The Best AI Model Right Now, by Job: Opus 5.5, Fable 5.2, GPT-6 Astra and Gemini 4 From the Same Builds

Claude Opus 5.5 is the best AI model right now for building apps, websites, games and videos. It placed first in 16 of the 21 builds Pat Simmons gave several models from identical prompts between September 23 and October 25, 2026, 18 of them ranked blind, and it placed second to the newer Fable 5.2 on all three builds of the Fable 5.2 review at about a quarter of the cost.

For everyday office work, Sonnet 5.5 got all six of AI for Mortals' office jobs right for about a fifth of what Fable 5.2 cost, and on Google's side Gemini 4 Hydrogen got all seven of our jobs right.

Last checked November 13, 2026: plan prices and which plan includes each model on Claude's plans page, ChatGPT's and Google's plan pages, and per-token prices on Anthropic's and Google's API pricing pages, plus the newest results from the Fable 5.2 and Gemini 4 Carbon reviews.

Which AI is best right now? The table by job

Opus 5.5 is the pick for building, Fable 5.2 for the hardest long builds when cost is no object, and Sonnet 5.5 or Gemini 4 Hydrogen for everyday office work. Each row links to the post with the full results.

The building results come from Pat's tests. The office-job results come from runs AI for Mortals did in Claude Code and through Google's API, and Pat's builds didn't cover writing, research or chat.

Here are the first places and the newest costs side by side:

Costs are API-equivalent: the builds ran on subscriptions, and the dollar figures are what the same tokens (the small chunks of text a model reads and writes) would cost at each company's published rates.

Is ChatGPT still the best AI?

For building apps, sites and games, ChatGPT's top model came second to Claude's in Pat's tests. GPT-6 Astra won one of the 21 builds, an Adidas running shoe modeled in Blender (the free 3D program), and Opus 5.5 placed above it in the other 20.

Astra still beat Google's model. In the four builds of Pat's Gemini 4 Argon review, Opus 5.5 placed first every time, Astra placed above Gemini 4 Argon every time, and Argon came last in all four (Claude vs ChatGPT vs Gemini on the same builds). In the newer Gemini 4 Carbon review, Carbon placed above Astra once, on a flight simulator game.

Which GPT-6 model you get depends on your plan. ChatGPT's pricing page lists GPT-6 Luna on Free and Go, adds GPT-6 Sol and GPT-6 Astra on Plus, and gives Pro expanded use of all three, with Pro's top reasoning option running on Astra. Pat tested Astra. He didn't run Sol or Luna head-to-head, so this page makes no call on them.

Which AI model is best for building apps, sites and games?

Opus 5.5. Across six head-to-heads, it won five of the six website rebuilds, every launch video and ad the models made in code, a playable theme-park game and a coffee-ordering iPhone app. Fable 5.1 won four builds, three of them mostly 3D, and cost the most: $1,613.14 for its 21 builds, against $977.57 for Opus 5.5 and $999.78 for GPT-6 Astra.

The 3D work is where Opus 5.5 pulled furthest ahead. When Pat revealed his blind ranking for a photoreal 3D coffee bottle you can open and pour, he said: "Opus just crushes 3D work."

Fable 5.2, which Anthropic released on November 9, is the only model that has placed above Opus 5.5 in Pat's builds since. Pat gave it three builds against Opus 5.5, Fable 5.1, GPT-6 Astra and Gemini 4 Argon: an interactive 3D history of Central Park with a narrated film, a rebuild of Shopify's Winter '26 Editions page, and a drivable 3D model of the new Tesla Roadster. He ranked Fable 5.2 first and Opus 5.5 second on all three, blind.

Fable 5.2 won by a small margin and at a much higher bill. Pat's verdict: "I don't think it's a sizable jump over Opus 5.5." The three Fable 5.2 builds cost $998.12 and ran 752 minutes, against $250.66 and 518 minutes for Opus 5.5, and Pat stopped each Fable 5.2 run once it reached its final polish.

