Models · Which model

Which model should I use?

A plain read of the public boards, plus a short speedrun of one builder’s advice. Boards measure different things, so read each card before you pick.

Today’s read

Checked 22 Nov 2026

These lines are generated from the boards below, not written by hand. Leaders count only the benchmark and leaderboard boards.

Speedrun: Theo’s video

Stop Picking The Wrong Model · Theo - t3.gg · 13 Nov 2026 · 42:19 · Watch on YouTube

The short version. For coding work, stop shopping around: pick between Claude Opus 5.5 (the careful default) and Claude Fable 5.2 (the bolder one for hard, creative or stuck problems), and run them on the biggest Claude plan you can afford.

What he says, with timestamps

  1. 00:00Theo says there are many strong coding models now, but unless you have a very specific reason, he thinks the real choice is between Opus 5.5 and Fable 5.2. He calls everything else a distraction.
  2. 03:03His budget advice: the most expensive Claude plan you can afford gives the most work per dollar, because plan usage is subsidised. He says cheaper per-token models are not cheaper once you count what a plan gives you.
  3. 06:07He walks through Anthropic's own model guide, which says to start most work on Opus 5.5 at its default medium effort. He mostly agrees, though he finds medium a little quick to stop checking.
  4. 07:40On effort levels, his ranking is: high is the best balance, medium and extra-high are both good, low is poor and max is not worth using. He says max forces too much thinking even on simple tasks, and that ultracode is now a toggle.
  5. 10:14He would not pick Sonnet 5.5 in the model picker. On open-ended tasks it writes so much that it can cost about as much as Opus, so he sees it as a tool for scoped jobs. Haiku 5.5 is fine for bulk pattern-matching across lots of data, but not for writing code.
  6. 16:26On the $200 plan, only half of your weekly allowance can go to Fable. By his own chart, that works out to a fraction of the Opus usage you would get, so Fable's lower cost per task does not decide the matter.
  7. 18:01Many people assume Fable is slower than Opus. In his own daily use, Fable 5.2 runs at about the same speed as Opus 5.5, so he says speed is not a reason to choose one over the other.
  8. 21:04His picture of Opus 5.5 is the teacher's pet: very thorough, rarely misses anything, but keen to please. It lets scope grow and seldom pushes back. His fix is to add a line to prompts: do it if you think it is good, otherwise suggest something else.
  9. 32:32He compares the two to two senior colleagues. Opus is the careful long-timer you ask when a change must be perfect. Fable is the experienced newcomer who glances at it, flags what really matters and says ship it.
  10. 35:41His simplest rule: if the work can be checked objectively, such as a rewrite with tests, Opus is enough. For creative or open-ended work, or an old thread that keeps going in circles, try Fable. Otherwise stay on Opus until you want more from it.

His verdict, model by model

ModelUse forAvoid for
Claude Opus 5.5at 21:04Most coding work, at medium or high effort. Best where the result can be verified, such as big migrations and rewrites, and where thoroughness matters.Open-ended or creative problems where you want unusual ideas. It tends to over-build and rarely pushes back unless you tell it to.
Claude Fable 5.2at 35:41The hardest problems, creative or exploratory work, and old threads that keep going in circles and need a decision.Work that needs every small edge case covered, such as building a test setup from scratch. It can let minor issues slip.
Claude Sonnet 5.5at 10:14Well-scoped jobs, mainly as a tool that your agents call.Being your default in the model picker on open-ended tasks, where it can cost about as much as Opus.
Claude Haiku 5.5at 11:17Bulk analysis, such as pattern-matching across many files or a lot of data.Writing code.

Source: our speedrun of Theo’s video, summarised in plain words. These are his opinions after about two days of use, not our tests.

Sites he showed

His own numbers

For Indian builders. The plans he discusses are the $100 and $200 a month Claude tiers, priced in US dollars, so an Indian builder should check the rupee cost on their own card before copying his advice. His case rests on heavy daily coding; if you use far less, test whether a smaller plan covers your work first.
Caveats
  • Captions are auto-generated; model names were matched to our tracked list and unclear ones were left out.
  • Evalboard is private, so what he shows is described as seen on screen, without figures.
  • Numbers he reads from Artificial Analysis, Priors and Slopalytics charts are left out; those charts are described in words.
  • The Browserbase sponsor segment, his Twitch chat, the speed comparison for Fable 5.1 and the 'Sonnet versus Opus' cost passage near 10:46 were left out because the captions were unclear or the figures did not match the screen.
  • His verdicts come from about two days of use with Fable 5.2, as he says himself.

The boards

One card per source. Each board is its owner’s own ranking, so we never blend them into one score.

