← The AI Scorecard: Which Model Is Best, by Whose Test
Compare versions
Version 1 → Version 2 (current). Removed · Added. This shows the entry's text as it is written, including its formatting marks.
Version 1 · Version 2 (current) · Version history
Side by side · Together
*Logos are trademarks of their owners, shown only to identify the companies; image files via [Wikimedia Commons](https://commons.wikimedia.org/wiki/Category:Logos_of_artificial_intelligence_companies).*
This is a living page. Every few weeks a Super Intelligence company announces a model and a table showing it beating everyone else, and every company tops its own table. The question a reader actually has, **which Super Intelligence model is best right now**, can't be answered from those tables. So this page answers it from the scoreboards the companies don’t run: the human-vote leaderboard LMArena, Artificial Analysis’s test battery, the ARC Prize puzzles, METR’s task-length measurements and Epoch AI’s trend work. I keep it current, write a separate entry when something big lands, and link it from here. **Last updated October 8, 2026.**
This is a living page. Every few weeks an AI company announces a model and a table showing it beating everyone else, and every company tops its own table. The question a reader actually has, **which AI model is best right now**, can't be answered from those tables. So this page answers it from the scoreboards the companies don’t run: the human-vote leaderboard LMArena, Artificial Analysis’s test battery, the ARC Prize puzzles, METR’s task-length measurements and Epoch AI’s trend work. I keep it current, write a separate entry when something big lands, and link it from here. **Last updated November 9, 2026.**
The logos at the top are ten of the labs competing for the top of these boards. The rule here is the one you would use on any machine that claims to be smart: before believing it, find out who is doing the scoring.
## Which Super Intelligence model is best right now?
## Which AI model is best right now?
*Standing page · updated Oct 8, 2026*
*Standing page · updated Nov 9, 2026*
**It depends whose test you trust. Human voters put Google’s new Gemini 4 Argon first. The biggest test battery puts Anthropic’s Claude Opus 5.5 five points clear.**
**It depends whose test you trust, and today the answer moved. The biggest test battery and the hardest puzzle test both put Anthropic’s Claude Fable 5.2, released today, first. Human voters put Google’s Gemini 4 Carbon first, a model most people can’t use yet.**
Three independent scoreboards, three different leaders, and a fourth that has run out of ruler. The disagreement is the useful part: each one measures something different, and the “best model” changes with the job.
Three independent scoreboards, two leaders, and a fourth that has stopped keeping score. The disagreement is still the useful part: each board measures something different, and the “best model” changes with the job. Both of Fable 5.2’s first places are first-day measurements, and I found no human-vote rating for it yet.
Human votes: Gemini 4 Argon · Test battery: Claude Opus 5.5 · New puzzles: GPT-6 Astra
Human votes: Gemini 4 Carbon · Test battery: Claude Fable 5.2 · New puzzles: Claude Fable 5.2
- **1525** Gemini 4 Argon’s rating, #1, but preliminary on 4,892 votes (Human votes, LMArena)
- **58** Claude Opus 5.5 on the Intelligence Index, the top score of 226 entries (Test battery, Artificial Analysis)
- **62.7%** GPT-6 Astra on the standard harness; 99.9% when allowed its own provider’s tools (New puzzles, ARC-AGI-3)
- **~4 mo** how far the best open-weight models trail the frontier (2026 average) (Open-model lag, Epoch AI)
- **8×** what the top score costs ($5.98) vs a model six points behind it ($0.72) (Price spread, Cost per task)
- **$172B** in Q2 2026 alone, more than their $155B for all of 2023 (Super Intelligence spending, Big-four capex)
- **1526** Gemini 4 Carbon’s rating, #1 but preliminary on 4,282 votes, and statistically tied with Gemini 4 Argon (Human votes, LMArena)
- **65** Claude Fable 5.2 at its max setting on the Intelligence Index, five points clear of Carbon’s 60 (Test battery, Artificial Analysis)
- **81.2%** Claude Fable 5.2 on the standard harness, past GPT-6 Astra’s 62.7% record (New puzzles, ARC-AGI-3)
- **19 pts** how far the best open-weight model (Kimi K3.1 Pro, 46) trails the top score (Open-model gap, Artificial Analysis)
- **3.4×** what the top score costs per task ($26.54) vs the model five points behind it ($7.76) (Price spread, Cost per task)
- **$210B** in Q3 2026 alone, up 84% on a year earlier (AI spending, Big-four capex)
### The scoreboard of scoreboards
- **LMArena · human votes:** Gemini 4 Argon. Blind head-to-head votes from the public: which answer reads better. #2 Claude Opus 5.5. Rewards answers people like. As of Oct 8.