Pat's earlier round with Opus 5.5, Fable 5.1 and GPT-6 Astra, which Opus 5.5 won, is below: scenes from Moby-Dick drawn in code, a launch film with every sound made in code, and a 3D skateboarding game.

Which AI is best for everyday work like email and spreadsheets?

The cheaper models did as well as the expensive ones. AI for Mortals gave Fable 5.2, Opus 5.5 and Sonnet 5.5 six office jobs, twice each: a long client email, a meeting write-up, a messy spreadsheet question, a project schedule with tight rules, a dense policy rewritten in plain language, and matching bank deposits to invoices. All three got all six right.

One full set of six cost $0.98 and took 3.2 minutes on Fable 5.2, $0.40 and 2.5 minutes on Opus 5.5, and $0.19 and 1.7 minutes on Sonnet 5.5, at API prices. Fable's one clear edge was length: it was the only model to stay under the word limit on the policy rewrite both times.

On Google's side, we gave Gemini 4 Hydrogen, released November 12, seven jobs: five office jobs and two with one provable right answer, a scheduling puzzle and a payroll function with eight planted bugs. Hydrogen got all seven right and was the fastest of the three Gemini 4 models on five. It was slower than Argon on cleaning up a 172-row contact list, and two of its four slide decks mislabeled one number.

Which plan gets you these models, from each company's plan page on November 13, 2026:

Google's prices and models are from its Google AI plans page. ChatGPT's pricing page didn't show dollar amounts when we checked, so its plans are listed by model in the ChatGPT section above.

Which AI model gives the most for the money?

Opus 5.5 won the most builds for about the same average cost as GPT-6 Astra. It averaged $46.55 a build across its 21 builds, against $47.61 for GPT-6 Astra and $76.82 for Fable 5.1, and it won 16 of them.

The cheapest models placed low. DeepSeek V4.1 Flash cost $1.00 to $1.43 a build and placed 3rd, 4th and 4th of five, and Kimi K3.1 Pro cost $3.46 to $17.69 and came last all three times. Gemini 4 Argon cost $39.54 across four builds against $189.04 for Opus 5.5 and came last in all four. Argon, Muse Wonder Full and Kimi K3.1 Pro came last in every build they ran.

Per-token price doesn't predict the bill on its own, because the number of steps a model takes moves it more. Fable 5.2 lists at twice Opus 5.5's price and cost about four times as much on the same three builds. These are the per-token prices on Anthropic's API pricing page and Google's Gemini API pricing page, checked November 13, 2026, in dollars per million tokens:

For office work, Sonnet 5.5 is the value pick: it matched Fable 5.2 on all six of our jobs for about a fifth of the cost.

Are Fable 5.2 and Gemini 4 Carbon worth switching to?

Keep Opus 5.5 as the everyday building model, bring Fable 5.2 in only for the biggest jobs, and skip Carbon for building, since it placed below Opus 5.5 in all three of its builds. Pat's advice on Fable 5.2: "My recommendation is to treat Fable 5.2 as an orchestrator," meaning a model that manages a group of Opus 5.5 agents doing the work, brought in only when you need it.

Fable 5.2 isn't on Claude's Free plan. Under Anthropic's Fable 5.2 launch offer, Pro and Team standard seats get it inside the plan through November 16, 2026 at 11:59 pm Pacific, for up to half the weekly limit, plus one Fable limit reset. After that, Claude's plans page lists Fable on Pro as paid with usage credits, while Max and Ultra include it for up to half the weekly limit.

Gemini 4 Carbon was the cheapest model in all three builds of Pat's Carbon review, $27.68 in total, but placed 4th, 4th and 3rd of four. Opus 5.5 ranked first on the launch film and the flight simulator, and Fable 5.2 first on a website clone. The totals were $883.30 for Fable 5.2, $181.29 for Opus 5.5 and $126.43 for GPT-6 Astra, and part of the Fable 5.2 flight simulator's log was lost, so its cost may be undercounted.