Benchmarks and leaderboards

Vibe Code Bench v1.1

Vals AI

Measures: How well a model, working inside the OpenHands coding agent, builds a working web app from scratch.

How to read it: Higher is better. 93% means the apps it built passed about 93 of every 100 automated checks.

Top 12 on 22 Nov 2026
#ModelAccuracy
1Claude Fable 5.293.8%
2google/cobalt-c-high-1p0093.7%
3Claude Sonnet 5.592.4%
4Gemini 4 Argon91.9%
5Claude Haiku 5.590.4%
6Claude Fable 590.3%
7Claude Opus 5.590.3%
8Claude Fable 5.190.3%
9GPT-6.2 Astra89.9%
10GPT-6 Astra89.6%
11Gemini 4 Carbon89.1%
12GPT-6.1 Sol88.9%

Source: Vals AI. Top 12 shown.

Checked 22 Nov 2026 · The source’s own date: 19 Nov 2026 · CC BY 4.0 (declared on the page). Credit: Vals AI.

Vals Index

Vals AI

Measures: One combined score for agent work on finance, coding, legal and tax tasks, weighted by each sector's share of the US economy.

How to read it: Higher is better. It says how well a model does office and coding work overall, not coding alone.

Top 12 on 22 Nov 2026
#ModelAccuracy
1Claude Fable 5.271.3%
2google/cobalt-c-high-1p0070.9%
3Claude Sonnet 5.570.8%
4Gemini 4 Carbon70.5%
5Claude Opus 5.570.1%
6Gemini 4 Argon69.4%
7Gemini 4 Hydrogen68.5%
8GPT-6.2 Astra67.7%
9Muse Wonder Full66.8%
10GPT-6.2 Sol66.1%
11GPT-6 Astra65.0%
12Muse Wonder64.9%

Source: Vals AI. Top 12 shown.

Checked 22 Nov 2026 · The source’s own date: 19 Nov 2026 · CC BY 4.0 (declared on the page). Credit: Vals AI.

LiveBench 2.0

LiveBench team (Abacus.AI and academic collaborators)

Measures: Long work tasks (agent work, coding, browsing, data analysis, research, design, reasoning) that are kept fresh so models cannot have seen them, graded on what the model hands back.

How to read it: Higher is better. Scores run 0 to 100. The overall number weights each category the way the site does (long agentic work 2, coding 1.5, design 1.5, others 1).

Top 5 on 22 Nov 2026
#ModelOverall (site's category weights)
1Gemini 4 Neoncoding 89.787.4
2Claude Fable 5.2coding 79.284.9
2GPT-6.2 Astracoding 91.984.9
4GPT-6 Astracoding 87.584.2
5Claude Fable 5.1coding 89.784.0

Source: LiveBench team (Abacus.AI and academic collaborators). Top 5 shown.

Checked 22 Nov 2026 · The source’s own date: 6 Nov 2026 · No licence stated. Top 5 rows with attribution and a link.

Terminal-Bench 4.1

Terminal-Bench team (Laude Institute)

Measures: How often a coding agent finishes hard command-line tasks (software, machine learning, security, operations) in a sandbox.

How to read it: Higher is better. 65% means the agent finished about 2 of every 3 tasks. Each run pairs a model with an agent such as Claude Code or Codex.

Top 5 on 22 Nov 2026
#ModelAccuracy
1Copilot FusionGitHub Copilot, run 2026-11-1771.2%
2Claude Opus 5.5Claude Code, run 2026-09-2265.8%
2Claude Fable 5.2Claude Code, run 2026-11-0965.8%
4Gemini 4 CarbonAntigravity CLI, run 2026-10-2864.5%
5Claude Sonnet 5.5Claude Code, run 2026-09-2863.9%

Source: Terminal-Bench team (Laude Institute). Top 5 shown.

Checked 22 Nov 2026 · The source’s own date: 20 Nov 2026 · No licence stated. Top 5 rows with attribution and a link.

ARC Prize (v3 semi-private)

ARC Prize Foundation

Measures: How well a model solves new puzzles it has not seen (fluid reasoning), scored by ARC Prize on a held-out set. Cost is also reported.

How to read it: Higher is better. 81% means it solved about 8 of 10 unseen puzzles. Check the cost too, because top scores can be expensive to run.

Top 5 on 22 Nov 2026
#ModelScore
1Claude Fable 5.2cost $21,948.8381.2%
2GPT-6.2 Astracost $24,168.9170.3%
3Gemini 4 Carboncost $7,244.2865.1%
4GPT-6 Astracost $26,097.5062.7%
5GPT-6.2 Solcost $5,466.2154.9%

Source: ARC Prize Foundation. Top 5 shown.