- **Artificial Analysis · test battery:** Claude Opus 5.5. A battery of hard evaluations run by an independent firm, folded into one index. #2 Claude Sonnet 5.5. Weighted to reasoning, agents and code. As of Oct 8.
- **ARC-AGI-3 · new puzzles:** GPT-6 Astra. Puzzle games the model has never seen, scored against 500 members of the public. The harness changes the score from 62.7% to 99.9%. As of Sep 3.
- **METR · task length:** Ruler ran out. How long a task, measured in human working time, a model can finish half the time. The newest model measured was Claude Mythos Preview in May, and METR says anything above 16 hours is now beyond its tasks. Last update May 8.
- **LMArena · human votes:** Gemini 4 Carbon. Blind head-to-head votes from the public: which answer reads better. #2 Gemini 4 Argon, inside Carbon’s margin. I found no rating for Fable 5.2 yet. As of Nov 9.
- **Artificial Analysis · test battery:** Claude Fable 5.2. A battery of hard evaluations run by an independent firm, folded into one index. #2 Gemini 4 Carbon. The index was rebuilt in October (now v5.0.1), so these scores can’t be compared with the Oct 8 panel. As of Nov 9.
- **ARC-AGI-3 · new puzzles:** Claude Fable 5.2. Puzzle games the model has never seen. 81.2% on the standard harness, 97.7% with Anthropic’s own adapter; verified by ARC Prize on Nov 9.
- **METR · task length:** Ruler retired. On Sep 8 METR said its time-horizons page is no longer actively updated. The last model added was Claude Mythos Preview, on May 8.
### Human votes: LMArena’s top eight Oct 8, 2026
### Human votes: LMArena’s top eight Nov 9, 2026
| | |
| --- | --- |
| Gemini 4 Argon (high) | 1525 ±9 · prelim. |
| Claude Opus 5.5 (high) | 1507 ±8 |
| Claude Opus 4.6 (high) | 1504 ±3 |
| Gemini 4 Carbon (high) | 1526 ±10 · prelim. |
| Gemini 4 Argon (high) | 1522 ±5 |
| Muse Wonder (max) | 1511 ±6 |
| Claude Opus 5.5 (high) | 1506 ±6 |
| Claude Fable 5 (high) | 1504 ±4 |
| Claude Opus 4.6 (high) | 1503 ±3 |
| Claude Fable 5.1 (max) | 1503 ±6 |
| Claude Opus 4.7 (high) | 1501 ±4 |
| Claude Fable 5.1 (max) | 1501 ±6 |
| Claude Opus 4.6 | 1498 ±3 |
| Gemini 3.8 Flash (high) | 1497 ±5 · prelim. |
*Each bar is a rating’s confidence range on an axis from 1480 to 1540. Ranges that overlap are a statistical tie. Argon’s range clears everyone else’s, so its lead is real on today’s votes, but it rests on 4,892 votes against tens of thousands for most of the models below it. Note how little separates places two through eight.*
*Each rating has a confidence range; ranges that overlap are a statistical tie. Carbon (1516 to 1536) and Argon (1517 to 1527) overlap, so first place is a tie on today’s votes, and Carbon’s rating rests on 4,282 votes. Argon, preliminary on 4,892 votes when this page opened, now has 19,787 and has settled from 1525 to 1522. Muse Wonder is new to the top three. Fable 5.2, released today, has no rating on the board that I could find.*
### Test battery: Artificial Analysis Intelligence Index Oct 8, 2026
### Test battery: Artificial Analysis Intelligence Index Nov 9, 2026
| | |
| --- | --- |
| Claude Opus 5.5 | 58 |
| Claude Sonnet 5.5 | 56 |
| Claude Fable 5.1 | 53 |
| GPT-6 Astra | 53 |
| Gemini 4 Argon (tested at high) | 53 |
| GPT-6.1 Sol | 52 |
| MiMo-V2.6-Pro (Xiaomi) · best open | 46 |
| GLM-5.3 (Z AI) · open | 45 |
| Kimi K3 (Moonshot) · open | 44 |
| Claude Fable 5.2 | 65 |
| Gemini 4 Carbon (tested up to high) | 60 |
| Claude Opus 5.5 | 59 |
| GPT-6 Astra | 57 |
| Claude Sonnet 5.5 | 55 |
| GPT-6.1 Sol | 55 |
| Gemini 4 Argon (tested at high) | 50 |
| Kimi K3.1 Pro (Moonshot) · best open | 46 |
| GLM-5.3 (Z AI) · open | 32 |
*Each model’s best reasoning setting, on a scale starting at zero so the gaps aren’t exaggerated. Artificial Analysis lists every setting separately; I keep one bar per model. Anthropic holds the top two places outright and shares third with Google and OpenAI, and the best open-weight model sits 12 points back.*
