Claude vs GPT vs Gemini for product work
By Imran Sheikh · 2026-11-12 · Models and tools · 5 min read
Key points
- All three flagships are close on broad benchmarks this month.
- Claude leads coding and web build rankings; Gemini leads general chat.
- Pick by task and price, then keep a second model on hand.
Every few weeks someone asks us which model is best. The honest answer changes faster than most product roadmaps. So this post is a snapshot, as of November 2026, of the three flagship lines we use: Anthropic’s Claude, OpenAI’s GPT and Google’s Gemini.
We look at them the way a product team would. What does each cost, how much can it read at once, what do the public rankings say, and which one do we reach for when we have a real task on the table. We build apps with AI coding agents every day, so the opinions here come from use. The numbers come from the sources listed at the end.
The current flagships
Each company now ships a family of models rather than one. The flagship is the one each vendor points you to for hard work. Anthropic’s docs suggest starting with Claude Opus 5.5 and moving up to Claude Fable 5.2 for the most demanding reasoning and long agentic runs. OpenAI’s docs say to start with GPT-6 Astra if you’re unsure, and point to GPT-6.1 Sol for a balance of quality and cost. Google lists Gemini 4 Carbon as its most capable reasoning model, with Gemini 4 Argon as the more efficient sibling.
| Model | Input / output per 1M tokens | Context window | Max output |
|---|---|---|---|
| Claude Opus 5.5 | $4 / $20 | 1M tokens | 128K tokens |
| Claude Fable 5.2 | $8 / $40 | 1M tokens | 128K tokens |
| GPT-6 Astra | $10 / $50 (up to 272K input) | 1.05M tokens | 128K tokens |
| GPT-6.1 Sol | $2 / $10 (up to 272K input) | 1.05M tokens | 128K tokens |
| Gemini 4 Carbon | $2 / $10 until Nov 30, then $4 / $20 | 1M tokens | See Google’s docs |
| Gemini 4 Argon | $2 / $10 until Nov 30, then $4 / $20 | 1M tokens | See Google’s docs |
Two pricing details matter more than they look. OpenAI charges more once a prompt passes 272K input tokens. GPT-6 Astra goes to $20 input and $75 output per million in that range. Google’s Gemini 4 Carbon and Argon prices double on December 1, 2026. If you budget a feature this month, budget it at the December rate.
What the public rankings say
No single benchmark tells you which model is best for your product. Two public sources are still worth a look because they measure different things.
LMArena (now at arena.ai) ranks models from blind, side-by-side human votes. On the text leaderboard dated November 11, 2026, Gemini 4 Carbon sits first with a preliminary score of 1526. Gemini 4 Argon is second at 1522 and Claude Opus 5.5 is fourth at 1507. The WebDev leaderboard, where people vote on generated web apps, looks different. Claude Fable 5.2 leads at 1838, Claude Opus 5.5 is second at 1815, Gemini 4 Carbon is fourth at 1768 and GPT-6 Astra fifth at 1764.
Artificial Analysis runs its own test suite and reports an Intelligence Index for each model at each effort setting. At their top settings, as of November 2026, Claude Fable 5.2 scores 65, Gemini 4 Carbon 60, Claude Opus 5.5 59 and GPT-6 Astra 57. The gaps between the top models are small. The gap between effort settings on the same model is often larger. Opus 5.5 scores 59 at max effort and 48 at medium.
The effort setting you choose can matter as much as the model you choose.
How they feel on real product tasks
Benchmarks set the shortlist. Daily use decides the pick. Here is how the three lines tend to sit for the work a product studio does. This is our opinion from regular use. It is not a scored test.
- Building and changing code across a real app: we reach for Claude first. It follows a codebase’s conventions well, and the WebDev ranking above lines up with what we see.
- Writing product copy, specs and research summaries: all three are good. We judge on tone. Claude and GPT tend to need less editing for our voice, but this is taste.
- Reading very long inputs, like a full transcript set or a large export: all three now take about a million tokens. Watch OpenAI’s price step above 272K input tokens.
- Images, video and mixed media inside one flow: Gemini is the natural fit, since Google ships image, video and live audio models on the same API.
- Second opinions on a tricky design or architecture call: we ask a second vendor’s model. Different training shows up as different blind spots.
Cost in practice
List price per million tokens is only part of the bill. Three things change what you actually pay.
First, effort and thinking. All three vendors now let the model think before it answers, and those thinking tokens are billed as output. A higher effort setting can raise quality and cost at the same time. Second, caching. Anthropic bills cache reads on Opus 5.5 at 5% of the base input price, and OpenAI lists a cached input price for every flagship. If your feature sends the same long system prompt each time, caching changes the math. Third, batch. Anthropic and OpenAI both take 50% off for batch requests that can wait.
For most in-app features you will not run a flagship at all. A mid-tier model like Claude Sonnet 5.5 or GPT-6.1 Sol, both at $2 input and $10 output per million, covers a lot of ground. We wrote a separate post on the fast, cheap tiers.
How we choose
We don’t pick one winner and stop thinking. We pick per job, and we re-check when a new model ships. Our routine is short.
- Name the task in one sentence, like “rewrite onboarding copy” or “add a settings screen to this React Native app”.
- Shortlist two models using the rankings that match the task: WebDev for building, text for writing and reasoning.
- Run both on five to ten real examples from the project. Read every output.
- Price the winner at the effort level that passed, with caching on, at next month’s rate.
- Keep the runner-up wired in as a fallback for outages and second opinions.
Right now that routine puts Claude at the center of our build work and keeps GPT and Gemini close by. Ask us again in a quarter. The ranking will have moved, and that’s fine. The routine still works.