Checked 22 Nov 2026 · No licence stated. Top 5 rows with attribution and a link.

SWE Atlas (Scale)

Scale AI

Measures: How well coding agents understand a codebase, write tests and refactor, graded against expert checklists.

How to read it: Higher is better. 80% means the agent met about 8 of 10 expert criteria. Close ranks are ties because each score has a margin of error.

Top 5 on 22 Nov 2026

Source: Scale AI. Top 5 shown.

Checked 22 Nov 2026 · No licence stated. Top 5 rows with attribution and a link.

SWE-Bench Pro V2 (Scale)

Scale AI

Measures: The share of long, multi-file software tasks from real open-source projects that an agent resolves.

How to read it: Higher is better. Each row names the agent it ran in, so rows are not strictly like-for-like. Scores near the top may be close to the ceiling.

Top 5 on 22 Nov 2026
#ModelResolve rate (default tab)
1Claude Opus 598.0%
2Claude Fable 5.192.2%
3Gemini 4 Argon92.2%
4GPT-6 Astra90.2%
4Muse Wonder 1.0 (mini-swe-agent) xhigh90.2%

Source: Scale AI. Top 5 shown.

Checked 22 Nov 2026 · No licence stated. Top 5 rows with attribution and a link.

MCP-Atlas (Scale)

Scale AI

Measures: How reliably a model uses real tools through the Model Context Protocol across multi-step tasks (36 servers, 220 tools).

How to read it: Higher is better. 90% means it finished about 9 in 10 tool-use tasks correctly.

Top 5 on 22 Nov 2026

Source: Scale AI. Top 5 shown.

Checked 22 Nov 2026 · No licence stated. Top 5 rows with attribution and a link.

SimpleBench

SimpleBench team (AI Explained)

Measures: Everyday common-sense and trick questions where an ordinary person beats most models.

How to read it: Higher is better. The human baseline is 83.7%, so a model above that beats an average person on these questions.

Top 5 on 22 Nov 2026
#ModelScore (average of 5 runs)
1Claude Fable 5.294.9%
2Gemini 4 Neon94.4%
3Gemini 4 Carbon92.6%
4Gemini 4 Argon91.7%
4Gemini 4 Hydrogen91.7%

Source: SimpleBench team (AI Explained). Top 5 shown.

Checked 22 Nov 2026 · The source’s own date: 19 Nov 2026 · No licence stated. Top 5 rows with attribution and a link.

Vending-Bench 2

Andon Labs

Measures: Whether an AI agent can run a simulated vending-machine business for a year without losing the plot, scored by final bank balance.

How to read it: Higher is better. The number is the dollar balance after a simulated year, averaged over runs.

Top 5 on 22 Nov 2026
#ModelAverage bank balance
1Gemini 4 Neon$16,843
2Gemini 4 Carbon$16,484
3GPT-6 Astra$15,515
4Gemini 4 Hydrogen$14,549
5GPT-6 Sol$14,428

Source: Andon Labs. Top 5 shown.

Checked 22 Nov 2026 · No licence stated. Top 5 rows with attribution and a link.

Artificial Analysis

Artificial Analysis

In Theo’s video at 18:38, 18:48, 19:00

Measures: Independent benchmarking of AI models and providers on intelligence, price, speed and latency.

How to read it: Use its model pages to compare intelligence, price and speed per provider. Read the numbers on its own site.

No numbers here. Its terms bar copying and republishing, so this site carries none of its numbers. We link to it instead.

Source: Artificial Analysis.

Listed 22 Nov 2026 · link only, not collected

DesignArena

Arcada Labs

Measures: Crowd-voted head-to-head preference for websites, apps and games built by AI coding agents.

How to read it: Higher Elo is better. A 100-point gap means the higher model wins about 64 of 100 votes. Models with few votes are less certain.

No numbers here. No reuse licence, so link only. We link to it instead.

Source: Arcada Labs.

Listed 22 Nov 2026 · link only, not collected

SkateBench

Theo (t3.gg)

Measures: How well models name skateboarding tricks from their technical definitions.

How to read it: Higher is better, but it is a fun knowledge test, not a coding or work benchmark.

No numbers here. Stale: no current models are listed, so it is linked only. We link to it instead.

Source: Theo (t3.gg).

Listed 22 Nov 2026 · link only, not collected

Evalboard

Theo (t3.gg), per the video

In Theo’s video at 39:46, 40:00, 40:20 and 2 more

Measures: Theo's own coding-model tests combined into a mean standing, where 100 means best on every test.