*Each model’s best reasoning setting, on a scale starting at zero so the gaps aren’t exaggerated. Artificial Analysis lists every setting separately; I keep one bar per model. In October it rebuilt the index (v5.0, then v5.0.1 in November): new tests, each category weighted 25%. Most models scored lower on the new version (Argon went from 53 to 50), so only compare bars within this panel. Google’s launch post put Carbon at 59 on v5.0; the board now shows 60.*
### What a point of intelligence costs Oct 8, 2026
### What a point of intelligence costs Nov 9, 2026
| | |
| --- | --- |
| Claude Fable 5.1 (max) · scores 53 | $7.63 |
| Claude Opus 5.5 (max) · 58 | $5.98 |
| Claude Sonnet 5.5 (max) · 56 | $5.46 |
| GPT-6 Astra (max) · 53 | $3.26 |
| Claude Sonnet 5.5 (xhigh) · 52 | $2.01 |
| Gemini 4 Argon (high) · 53 · intro price | $1.99 |
| GPT-6.1 Sol (max) · 52 | $0.72 |
| Claude Fable 5.2 (max) · scores 65 | $26.54 |
| Claude Fable 5.2 (xhigh) · 62 | $14.83 |
| Claude Opus 5.5 (max) · 59 | $10.04 |
| Claude Sonnet 5.5 (max) · 55 | $10.02 |
| Gemini 4 Argon (high) · 50 | $9.95 |
| Claude Fable 5.2 (high) · 62 | $9.08 |
| Gemini 4 Carbon (high) · 60 · intro price | $7.76 |
| GPT-6 Astra (max) · 57 | $6.24 |
| GPT-6.1 Sol (max) · 55 | $1.27 |
| GPT-6.1 Sol (high) · 50 | $0.59 |
*Artificial Analysis’s cost to run one Intelligence Index task, in US dollars, with each model’s index score in the label. This measure folds in how many tokens a model burns thinking, which a per-token price list hides. The top score costs about eight times the cheapest score six points below it. Argon’s $1.99 is an introductory price that Google says will double.*
*Artificial Analysis’s cost to run one Intelligence Index task, in US dollars, with each model’s index score in the label. This measure folds in how many tokens a model burns thinking, which a per-token price list hides. The top score costs about 3.4 times Carbon’s 60. Fable 5.2 at its high setting scores 62 for about a third of its max cost. GPT-6.1 Sol reaches 55 for $1.27. Carbon’s price is introductory, and Google says it will double.*
| Flagship list price, per million input tokens | |
| --- | --- |
| GPT-4 · Mar 2023 | $30 |
| Claude Opus 4.1 · 2025 | $15 |
| Claude Fable 5.1 · 2026 | $10 |
| Claude Fable 5.2 · Nov 2026 | $8 |
| Gemini 4 Argon · 2026, standard | $4 |
| Gemini 4 Carbon · Oct 2026, intro | $2 |
*The newest flagship lists at under a seventh of what GPT-4 cost three and a half years ago, and it is far more capable. The full story of the two speeds at which this price falls is in [Six Million Tokens in Eleven Minutes](/t/82/six-million-tokens-in-eleven-minutes).*
*Anthropic’s newest flagship lists at about a quarter of what GPT-4 cost in 2023, and Google’s introduces at a fifteenth (rising to $4). The full story of the two speeds at which this price falls is in [Six Million Tokens in Eleven Minutes](/t/82/six-million-tokens-in-eleven-minutes).*
| Who pays for it: big-four capex | |
| --- | --- |
| All of 2023 | $155B |
| Q2 2026, one quarter | $172B |
| 2026 guidance | $735–750B |
| Q3 2026, one quarter | $210B |
| 2026 guidance | $720–740B |
*Capital spending by Amazon, Microsoft, Alphabet and Meta, per the Platformonomics tracker. 2026 is guidance as of late July, and the total has risen at each quarterly update this year. Q3 updates arrive with late-October earnings.*
*Capital spending by Amazon, Microsoft, Alphabet and Meta, per the Platformonomics tracker. Q3 was up 22% on Q2 and 84% on a year earlier: Amazon $78.7B, Alphabet $51.4B, Microsoft $45.2B, Meta $34.9B. Full-year guidance slipped for the first time this year, to $720–740B from $720–745B in July, after Microsoft cut its number from $175B to $170B.*
### How fast is it improving? METR, Jan–May 2026
### How fast is it improving? METR, Mar 2025–Sep 2026
- **Mar 2025** — METR: the length of tasks Super Intelligence agents can finish has doubled about every 7 months for six years (Measured in human working time, at 50% reliability, on software, ML and security tasks.)
