Leaderboard · Intelligence Index v4.3.2 · captured
AI Model Leaderboard
The honest answer to “which model is best right now” is that the leader is Claude Opus 5.5, at 58. All 34 models below are ranked on that same independent score, with the measured cost of one task beside every one of them — and settling the ties the score leaves.
- Models ranked
- 34
- one measured configuration each, v4.3.2
- Top score
- 58
- Claude Opus 5.5
- Best value
- 633
- index points per $ — MiMo-V2.6-Flash at $0.06 a task
- Cost frontier
- 5
- models nothing else matches on score for the same or less money, across 12 makers
Scores and task costs are the source's. Rank, gap to the leader, index per dollar, the cost frontier and the provider bests are computed by BitsMinds from those two columns.
The leaders
Top of the board
Competition ranking. Models level on the index are separated by cost per task when one is more than 1.5× dearer; those closer than that still share a place, and the next one is skipped.
- 1stFrontier
Claude Opus 5.5
Anthropic · max with fallback
58Intelligence Index
v4.3.2Highest score on the board
- Cost / task
- $5.98
- Index / $
- 9.7
- 2nd
Claude Sonnet 5.5
Anthropic · max with fallback
56Intelligence Index
v4.3.22 points behind the top score
- Cost / task
- $7.62
- Index / $
- 7.3
- 3rdFrontier
Gemini 4 Argon
Google · high
53Intelligence Index
v4.3.25 points behind the top score
- Cost / task
- $1.99
- Index / $
- 27
The full ranking
All models, sorted your way
Bars are drawn from zero as a share of the highest score, so a longer bar is a genuinely higher number. Sort by capability, by what a task costs, or by how many index points a dollar buys.
| Rank | Model | |||
|---|---|---|---|---|
| 1 | Anthropic | 58 | $5.98 | 9.7 |
| 2 | Anthropic | 56 | $7.62 | 7.3 |
| 3 | Google | 53 | $1.99 | 27 |
| 4 | OpenAI | 53 | $3.26 | 16 |
| 5 | Anthropic | 53 | $7.63 | 6.9 |
| 6 | OpenAI | 52 | $0.72 | 72 |
| 7 | Muse Wonder(max) Meta | 52 | $4.52 | 12 |
| 8 | Anthropic | 51 | $5.86 | 8.7 |
| 9 | 48 | $1.05 | 46 | |
| 10 | Muse Spark 1.3(max) Meta | 48 | $1.60 | 30 |
| 11 | OpenAI | 47 | $1.99 | 24 |
| 12 | MiMo-V2.6-ProFrontier Xiaomi | 46 | $0.13 | 354 |
| 13 | 46 | $3.74 | 12 | |
| 14 | GLM-5.3(max) Z.AI | 45 | $2.01 | 22 |
| 15 | Qwen3.8 Max (0902) Alibaba | 45 | $5.41 | 8.3 |
| 16 | Step 5 Preview StepFun | 44 | $0.72 | 61 |
| 17 | Grok 4.6(high) SpaceXAI | 44 | $1.86 | 24 |
| 17 | Kimi K3(max) Moonshot AI (Kimi) | 44 | $2.00 | 22 |
| 19 | Anthropic | 43 | $0.21 | 205 |
| 20 | Z.AI | 42 | $0.25 | 168 |
| 21 | OpenAI | 42 | $1.40 | 30 |
| 22 | Gemini 3.8 Flash(high) Google | 41 | $1.24 | 33 |
| 23 | Qwen3.8-Flash-Next Alibaba | 40 | $0.37 | 108 |
| 24 | Qwen3.8 2.4T A95B Alibaba | 40 | $2.16 | 19 |
| 25 | DeepSeek V4.1 Flash(max) DeepSeek | 39 | $0.27 | 144 |
| 26 | MiMo-V2.6-FlashFrontier Xiaomi | 38 | $0.06 | 633 |
| 26 | OpenAI | 38 | $0.07 | 543 |
| 28 | Anthropic | 38 | $5.09 | 7.5 |