How to read it: A higher mean standing means a model finished nearer the top across his tests. Look at each test for where a model is weak.

Private, so no numbers and no link. This site needs a sign-in. We collected nothing from it. His own standings across four internal benchmarks, where Opus 5.5 and Sonnet 5.5 sit ahead of Fable 5.2.

Source: Theo’s video, as shown on screen.

Listed 22 Nov 2026 · from the video only

Usage

Vercel AI Gateway leaderboard

Vercel

Measures: Each model's share of all token volume sent through Vercel's AI Gateway, with spend, request and team-reach shares for the same day.

How to read it: Shares, not totals. Token share shows what is used most, spend share shows where the money goes, and reach shows how many teams touch a model.

Top 9 on 22 Nov 2026
#ModelShare of token volume
1Beam 1.021.0%
2DeepSeek V4.1 Flash 101719.9%
3GLM-5.311.3%
4GLM-5.3-Flash8.4%
5Kimi K34.9%
6Claude Opus 5.54.8%
7GPT-6 Luna3.8%
8Claude Sonnet 5.52.6%
9DeepSeek V4.1 Flash2.1%

Source: Vercel. Top 9 shown.

Checked 22 Nov 2026 · The source’s own date: 21 Nov 2026 · Vercel publishes this as anonymised data updated daily. Credit: Vercel.

Slopalytics usage

Slopalytics (T3 Code telemetry)

In Theo’s video at 10:26, 10:48, 11:00

Measures: Which models people actually send turns to inside T3 Code, a coding app, on the latest day.

How to read it: Numbers are turns per 100,000 turns that day, so they show share of use among T3 Code users, not quality.

Top 5 on 22 Nov 2026
#ModelTurns per 100,000 turns
1Claude Opus 5.556,223 per 100k turns
2Claude Fable 5.28,681 per 100k turns
3GPT-6.1 Sol4,246 per 100k turns
4Claude Sonnet 5.54,037 per 100k turns
5GPT-6 Astra1,015 per 100k turns

Source: Slopalytics (T3 Code telemetry). Top 5 shown.

Checked 22 Nov 2026 · The source’s own date: 20 Nov 2026 · No licence stated. Only the usage file is used. Top 5 rows with attribution and a link.

Hugging Face downloads

Hugging Face

Measures: Downloads over the last 30 days for open-weight text models hosted on Hugging Face.

How to read it: Shows which open models people pull most. It says nothing about closed models such as Claude or GPT, and small old models stay high.

Top 5 on 22 Nov 2026
#ModelDownloads (30 days)
1openai-community/gpt24,047 likes34,152,708
2Qwen/Qwen3-0.6B1,926 likes15,314,218
3Qwen/Qwen3-14B551 likes13,131,010
4deepseek-ai/DeepSeek-V3.21,610 likes10,707,628
5Qwen/Qwen2.5-7B-Instruct2,505 likes6,410,779

Source: Hugging Face. Top 5 shown.

Checked 22 Nov 2026 · The source’s own date: 22 Nov 2026 · No licence stated. Top 5 rows with attribution and a link.

OpenRouter rankings

OpenRouter

Measures: Tokens sent through OpenRouter to each model, per day, with weekly and category views.

How to read it: Bigger volume means more builders route traffic there. Cheap and free models inflate raw volume, so read it as popularity, not quality.

No numbers here. Its terms bar scraping, so this site does not collect it. We link to it instead.

Source: OpenRouter.

Listed 22 Nov 2026 · link only, not collected

Speed

Vercel AI Gateway model speed

Vercel

Measures: Live output speed (tokens per second) and time to first token for each model, measured on real AI Gateway requests over the last hour.

How to read it: Higher tokens per second means text appears faster. Lower first-token time means the answer starts sooner. The same model can differ a lot between the maker and a cloud reseller.

Top 5 on 22 Nov 2026
#ModelOutput speed (fastest provider, median)
1Inkling 2 Smallvia deepinfra, first token 309 ms575.5 tokens/s
2DeepSeek V4.1 Flashvia mistral, first token 727 ms443.0 tokens/s
3Muse Embervia meta, first token 1667 ms418.5 tokens/s
4Muse Spark Openvia baseten, first token 680.5 ms321.0 tokens/s
5Claude Fable 5.2via anthropic, first token 2036.5 ms306.0 tokens/s

Source: Vercel. Top 5 shown.

Checked 22 Nov 2026 · The source’s own date: 22 Nov 2026 · No licence stated. Top 5 rows with attribution and a link.

Cost and limits

Claude plans and limits

Anthropic

Measures: Anthropic's price and usage allowance for the Free, Pro, Max and Ultra plans, stated as multiples of Pro.