- **Mar 2025** — METR: the length of tasks AI agents can finish has doubled about every 7 months for six years (Measured in human working time, at 50% reliability, on software, ML and security tasks.)
- **Jan 29, 2026** — New task suite: since 2023 the doubling time is about 131 days, roughly four months (Top model then: Claude Opus 4.5, about 5 hours 20 minutes (with a 95% range of about 3 to 12 hours).)
- **May 8, 2026** — METR adds Claude Mythos Preview and warns that anything above 16 hours is unreliable on its tasks (The ruler is running out. No newer model has been added since, including Opus 5.5, GPT-6 Astra or Argon.)
- **May 8, 2026** — METR adds Claude Mythos Preview and warns that anything above 16 hours is unreliable on its tasks (The ruler is running out.)
- **Sep 8, 2026** — METR says its time-horizons page is no longer actively updated (No model after Mythos Preview has been added, so Opus 5.5, GPT-6 Astra, Argon, Carbon and Fable 5.2 have no time horizon.)
### Claim check
- Claude Fable 5.2 beats Opus 5.5 on FrontierCode 2.0, 49.1% to 45.5% (Anthropic’s own table, Nov 9. No outside board runs FrontierCode yet.) · *Vendor table*
- Claude Fable 5.2 leads on the hardest tests (Artificial Analysis 65, five ahead of Carbon; ARC-AGI-3 81.2% on the standard harness, verified by ARC Prize. Both measured on launch day; no human-vote rating found yet.) · *Independent, day one*
- Gemini 4 Carbon is #1 on the Artificial Analysis index (Google’s Oct 28 post, at 59 on v5.0. True at launch; on Nov 9 the board shows Carbon at 60, second to Fable 5.2.) · *Overtaken*
- Gemini 4 Argon is the best model in the world (#1 on LMArena on Oct 8, on preliminary votes. Now #2 there, tied with Carbon, and 50 on the rebuilt Artificial Analysis index.) · *No longer*
- Gemini 4 Argon beats Claude and GPT-6 on 13 of 19 benchmarks (Google’s table, Sep 30. On the nine rows scored by outside leaderboards Argon still wins seven. The biggest margins (long context, video) are rows Google ran itself. [Full check](/t/239/gemini-4-argon-google-s-benchmarks-checked).) · *Vendor table*
- Gemini 4 Argon is the best model in the world (#1 on LMArena human votes (preliminary). 53 on Artificial Analysis, five behind Claude Opus 5.5. True on one scoreboard, not the other.) · *Depends on the test*
- Argon hallucinates less than any other top model (Artificial Analysis: 15%, the lowest of models scoring 45+ on its index (as reported by *The Batch*).) · *Independent*
- GPT-6 Astra has essentially solved ARC-AGI-3 (99.9% only with OpenAI’s own context tools; 62.7% on the neutral harness everyone else is scored on. ARC Prize calls both state of the art.) · *Depends on harness*
- Open models have nearly caught the frontier (One article in circulation put the gap at 6 points. Artificial Analysis’s own board shows 12 (46 vs 58), and Epoch AI found the lag slightly widened in 2026, to about four months.) · *Not on current data*
- Super Intelligence’s task length is doubling every four months (METR’s post-2023 estimate on its new suite: 131 days, with a range of 107 to 161. METR’s own caveat is that its suite is near its ceiling.) · *Independent*
- Argon hallucinates less than any other top model (Artificial Analysis: 15%, the lowest of models scoring 45+ on its old index (as reported by *The Batch*, Oct 5).) · *Independent, Oct 5*
- GPT-6 Astra has essentially solved ARC-AGI-3 (99.9% only with OpenAI’s own context tools; 62.7% on the neutral harness everyone else is scored on. Fable 5.2 now holds the neutral-harness record at 81.2%.) · *Depends on harness*
- Open models have nearly caught the frontier (Artificial Analysis now shows a 19-point gap on its rebuilt index (46 vs 65), and Epoch AI found the lag slightly widened in 2026, to about four months.) · *Not on current data*
- AI’s task length is doubling every four months (METR’s post-2023 estimate on its new suite: 131 days, with a range of 107 to 161. METR has since stopped updating the measurement.) · *Independent*
### Who leads at what by the cited test, not my opinion
- **Coding agent in a terminal:** Claude Opus 5.5. 66.4% on Terminal-Bench 4.0, as printed in Google’s own table; Argon 57.4%.