| 29 | GPT-5.6 Luna(max) OpenAI | 37 | $0.18 | 206 |
| 30 | DeepSeek V4 Pro 0813(max) DeepSeek | 36 | $0.67 | 54 |
| 31 | DeepSeek V4 Flash Vision(max) DeepSeek | 35 | $0.31 | 113 |
| 32 | Qwen3.8 27B(xhigh) Alibaba | 34 | $1.01 | 34 |
| 33 | K2 Horizon 375B A23B Institute of Foundation Models | 31 | — | — |
| 34 | Google | 30 | $0.67 | 45 |
- Cost / task
- $5.98
- Index / $
- 9.7
- Cost / task
- $7.62
- Index / $
- 7.3
- Cost / task
- $1.99
- Index / $
- 27
- Cost / task
- $3.26
- Index / $
- 16
- Cost / task
- $7.63
- Index / $
- 6.9
- Cost / task
- $0.72
- Index / $
- 72
- 7Muse Wonder(max)Meta52Index
- Cost / task
- $4.52
- Index / $
- 12
- Cost / task
- $5.86
- Index / $
- 8.7
- Cost / task
- $1.05
- Index / $
- 46
- 10Muse Spark 1.3(max)Meta48Index
- Cost / task
- $1.60
- Index / $
- 30
- Cost / task
- $1.99
- Index / $
- 24
- 12MiMo-V2.6-ProFrontierXiaomi46Index
- Cost / task
- $0.13
- Index / $
- 354
- Cost / task
- $3.74
- Index / $
- 12
- 14GLM-5.3(max)Z.AI45Index
- Cost / task
- $2.01
- Index / $
- 22
- 15Qwen3.8 Max (0902)Alibaba45Index
- Cost / task
- $5.41
- Index / $
- 8.3
- 16Step 5 PreviewStepFun44Index
- Cost / task
- $0.72
- Index / $
- 61
- Cost / task
- $1.86
- Index / $
- 24
- 17Kimi K3(max)Moonshot AI (Kimi)44Index
- Cost / task
- $2.00
- Index / $
- 22
- Cost / task
- $0.21
- Index / $
- 205
- Cost / task
- $0.25
- Index / $
- 168
- Cost / task
- $1.40
- Index / $
- 30
- Cost / task
- $1.24
- Index / $
- 33
- 23Qwen3.8-Flash-NextAlibaba40Index
- Cost / task
- $0.37
- Index / $
- 108
- 24Qwen3.8 2.4T A95BAlibaba40Index
- Cost / task
- $2.16
- Index / $
- 19
- 25DeepSeek V4.1 Flash(max)DeepSeek39Index
- Cost / task
- $0.27
- Index / $
- 144
- 26MiMo-V2.6-FlashFrontierXiaomi38Index
- Cost / task
- $0.06
- Index / $
- 633
- Cost / task
- $0.07
- Index / $
- 543
- Cost / task
- $5.09
- Index / $
- 7.5
- Cost / task
- $0.18
- Index / $
- 206
- 30DeepSeek V4 Pro 0813(max)DeepSeek36Index
- Cost / task
- $0.67
- Index / $
- 54
- 31DeepSeek V4 Flash Vision(max)DeepSeek35Index
- Cost / task
- $0.31
- Index / $
- 113
- 32Qwen3.8 27B(xhigh)Alibaba34Index
- Cost / task
- $1.01
- Index / $
- 34
- 33K2 Horizon 375B A23BInstitute of Foundation Models31Index
- Cost / task
- —
- Index / $
- —
- Cost / task
- $0.67
- Index / $
- 45
Capability against price
The value map
Every model on the board plotted against what one task costs to run. Cost is on a log scale because the field spans roughly a 127-fold range, from $0.06 to $7.63 a task.
- Anthropic
- OpenAI
- Meta
- Xiaomi
- SpaceXAI
- Z.AI
- Alibaba
- StepFun
- Moonshot AI (Kimi)
- DeepSeek
A hollow dot sits on the staircase without being on the frontier: it matches the best score available at its price, but something cheaper got there first.
By maker
Provider scoreboard
12 makers hold a measured row on this board. Ordered by their best model, because that is the only comparison a per-model index supports.