Pat said: "The one thing that has me considering Carbon for my day-to-day is the cost," and then: "I still absolutely prefer Opus 5.5." Carbon's $2 and $10 introductory price ends on November 30, and from December 1 it costs the same per token as Opus 5.5. Carbon needs Google AI Ultra in the Gemini app. The full results are in the Gemini 4 Carbon review.

How the models were ranked

Every model in Pat's builds got the same prompt, ran once with no follow-up prompts, and the main models worked in their own company's coding tool: Claude Code for the Claude models, Codex for GPT-6 Astra and, in the two newest reviews, Antigravity for Gemini. Pat ranked 18 of the 21 builds from September and October blind, before he knew which model made which, and he ranked all three Fable 5.2 review builds blind. The Gemini 4 Carbon review post doesn't say whether its rankings were blind.

Tokens, cost and minutes were logged for every run. The builds are big one-pass jobs in 3D, web design, games and video, and Pat calls them limited tests. They don't cover writing, research or chat.

The office-job results are AI for Mortals' own runs: the six Claude jobs ran in Claude Code with its tools switched off, and the seven Gemini jobs ran through Google's API, each with an answer we could check. Every prompt is published in the linked posts, and every model comparison with what each build cost is in one place.

Questions people ask

What is the best LLM right now?

Claude Opus 5.5, for most building work. An LLM (large language model) is the kind of AI model behind ChatGPT, Claude and Gemini. Fable 5.2 placed higher on three builds at about four times the cost.

What is the best AI model for coding?

Opus 5.5. It came first in 16 of Pat's 21 builds and never placed below 3rd. Claude's plans page lists Opus on Pro and up, and not on Free.

Is Claude better than ChatGPT?

For building, yes in Pat's tests: Opus 5.5 placed above GPT-6 Astra in 20 of 21 builds. Astra usually cost less, coming in under Opus 5.5 on 17 of the 21.

Is Gemini better than ChatGPT?

Not for building. GPT-6 Astra placed above Gemini 4 Argon in all four of their builds, while Argon cost about a tenth as much. Gemini 4 Carbon beat Astra once in three builds.

What is the cheapest good AI model?

For office work, Sonnet 5.5, which got all six of our jobs right at $0.19 a set. For builds, DeepSeek V4.1 Flash costs about a dollar a build but placed 3rd or 4th of five.

Is Fable 5.2 free?

No. It isn't on Claude's Free plan; Pro and Team standard seats include it only through November 16, 2026, and after that Pro pays for it with usage credits.

Which model does ChatGPT's free plan use?

GPT-6 Luna. ChatGPT's pricing page lists Luna on Free and Go, and GPT-6 Sol and GPT-6 Astra from Plus up.

Which Gemini model should I use?

Gemini 4 Hydrogen for everyday jobs: it got all seven of ours right and was usually fastest. Hydrogen comes with Google AI Plus at $4.99 a month.

What moved: update log

November 13, 2026: Page published. It covers Pat's 21 builds from September 23 to October 25, the Fable 5.2 review builds (Fable 5.2 first, Opus 5.5 second on all three) and the Gemini 4 Carbon review builds (Opus 5.5 first on two of three), plus AI for Mortals' office-job tests of Fable 5.2, Opus 5.5, Sonnet 5.5 and Gemini 4 Hydrogen. Plan and API prices checked November 13, 2026.

Sources

• Claude: plans and pricing (checked November 13, 2026)
• Anthropic: Claude API pricing (checked November 13, 2026)
• Claude Help Center: Claude Fable 5.2 launch offers (checked November 13, 2026)
• OpenAI: ChatGPT plans and pricing (checked November 13, 2026)
• Google: Google AI plans (checked November 13, 2026)
• Google: Gemini API pricing (checked November 13, 2026)