- **Office and professional work:** Gemini 4 Argon. Leads Vals Index, Finance Agent and Harvey’s legal benchmark, all scored by Vals AI.
- **Very long documents:** Gemini 4 Argon. 84.2% on GraphWalks at 256k–1M tokens vs 71.8% next best. Google’s own run.
- **Puzzles never seen before:** GPT-6 Astra. 62.7% on ARC-AGI-3’s standard harness, verified by ARC Prize.
- **Near-frontier on a budget:** GPT-6.1 Sol. Index score 52 for $0.72 a task, per Artificial Analysis.
- **Weights you can download:** MiMo-V2.6-Pro. Xiaomi’s model, 46 on the index, top of 81 open-weight entries.
- **Hardest test battery:** Claude Fable 5.2. 65 on the Artificial Analysis index at its max setting; Carbon 60. Nov 9.
- **Puzzles never seen before:** Claude Fable 5.2. 81.2% on ARC-AGI-3’s standard harness, verified by ARC Prize; GPT-6 Astra 62.7%. Nov 9.
- **Answers people prefer:** Gemini 4 Carbon and Argon, tied on LMArena human votes. Carbon is not generally available yet. Nov 9.
- **Coding agent in a terminal:** Claude Opus 5.5. 66.4% on Terminal-Bench 4.0, as printed in Google’s own table; Argon 57.4%. Oct 8; no outside run of Fable 5.2 or Carbon yet.
- **Office and professional work:** Gemini 4 Argon. Leads Vals Index, Finance Agent and Harvey’s legal benchmark, all scored by Vals AI. Oct 8.
- **Very long documents:** Gemini 4 Argon. 84.2% on GraphWalks at 256k–1M tokens vs 71.8% next best. Google’s own run, Oct 8.
- **Near-frontier on a budget:** GPT-6.1 Sol. Index score 55 for $1.27 a task, or 50 for $0.59, per Artificial Analysis. Nov 9.
- **Weights you can download:** Kimi K3.1 Pro. Moonshot’s model, 46 on the index, top of Artificial Analysis’s open-weights board. Nov 9.
### What I am watching for
- **Argon’s LMArena rating as votes pile up** Preliminary ratings often settle. If the lead holds past 20,000 votes, the “best on human votes” chip stays.
- **Argon’s public release and model card** Also an Artificial Analysis run at its highest setting, which could move the 53.
- **Q3 earnings, late October** Whether the big four raise 2026 capex again, and what they say about 2027.
- **A new ruler from METR** The time-horizon chart is the clearest progress measure there is, and it has been quiet since May.
- **Fable 5.2 on LMArena** Its first human-vote rating, and whether Carbon’s preliminary lead holds once it passes 20,000 votes.
- **Gemini 4 Carbon’s public release** It is limited to cyber defenders and trusted partners for now, at an introductory $2/$10 per million tokens that Google says will rise to $4/$20.
- **Outside coding-agent runs** Terminal-Bench and Vals results for Fable 5.2 and Carbon, to replace the three cards above that rest on Google’s own table.
- **ARC Prize 2026 results, December 4** The competition has closed; the winners’ scores show what open, cheap systems can do on the same puzzles.
- **A new ruler** METR’s time horizon was the clearest progress measure there was, and nobody has picked it up since September.
- **Q4 earnings, late January** Whether the big four’s 2027 capex plans keep rising after this year’s first trim.
### The story so far
[Oct 8, 2026 · newGemini 4 Argon: Google’s Benchmarks, Checked13 wins of 19, which rows Google ran itself, and why the leaderboards disagree.](/t/239/gemini-4-argon-google-s-benchmarks-checked)[Sep 22, 2026Claude Opus 5.5 Is Cheaper and FasterThe current Artificial Analysis leader, and a launch that opened with an alignment score.](/t/147/claude-opus-5-5-is-cheaper-and-faster-but-the-real-news-is-what-anthropic-led-wi)[Sep 21, 2026The Most-Downloaded AI Models in the World Are Alibaba’sThe open-weight side of the race, and Qwen’s three billion downloads.](/t/141/the-most-downloaded-ai-models-in-the-world-are-alibaba-s)[Sep 12, 2026The Week the AI Labs Asked to Be RegulatedGPT-6 Astra’s launch, and OpenAI’s turn on safety rules.](/t/99/the-week-the-ai-labs-asked-to-be-regulated)[Sep 10, 2026 · the LedgerSix Million Tokens in Eleven MinutesWhat intelligence costs, and who pays for the free version.](/t/82/six-million-tokens-in-eleven-minutes)
- **Nov 9, 2026 · new** [What is Claude Fable 5.2?](/t/1795/fable-5-2) Anthropic’s new top model: price, speed and its own benchmark claims.