Anthropic
6 models58Best measuredClaude Opus 5.5Cheapest taskClaude Haiku 5.5 $0.211 on the cost frontier
Google
3 models53Best measuredGemini 4 ArgonCheapest taskGemini 3.1 Pro Preview $0.671 on the cost frontier
OpenAI
7 models53Best measuredGPT-6 AstraCheapest taskGPT-6 Luna $0.071 on the cost frontier
Meta
2 models52Best measuredMuse WonderCheapest taskMuse Spark 1.3 $1.60SpaceXAI
2 models46Best measuredGrok 4.7Cheapest taskGrok 4.6 $1.86Xiaomi
2 models46Best measuredMiMo-V2.6-ProCheapest taskMiMo-V2.6-Flash $0.062 on the cost frontier
Alibaba
4 models45Best measuredQwen3.8 Max (0902)Cheapest taskQwen3.8-Flash-Next $0.37Z.AI
2 models45Best measuredGLM-5.3Cheapest taskGLM-5.3-Flash $0.25Moonshot AI (Kimi)
1 model44Best measuredKimi K3Cheapest taskKimi K3 $2.00StepFun
1 model44Best measuredStep 5 PreviewCheapest taskStep 5 Preview $0.72DeepSeek
3 models39Best measuredDeepSeek V4.1 FlashCheapest taskDeepSeek V4.1 Flash $0.27Institute of Foundation Models
1 model31Best measuredK2 Horizon 375B A23B
Our own tests
BitsMinds Lab verdicts
18 head-to-heads we ran ourselves, newest first, that involve models on this board. Scores are in each test's own scheme — they are not index points and do not convert into them.
The index measures ten benchmarks. BitsMinds Lab measures one thing it cannot: what a model produces when it gets the identical brief and one attempt, with every build live in the article. How the tests work →
Gemini 4 Argon vs Opus 5.5, Astra and Grok: A New King?
- Claude Opus 5.5 · Claude Codeidx 5812 pts
- GPT-6 Astra · OpenAI Codexidx 539 pts
- Gemini 4 Argon · Gemini appidx 538 pts
- Grok 4.7 · Grok Buildidx 464 pts
Opus 5.5 wins with 12 points, Astra 6 takes 9, Argon takes 8, Grok 4.7 takes 4. Argon was by far the fastest on the fairground and the launch and staged a full launch to orbit, but its carousel never goes round and its launch scene is plain.
Read the test →Claude Haiku 5.5 vs GPT-6 Luna: Lost in Thought
- GPT-6 Luna · OpenAI Codexidx 384 pts
- Claude Haiku 5.5 · Claude Codeidx 433 pts
Luna wins with 4 points, Haiku 5.5 takes 3. Haiku 5.5 won the fairground, the one build it finished at maximum effort. On the interchange and the launch it filled Claude Code's 128,000-token response limit with reasoning again and again; one step down it finished the interchange, and it never produced a launch.
Read the test →Opus 5.5 vs GPT-6 Astra vs Grok 4.7 Build No Man's Sky
- Claude Opus 5.5 · Claude Codeidx 5818 pts
- GPT-6 Astra · OpenAI Codexidx 5311 pts
- Grok 4.7 · Grok Buildidx 467 pts
Claude Opus 5.5 wins with 18 points, GPT-6 Astra takes 11, Grok 4.7 takes 7. All three built the whole sequence from orbit to footfall; Opus 5.5 is the only one that turned it into a place worth staying in.
Read the test →Claude Opus 5.5 vs GPT-6.1 Sol: A Familiar Hill Climb
- Claude Opus 5.5 · Claude Codeidx 5819 pts
- GPT-6.1 Sol · OpenAI Codexidx 5213 pts
Claude Opus 5.5 wins with 19 points, GPT-6.1 Sol takes 13. GPT-6.1 Sol built a far longer and harder game than GPT-6 Sol did, but one whose coins and fuel ask little of the player, and whose layout and physics constants match GPT-6 Astra's earlier build.
Read the test →GPT-6.1 Sol vs GPT-6 Astra and GPT-6 Sol: A Dead Heat
- GPT-6.1 Sol · OpenAI Codexidx 529 pts
- GPT-6 Astra · OpenAI Codexidx 539 pts
- GPT-6 Sol · OpenAI Codexidx 484 pts
GPT-6.1 Sol and Astra draw with 9 points each, GPT-6 Sol takes 4. At GPT-6 Sol's per-token price, GPT-6.1 Sol matched OpenAI's flagship on these briefs, but it took as long as Astra to do it.
Read the test →Claude Sonnet 5.5 vs Opus 5.5, Astra, Sol: A Point Short
- Claude Opus 5.5 · Claude Codeidx 5813 pts
- Claude Sonnet 5.5 · Claude Codeidx 5612 pts
- GPT-6 Astra · OpenAI Codexidx 5311 pts
- GPT-6 Sol · OpenAI Codexidx 485 pts
Opus 5.5 wins with 13 points, Sonnet 5.5 takes 12, Astra takes 11, Sol takes 5. Sonnet 5.5 finished second in every round at half Opus 5.5's price and beat it on the interchange; it was quicker than Opus 5.5 on every brief and far slower than either OpenAI model.