How to read it: Pro is the baseline. Max gives 5 times Pro usage and Ultra 10 times Pro weekly. Real limits depend on the model and how long your chats are.

Top 4 on 22 Nov 2026
PlanMonthly price (billed monthly)
Freeweekly limit Limited$0/month
Pro$17/month on annual billing; weekly limit 1x$20/month
Maxweekly limit 5x$100/month
Ultraweekly limit 10x$200/month

Source: Anthropic. Top 4 shown.

Checked 22 Nov 2026 · No licence stated. Plan table with attribution and a link.

Priors

Adam Holter

In Theo’s video at 03:50, 04:20, 05:00 and 1 more

Measures: Ranks models using Artificial Analysis scores and prices each model's cost per task the way a subscriber pays, using the share of a plan's allowance.

How to read it: On its cost chart, models toward the top-left are smarter and cheaper per task. The numbers come from Artificial Analysis.

No numbers here. Core numbers are Artificial Analysis data, so this site carries none of them. We link to it instead.

Source: Adam Holter.

Listed 22 Nov 2026 · link only, not collected

Slopalytics benchmark charts

Slopalytics

In Theo’s video at 10:26, 10:48, 11:00

Measures: Per-model, per-effort-level intelligence, cost per task, output tokens per task and speed.

How to read it: Higher bars on tokens per task mean a model talks and bills more to finish the same job, so compare cost per task, not just price per token.

No numbers here. The footer says the data comes from Artificial Analysis, so this site carries none of its numbers. We link to it instead.

Source: Slopalytics.

Listed 22 Nov 2026 · link only, not collected

ChatGPT plans

OpenAI

Measures: OpenAI's plans and usage tiers for ChatGPT.

How to read it: Compare plan tiers and Codex usage on the page itself.

No numbers here. Prices are drawn by JavaScript and cannot be read reliably, so this is a link only. We link to it instead.

Source: OpenAI.

Listed 22 Nov 2026 · link only, not collected

Anthropic’s own guide

Which Claude model should you build with? · Anthropic · published 9 Nov 2026 · Theo walks through it from 06:20

The short version

Which model for which task

TaskStart withWhy
Everyday engineering, debugging, multi-PR feature workOpus 5.5 (medium, then high)Best balance of cost and capability for routine work.
Classification, routing, tagging, high-volume extractionHaiku 5.5 (medium)Cheapest and fastest in the family.
Subagents that search files, read logs or summariseHaiku 5.5 or Sonnet 5.5 (default effort)Mistakes are cheap to spot and the main model makes the real judgement calls.
Chat and interactive sessionsSonnet 5.5 (low/medium), Haiku 5.5 for simple volume chatSonnet is quick at low and medium effort.
Well-scoped coding: bug fixes, feature iteration, spec-driven workSonnet 5.5 (medium, high if harder)Handles scoped tasks well at half Opus 5.5's token price.
Everyday documents, slides, spreadsheetsSonnet 5.5 (high; medium in Claude Code)Iterates fast with good design sense; start on Opus 5.5 for complex financial models or big source piles.
Repeated agent jobs such as investigation, review, draftingSonnet 5.5 (medium)Well-defined repeat tasks rarely need a bigger model.
Computer use and vision-heavy workflowsOpus 5.5 (medium)Long agent loops over many screenshots suit Opus.
Major refactors, migrations, architecture planning, leading many subagentsFable 5.2 (medium, then high; API default high)Most capable generally available model for technical fields.
Deep research, scientific or quantitative analysisFable 5.2 (medium, then high)Built for the hardest coding, finance and science projects.
Not sure yetOpus 5.5 (medium)Default in Claude Code and the platform docs' recommended starting point.

Source: Anthropic’s guide, linked above. This is our paraphrase, not a copy. Vendor guide. Paraphrased with a link; no long quotes. It covers Claude models only.

How we use these sources

Full: 3 sources

The source states an open licence. We show up to 12 rows, with attribution and the licence named.

Snippet: 12 sources

No licence is stated, but the page allows robots and its terms do not ban reuse. We show the top 5 rows, the date and a link.

Link: 7 sources

Their terms bar copying, or the data is derived from someone else’s, or the page is out of date. We name them and link. No numbers.

Video only and guide

One private site, described only as seen in Theo’s video. One vendor guide, paraphrased with a link.

Artificial Analysis, Priors and Slopalytics’ benchmark charts are linked but not copied. Priors and Slopalytics draw on Artificial Analysis data, and its terms do not allow republishing. Open one of those cards to read what its chart shows, then go to the source for the numbers.

Boards are read at most once a day. A board that fails to load keeps its last good rows and says so.