- **Nov 9, 2026** [What do AI benchmarks actually measure?](/t/1788/what-do-ai-benchmarks-actually-measure) A question in the AI Benchmarks square.
- **Oct 8, 2026** [Gemini 4 Argon: Google’s Benchmarks, Checked](/t/239/gemini-4-argon-google-s-benchmarks-checked) 13 wins of 19, which rows Google ran itself, and why the leaderboards disagree.
- **Sep 22, 2026** [Claude Opus 5.5 Is Cheaper and Faster](/t/147/claude-opus-5-5-is-cheaper-and-faster-but-the-real-news-is-what-anthropic-led-wi) The Artificial Analysis leader of its day, and a launch that opened with an alignment score.
- **Sep 21, 2026** [The Most-Downloaded AI Models in the World Are Alibaba’s](/t/141/the-most-downloaded-ai-models-in-the-world-are-alibaba-s) The open-weight side of the race, and Qwen’s three billion downloads.
- **Sep 12, 2026** [The Week the AI Labs Asked to Be Regulated](/t/99/the-week-the-ai-labs-asked-to-be-regulated) GPT-6 Astra’s launch, and OpenAI’s turn on safety rules.
- **Sep 10, 2026 · the Ledger** [Six Million Tokens in Eleven Minutes](/t/82/six-million-tokens-in-eleven-minutes) What intelligence costs, and who pays for the free version.
*Every number on this page is from a named outside source with the date I read it. A lab’s own claim is labelled as the lab’s until someone else reproduces it. Scores move daily on some of these boards, so check the date in each panel heading.*
## How this page works
Three rules. **Every ranking comes from someone else’s test**, named and dated. **A company’s own benchmark table is a claim**, and it goes in the claim check until an independent scoreboard confirms or contradicts it. **No single number decides “best”**, because the scoreboards measure different things, and their disagreements are where the information is.
One disclosure belongs here. I use Anthropic’s Claude for nearly everything, and Claude models do very well on several of these boards. That is the reason the page runs on other people’s tests. If Claude slips, the bars will say so.
One disclosure belongs here. I use Anthropic’s Claude for nearly everything, and Claude models do very well on several of these boards. That is the reason the page runs on other people’s tests. If Claude slips, the bars will say so. The disclosure matters more than usual in this update, because a Claude model took first place on two boards on its launch day. Those are Artificial Analysis’s and ARC Prize’s measurements, not Anthropic’s and not mine, and first-day results can move when the boards re-run them.
> **The four scoreboards in one line each.** *LMArena* shows two anonymous answers side by side and asks a person which is better. It measures what people prefer, which rewards style as well as substance. *Artificial Analysis* runs a fixed battery of hard tests itself and averages them into one index, and it also prices each run. *ARC-AGI-3* drops a model into puzzle games it has never seen and compares it with ordinary people. It is the closest thing to a test of learning on the spot. *METR* times how long the tasks a model can finish would take a skilled human. It is the most intuitive measure of progress there is, and the hardest to keep running.
> **The four scoreboards in one line each.** *LMArena* shows two anonymous answers side by side and asks a person which is better. It measures what people prefer, which rewards style as well as substance. *Artificial Analysis* runs a fixed battery of hard tests itself and averages them into one index, and it also prices each run. *ARC-AGI-3* drops a model into puzzle games it has never seen and compares it with ordinary people. It is the closest thing to a test of learning on the spot. *METR* timed how long the tasks a model can finish would take a skilled human. It was the most intuitive measure of progress there is, and it proved the hardest to keep running.
## Why the scoreboards disagree
Gemini 4 Argon is the clean example this week. People voting on LMArena prefer its answers to everyone else’s. On Artificial Analysis’s battery, which leans heavily on multi-step reasoning, agent work and code, it ties for the second tier. Google’s own table says the same thing in its own way: Argon wins the knowledge-work and long-document rows and loses the terminal and coding-agent rows. There is no contradiction there. A model can write the answer you’d pick and still be second-best at grinding through a software job.
This week’s example is the two new flagships. People voting on LMArena prefer Gemini 4 Carbon’s answers, by a margin too small to separate it from Argon. On Artificial Analysis’s battery, which leans on multi-step reasoning, agent work, science and code, Claude Fable 5.2 is five points ahead of Carbon. There is no contradiction there. A model can write the answer you’d pick and still be second-best at grinding through a hard technical job, and the reverse. Fable 5.2 hasn’t faced the voters yet, so the human-vote side of the comparison is still open.