Read the test →GPT-6 Sol vs GPT-5.6 Sol: A Small Step for Half the Price
- GPT-6 Sol · OpenAI Codexidx 486 pts
- GPT-5.6 Sol · OpenAI Codexidx 474 pts
GPT-6 Sol wins with 6 points, GPT-5.6 Sol takes 4. No regression on these briefs, and only a small, uneven step forward, from a model that was quicker on every brief and costs half as much per token.
Read the test →Opus 5.5 vs GPT-6 Sol vs Grok 4.7: Finally, a Real Game
- Claude Opus 5.5 · Claude Codeidx 5819 pts
- Grok 4.7 · Grok Buildidx 4610 pts
- GPT-6 Sol · OpenAI Codexidx 486 pts
Claude Opus 5.5 wins with 19 points, Grok 4.7 takes 10, GPT-6 Sol takes 6. Opus 5.5 is the first model on this brief whose fuel and collectibles reward driving well, and it drove its own level to check them; GPT-6 Sol built a shorter, easier game than the GPT-5.6 Sol it succeeds.
Read the test →GPT-6 Astra vs Sol vs Luna: Is the Discount Worth It?
- GPT-6 Astra · OpenAI Codexidx 5310 pts
- GPT-6 Sol · OpenAI Codexidx 485 pts
- GPT-6 Luna · OpenAI Codexidx 383 pts
Astra wins with 10 points, Sol takes 5, Luna takes 3, the same order as the price list. Sol and Luna finished each brief two to six times faster, and each of their six builds shipped a fault a single look would catch.
Read the test →Opus 5.5 vs Fable 5.1 vs Astra 6: We Have a New Leader
- Claude Opus 5.5 · Claude Codeidx 5810 pts
- Claude Fable 5.1 · Claude Codeidx 539 pts
- GPT-6 Astra · OpenAI Codexidx 538 pts
Opus 5.5 wins with 10 points, Fable 5.1 takes 9, Astra 6 takes 8. Opus 5.5 was the slowest entrant on every round and spent close to an hour thinking on each brief before it wrote a line; its one fault, an edge line across its exits, is the kind only looking at the picture would catch.
Read the test →Opus 5 vs GPT-5.6 Sol vs Grok 4.7: Can They Build a Game?
- Claude Opus 5 · Claude Codeidx 5112 pts
- GPT-5.6 Sol · OpenAI Codexidx 4710 pts
- Grok 4.7 · Grok Buildidx 466 pts
Claude Opus 5 wins with 12 points, GPT-5.6 Sol takes 10, Grok 4.7 takes 6. None of the three repeated the unreachable-coins fault that decided the first field; instead all three shipped collectibles that cost the player nothing, and Grok 4.7 again built working mechanisms behind a visible layer that is not joined to them.
Read the test →Grok 4.7 vs Fable 5.1 vs Astra 6: Can It Compete?
- Claude Fable 5.1 · Claude Codeidx 5310 pts
- GPT-6 Astra · OpenAI Codexidx 539 pts
- Grok 4.7 · Grok Buildidx 464 pts
Fable 5.1 wins with 10 points, Astra 6 takes 9, Grok 4.7 takes 4. Every mechanism the briefs asked Grok 4.7 for is present and computed correctly; in all three rounds the layer the viewer actually sees was left disconnected from it.
Read the test →GPT-6 Astra vs Claude Fable 5.1: Hill Climb
- Claude Fable 5.1 · Claude Codeidx 5315 pts
- GPT-6 Astra · OpenAI Codexidx 5314 pts
15-14 to Fable 5.1, the narrowest result the series has recorded. Both models verified at length and both still shipped a visible fault, because what they checked was the engine and never the level.
Read the test →Claude Sonnet 5 vs GPT-5.6 Terra: Middle-Class Fight
- Claude Sonnet 5 · Claude Codeidx 386 pts
- GPT-5.6 Terra · OpenAI Codexidx 422 pts
6-2 to Sonnet 5, on a low bar: nothing in the round scored above two out of four, and Terra's Pac-Man never parses.
Read the test →Claude Opus 5 vs GPT-5.6 Sol: Not Even Close
- Claude Opus 5 · Claude Codeidx 5112 pts
- GPT-5.6 Sol · OpenAI Codexidx 474 pts
12-4 to Opus 5, which won every round. Sol wrote an arcade-accurate Pac-Man rule set in ten minutes and shipped ghosts that drive through walls.
