled it, but Claude's image input is currently text-and-image only. Gemini accepts text, images, audio, and video.
Winner: Gemini 3.5 Flash — and it's not close. For image, video, and document understanding, this is Google's lane.
Task 5: Write a Long-Form Blog Post Intro (500 words) 🔵 Claude Wins
I asked all three to write the opening of an article about the rise of robotaxis in America. Target reader: curious adult, not a tech bro.
Claude's intro opened with a scene in Phoenix that felt like you were there. Smart hook, smooth transition into the data.
GPT-5.5 gave me a competent intro but opened with a statistic — the kind of thing an editor would push back on.
Gemini came back in about two seconds (seriously, it's fast), but the prose had that slightly encyclopedic tone that doesn't grip a general reader.
Winner: Claude — for any writing that'll live under a human's name, Claude is still the top choice.
Task 6: Research Question — Who Won the 2025 NBA Finals? 🟠 GPT-5.5 / 🟡 Gemini Tie
I asked all three this factual question with web search enabled.
Both GPT-5.5 and Gemini retrieved current information quickly and accurately.
Claude with web search also performed well, but its raw speed at returning the answer was the slowest of the three.
Winner: Tie (GPT-5.5 / Gemini) — all three got it right; Gemini was fastest.
Task 7: Build a Basic HTML Landing Page from a Description 🔵 Claude Wins
"Build me a simple landing page for a dog walking service. Include a hero section, pricing, and a contact form."
Claude's output was clean, semantic, and actually readable by a non-developer. The layout worked. The CSS was minimal and sensible.
GPT-5.5 produced more code, but it was over-engineered for the task — JavaScript included where none was needed.
Gemini produced fast output but had a structural HTML error in the form section that would break validation.
Winner: Claude — coding quality for front-end work, plus accuracy. Claude Opus 4.8 holds a 69.2% SWE-bench Pro score vs GPT-5.5's 58.6%, and that gap shows in complex builds.
Task 8: Explain a Complex Concept Simply (Quantum Computing) 🔵 Claude Wins (Narrowly)
Asked: "Explain quantum computing to a 15-year-old who's good at math."
All three did a solid job here, honestly. But Claude's explanation used a better analogy — comparing qubit superposition to a coin spinning in the air before it lands. Concrete, memorable, age-appropriate.
GPT-5.5 was technically accurate but slightly more textbook-y.
Gemini was fast and clear but felt like it was pulling from a Wikipedia summary.
Winner: Claude — just slightly more thoughtful in how it communicates.
Task 9: Agentic Multi-Step Task (Research + Summarize + Format) 🟡 Gemini Wins
This was the most revealing test. I gave each model a multi-step task: find recent data on EV adoption in the US, summarize it, and format it into a table with three columns.
Gemini 3.5 Flash was the clear winner here — and this is exactly what Google built it for. Agentic benchmarks show Gemini leading on MCP Atlas (83.6% vs GPT-5.5's 78.2%) and Finance Agent v2 (57.9% vs 43%). In practice, it chained the steps together smoothly and produced the table without me having to re-prompt.
GPT-5.5 handled it but needed a follow-up nudge to complete the formatting.
Claude completed it but was slowest, and its table formatting needed light cleanup.
Winner: Gemini 3.5 Flash — for multi-step agentic workflows, it's now the default recommendation. The speed compound effect across 20+ sequential steps is massive.
Task 10: Spot the Hidden Risk in a Contract Clause 🔵 Claude Wins
I pasted a vendor contract clause with a buried auto-renewal provision and asked each model to flag any risks.
Claude caught it immediately — and flagged two secondary concerns I hadn't even thought of, including jurisdiction ambiguity in the dispute resolution section.
GPT-5.5 also caught the auto-renewal but missed the jurisdiction issue.
Gemini flagged auto-renewal but added a suggested fix that was legally inaccurate — a notable hallucination on a task where accuracy really matters.
Winner: Claude — its hallucination rate is significantly lower (around 36% vs Gemini's ~61% and GPT-5.5's ~86% on the AA-Omniscience benchmark). For anything where being wrong has real consequences, that gap isn't a footnote. It's the whole decision.
The Scorecard
| Task | Winner |
|---|---|
| Marketing Email | 🔵 Claude |
| Python Debugging | 🟠 GPT-5.5 |
| PDF Summarization | 🔵 Claude |
| Image/Chart Analysis | 🟡 Gemini 3.5 Flash |
| Blog Post Writing | 🔵 Claude |
| Factual Research | 🔵🟡 Tie |
| HTML Landing Page | 🔵 Claude |
| Concept Explanation | 🔵 Claude |
| Multi-Step Agentic Task | 🟡 Gemini 3.5 Flash |
| Contract Risk Analysis | 🔵 Claude |
Final tally: Claude 6.5 / Gemini 2 / GPT-5.5 1.5
The Part Everyone Skips: Hallucination Rates
Most comparison articles don't put this front and center. I'm going to.
Hallucination rates (AA-Omniscience benchmark):
- Claude Opus 4.7/4.8: ~36%
- Gemini 3.5 Flash: ~61%
- GPT-5.5: ~86%
Here's the ugly truth about GPT-5.5: it's simultaneously the most knowledgeable model AND the most likely to confidently make something up. When it doesn't know an answer, it tends to answer anyway — with the same confident tone it uses when it does know.
Gemini 3.5 Flash sits in the middle, but a 61% hallucination rate in production code without human review is a real operational risk.
Claude hallucinates too — but it's more likely to say "I'm not certain about this" than to invent a plausible-sounding lie.
For legal, financial, medical, or any high-stakes writing: the hallucination gap is the whole ballgame.
The Pricing Reality Check
Here's where Gemini 3.5 Flash makes a genuinely compelling case, especially for builders and API users:
| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| Gemini 3.5 Flash | $1.50 | $9.00 |
| Claude Opus 4.8 | $5.00 | $25.00 |
| GPT-5.5 | $5.00 | $30.00 |
Gemini is 3.3x cheaper than GPT-5.5 and roughly 3x cheaper than Claude on input. At high volume — running thousands of API calls — that difference reshapes your entire budget.
So Who Should You Actually Use?
After 10 tasks and a lot of back-and-forth, here's the honest breakdown:
Use Claude if you write for a living, need high accuracy on complex documents, do legal or financial analysis, or just want the output that needs the least editing.
Use Gemini 3.5 Flash if you're a developer running high-volume agentic pipelines, working with images or video, need raw speed, or are cost-sensitive at scale.
Use GPT-5.5 if your work lives in terminals, you do complex multi-step coding with Codex, or you need the strongest pure reasoning on abstract problems (it still leads on ARC-AGI-2 with an 84.6% score).
The smartest teams in 2026 aren't picking one model. They're routing tasks — Claude for writing and analysis, Gemini Flash for speed-critical pipelines, GPT-5.5 for terminal-heavy engineering. That's not a cop-out; that's just good engineering.
The Bottom Line
Gemini 3.5 Flash is a genuinely disruptive release. A Flash-tier model beating Pro-tier flagships on agentic benchmarks at a fraction of the price — that wasn't supposed to happen. Google quietly changed the routing math for the entire industry.
But "fastest" doesn't mean "best for everything." Claude is still the most reliable tool for writing, accuracy, and anything where a hallucinated answer does real damage. And GPT-5.5 remains the terminal coding champ.
The AI race isn't over. Gemini 3.5 Pro is still coming. And if Flash is any indication of what Google's Pro tier will look like — the next few months are going to be very interesting.
Want more no-BS AI comparisons? Follow Ampick for the latest — ampick.xyz
0 Comments