The ARC-AGI-3 result for GPT-6 Astra is the other lesson: **the harness is part of the score**. The same model scores 62.7% when every lab plays by the same neutral rules and 99.9% when it can use OpenAI’s own context-management features. Both are real, and they answer different questions. When a headline quotes one number, it is worth asking which harness produced it.
ARC-AGI-3 keeps teaching the other lesson: **the harness is part of the score**. GPT-6 Astra scores 62.7% under the neutral rules and 99.9% with OpenAI’s own context tools; Fable 5.2 scores 81.2% and 97.7% the same two ways. With their own makers’ adapters the two are close to the ceiling, and under the neutral harness Fable 5.2 leads by 18.5 points. There is a smaller lesson in Fable 5.2’s own results too: its max setting scored lower than xhigh (65.7% against 81.2%), because thinking longer ran 28 of the 55 games into the 12-hour limit. When a headline quotes one number, it is worth asking which harness and which setting produced it.
## Update log
- **November 9, 2026.** Claude Fable 5.2 released; Gemini 4 Carbon (Oct 28) added. Scoreboards read the same day: LMArena text, the Artificial Analysis Intelligence Index (now v5.0.1) and open-weights board, ARC Prize’s leaderboard and Fable 5.2 result page, METR’s time-horizons page (frozen since Sep 8), and Platformonomics’ Q3 capex scoreboard (Oct 30). Fable 5.2 and Carbon now lead the boards; Argon dropped to second on human votes; the open-model gap is 19 points on the rebuilt index; capex guidance slipped for the first time this year. Correction: the first version gave July’s 2026 capex guidance as $735–750B; Platformonomics’ Q3 post gives July’s range as $720–745B, and this version uses its figures. Wording changed from “Super Intelligence” to “AI” throughout.
- **October 8, 2026.** Page created. Scoreboards read the same day: LMArena text (updated Oct 8), the Artificial Analysis Intelligence Index and open-weights leaderboard, ARC Prize’s Sep 3 post on GPT-6 Astra, METR’s time-horizon page (last updated May 8) and Jan 29 TH1.1 post, Epoch AI’s May 29 open-vs-closed analysis, and the Platformonomics Q2 capex tracker. First entry: [Gemini 4 Argon, checked](/t/239/gemini-4-argon-google-s-benchmarks-checked).
## Where I could be wrong
- **These boards move daily.** LMArena in particular re-rates as votes arrive, and a preliminary rating can shift a lot. Every panel is a snapshot from the date in its heading.
- **“One bar per model” is my simplification.** Artificial Analysis scores each reasoning setting separately, and some models were tested at fewer settings than others. Argon was only tested at “high.”
- **Fable 5.2’s results are a day old.** Artificial Analysis and ARC Prize ran them at launch; a re-run, a regrade or a new index version can move them, as Carbon’s moved from 59 to 60.
- **The index changed under the page.** Artificial Analysis rebuilt its Intelligence Index in October, so the Nov 9 test-battery and cost panels can’t be lined up against the Oct 8 ones. Compare within a panel.
- **“One bar per model” is my simplification.** Artificial Analysis scores each reasoning setting separately, and some models were tested at fewer settings than others. Argon was only tested at “high,” Carbon up to “high.” Claude Fable 5.1 wasn’t in the rows of the current board I read.
- **Cost per task is one firm’s measurement** on one set of tasks. Real-world bills depend on the job, caching and how much thinking you let a model do.
- **The flagship price line mixes labs** and compares list prices for input tokens only. It shows the direction, not a like-for-like index.
- **The “who leads at what” cards inherit their sources’ limits.** Three of the six come from Google’s own table, and the long-document card is a test Google ran itself; each card says so.
- **Three of the “who leads at what” cards are a month old** and come from Google’s own table or Google’s own run; each card says so.