Read the test →GPT-6 Astra vs Claude Fable 5.1: There Can Be Only One
- Claude Fable 5.1 · Claude Codeidx 5310 pts
- GPT-6 Astra · OpenAI Codexidx 539 pts
10–9 to Fable 5.1 on the artefact. Weight the same rounds for cost and speed and it inverts to 10–9 Astra.
Read the test →Claude Fable 5.1 vs Opus 5 vs Sonnet 5: One Point Apart
- Claude Fable 5.1idx 5317 pts
- Claude Opus 5idx 5116 pts
- Claude Fable 5not on the board13 pts
- Claude Sonnet 5idx 386 pts
Fable 5.1 wins with 17 points, one ahead of Opus 5, after winning four of five rounds and losing the fifth completely.
Read the test →Claude Opus 5 vs Fable 5 vs Sonnet 5: One Prompt Each
- Claude Opus 5idx 518 pts
- Claude Fable 5not on the board6 pts
- Claude Sonnet 5idx 383 pts
Opus 5 wins with 8 points, Fable 5 takes 6 and Sonnet 5 takes 3; Opus 5 was also the fastest model on every round.
Read the test →
How to read this
Methodology
Where the scores come from
Every score is the Artificial Analysis Intelligence Index v4.3.2 (0–100, higher is better), which combines ten independent evaluations — AA-Briefcase, GDPval-AA v2, AutomationBench-AA, Terminal-Bench v4.0, SciCode, AA-LCR v1.1, AA-Omniscience, Humanity's Last Exam, GDP.pdf and CritPt (methodology). Data captured September 30, 2026. Each row is the highest-scoring configuration the source lists for that model, named exactly as the source labels it; rows the source marks as estimated — not independently run — are left out, as are models scoring under 30.
What the numbers do not say
Index versions are not comparable. Artificial Analysis re-bases the index when its evaluation set changes, and every model's number moves; earlier BitsMinds articles quote the version current when they were written. Cost per task is the source's measured average cost to run one index task at that configuration — the number that shows how many tokens a model burns to reach its score, not a list price. Want plans and features side by side instead? Use the interactive compare tool, or read the written head-to-head comparisons.
What BitsMinds computes
The source publishes two numbers per model: the index score and the cost per task. Everything else on this page is arithmetic we do on those two and label as ours — the competition rank and the gap to the top score, index points per dollar, the cost frontier, and each maker's best and cheapest row. No third measurement is introduced anywhere, and no score is rescaled, averaged or combined across makers.
The tie-break on price is ours too, and a deliberate departure from the source: Artificial Analysis ranks on capability alone, and this board does not. Where models are level on the index, the cheaper one takes the higher place — but only when the dearer one costs more than 1.5× as much per task. The index is published in whole numbers, so two models printed at the same score are equal only to within that rounding; splitting them over a narrower price gap would be false precision, so those keep a shared place instead.
Questions
Frequently asked
- Which AI model scores highest right now?
- Claude Opus 5.5, measured at 58 on the Artificial Analysis Intelligence Index v4.3.2, captured September 30, 2026. The index is capability only — it combines ten independent evaluations and takes no account of price — so the highest score is not automatically the model you should run.
- What does cost per task mean on this page?
- It is the source's measured average cost, in US dollars, to run one index task at the configuration listed for that model. It is not a list price per million tokens: two models on the same token price can differ several-fold here, because one burns far more tokens reaching its answer. Across this board one task ranges from $0.06 on MiMo-V2.6-Flash to $7.63 on Claude Fable 5.1.
- What is the cost frontier?
- A calculation BitsMinds runs on the source's two published columns. A model is on the frontier when no other model on this board is both at least as high-scoring and no more expensive per task. Five of the 34 models qualify — Claude Opus 5.5, Gemini 4 Argon, GPT-6.1 Sol, MiMo-V2.6-Pro and MiMo-V2.6-Flash — and every other model is matched or beaten on score by one of them at a lower cost per task.
- Why is a model I use missing from this leaderboard?
- Three filters. Rows the source marks as estimated, meaning it has not independently run all the required benchmarks, are left out entirely. Models scoring under 30 are left out to keep the board readable. And each model appears once, as the highest-scoring configuration the source lists for it — so a lower-effort setting of a model already here does not get a second row.
- Can I compare these scores with the ones in older BitsMinds articles?
- No. Artificial Analysis re-bases the index whenever it changes the evaluation set, and every model's number moves with it. An article quotes the version that was current when it was written; this board is v4.3.2, captured September 30, 2026. Compare within one version only.