## Sources
1. LMArena (arena.ai). Text leaderboard, overall view, updated October 8, 2026. [arena.ai/leaderboard/text](https://arena.ai/leaderboard/text)
2. Artificial Analysis. LLM leaderboard (Intelligence Index, cost per task, speed), read October 8, 2026. [artificialanalysis.ai/leaderboards/models](https://artificialanalysis.ai/leaderboards/models)
3. Artificial Analysis. Open-weights models leaderboard, read October 8, 2026. [artificialanalysis.ai/models/open-source](https://artificialanalysis.ai/models/open-source)
4. ARC Prize. *OpenAI’s GPT-6 Astra on ARC-AGI-3*, September 3, 2026. [arcprize.org/blog/astra](https://arcprize.org/blog/astra)
5. METR. *Time Horizon 1.1*, January 29, 2026. [metr.org](https://metr.org/blog/2026-1-29-time-horizon-1-1/)
6. METR. Time-horizons page, last updated May 8, 2026. [metr.org/time-horizons](https://metr.org/time-horizons/)
7. METR. *Measuring AI Ability to Complete Long Tasks*, March 19, 2025. [metr.org](https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/)
8. Epoch AI. Data Insight: open-weight vs closed models on the Epoch Capabilities Index, May 29, 2026. [epoch.ai](https://epoch.ai/data-insights/open-closed-eci-gap)
9. Google DeepMind. Gemini models page (Performance table) and *Gemini 4 Argon Model evaluation* methodology PDF. [deepmind.google/models/gemini](https://deepmind.google/models/gemini/)
10. DeepLearning.AI, *The Batch*. *Data Points: Gemini 4 Argon’s benchmarks and availability*, October 5, 2026. [deeplearning.ai](https://www.deeplearning.ai/the-batch/gemini-4-argons-benchmarks-and-availability)
11. Platformonomics. *Follow the CAPEX: Q2 2026 Scoreboard*, July 2026. [platformonomics.com](https://platformonomics.com/2026/07/follow-the-capex-q2-2026-scoreboard/)
12. The Ledger. [Six Million Tokens in Eleven Minutes](/t/82/six-million-tokens-in-eleven-minutes) (September 10, 2026), for the flagship price history and its sources.
1. LMArena (arena.ai). Text leaderboard, overall view, read November 9, 2026. [arena.ai/leaderboard/text](https://arena.ai/leaderboard/text)
2. Artificial Analysis. LLM leaderboard (Intelligence Index v5.0.1, cost per task, time per task), read November 9, 2026. [artificialanalysis.ai/leaderboards/models](https://artificialanalysis.ai/leaderboards/models)
3. Artificial Analysis. Open-weights models leaderboard, read November 9, 2026. [artificialanalysis.ai/models/open-source](https://artificialanalysis.ai/models/open-source)
4. Artificial Analysis. Intelligence benchmarking methodology and version history. [artificialanalysis.ai/methodology](https://artificialanalysis.ai/methodology/intelligence-benchmarking)
5. ARC Prize. Leaderboard, read November 9, 2026. [arcprize.org/leaderboard](https://arcprize.org/leaderboard)
6. ARC Prize. Claude Fable 5.2 results, November 9, 2026. [arcprize.org](https://arcprize.org/results/anthropic-claude-fable-5-2)
7. ARC Prize. *OpenAI’s GPT-6 Astra on ARC-AGI-3*, September 3, 2026. [arcprize.org/blog/astra](https://arcprize.org/blog/astra)
8. METR. *Time Horizon 1.1*, January 29, 2026. [metr.org](https://metr.org/blog/2026-1-29-time-horizon-1-1/)
9. METR. Time-horizons page, last updated September 8, 2026. [metr.org/time-horizons](https://metr.org/time-horizons/)
10. METR. *Measuring AI Ability to Complete Long Tasks*, March 19, 2025. [metr.org](https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/)
11. Epoch AI. Data Insight: open-weight vs closed models on the Epoch Capabilities Index, May 29, 2026. [epoch.ai](https://epoch.ai/data-insights/open-closed-eci-gap)
12. Anthropic. *Introducing Claude Fable 5.2 and Claude Mythos 5.2*, November 9, 2026. [anthropic.com](https://www.anthropic.com/claude-fable-and-mythos-5-2)
13. Google. *Gemini 4 Carbon: our next frontier model*, October 28, 2026. [blog.google](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/)
14. Google DeepMind. Gemini models page (Performance table) and *Gemini 4 Argon Model evaluation* methodology PDF. [deepmind.google/models/gemini](https://deepmind.google/models/gemini/)
15. DeepLearning.AI, *The Batch*. *Data Points: Gemini 4 Argon’s benchmarks and availability*, October 5, 2026. [deeplearning.ai](https://www.deeplearning.ai/the-batch/gemini-4-argons-benchmarks-and-availability)
16. Platformonomics. *Follow the CAPEX: Q3 2026 Scoreboard*, October 30, 2026. [platformonomics.com](https://platformonomics.com/2026/10/follow-the-capex-q3-2026-scoreboard/)
17. Platformonomics. *Follow the CAPEX: Q2 2026 Scoreboard*, July 2026. [platformonomics.com](https://platformonomics.com/2026/07/follow-the-capex-q2-2026-scoreboard/)
18. The Ledger. [Six Million Tokens in Eleven Minutes](/t/82/six-million-tokens-in-eleven-minutes) (September 10, 2026), for the flagship price history and its sources.
*First published on the Notebook, October 8, 2026.**First published on the Notebook, October 8, 2026.*