Best AI Model: 18 Models Ranked by Intelligence Index and Cost

1. The verdict

As of November 12, 2026, the best AI model overall is Claude Fable 5.2 — Intelligence Index 65, the highest of the 18 models we track, at $26.54 per evaluation task.

If you are paying for this out of your own budget, the best value is GPT-6.1 Terra: it scores 51, which is 78% of Fable's score for 3.6% of the cost ($0.96 per task). Both numbers come from the same table below, so you can check the arithmetic: 51 ÷ 65 = 0.785, and 0.96 ÷ 26.54 = 0.036.

If your input is audio or video, there is one answer in this table: Gemini 4 Carbon (60, $7.76). Only three models in this list are specified by AA as accepting speech and video, and Carbon is the only one that also has a cost figure from AA and a score at or above 60. A second source lists video input for three other rows — section 5 explains why that does not change the recommendation, and where it does narrow it.

See the full 18-model table ↓

2. All 18 models

Sorted by Intelligence Index. Every row is taken at that model's highest reasoning effort — the same model can lose up to 24 points at a lower setting, so mixing tiers would make the column meaningless.

Column definitions.

• N/A (AA: Unknown) — Artificial Analysis reports nothing in that cell. For Muse Wonder Full the cost card reads In – / Out – / N/A and Cost per Intelligence Index task Unknown out of 4 units for Cost; the same model's Performance card reads Unknown out of 4 units for Performance, so it has two empty fields, cost and time, not one. The gap is in the aggregator's data, not a finding about the model.

• Intelligence Index — Artificial Analysis Intelligence Index v5.0.1. A composite score, not a percentage; the useful comparison is between rows, not the absolute number.
• Tier — the reasoning-effort setting at which that score was recorded. We always use the model's highest available setting.
• Cost / task — what one full evaluation run cost in USD. This is not the same as token price: two models with identical per-token rates can land several dollars apart per task.
• $/M in / out — list price per million tokens, published on the same page and at the same tier as the index.
• Input modalities — from each model's own technical specification page, not from a vendor claim.
• Context — maximum context window. 16 of the 18 rows offer 1M, so context is a tiebreaker, not a differentiator.

3. The frontier: who is actually worth paying

Nine models score 50 or above. Four earn their price — Fable for the top score, Carbon for multimodal value, Sol and Terra on cost per index point. The other five lose on the axes that decide a bill: Argon is beaten on all four axes by Carbon; Sonnet pays 7.9× Sol's price for the same score; Opus and Astra each pay 4.9× to 7.9× Sol's price for two to four extra index points; and Full has no cost figure from AA at all. Each of those verdicts is about API access; none of them is a claim about a model you can download and run yourself.

Claude Fable 5.2 (65, $26.54) is the strongest model here, and also the slowest: 47.5 minutes per task, 275M output tokens, $8/$40 per million tokens. AA ranks it #1 of 74 on the Intelligence Index — first among the models scored in its comparison class, which holds 237. On cost it is #73 of 73, last among the models in that class that carry a cost figure. First on intelligence, last on cost. Per index point it is the most expensive model in the table at $0.41 — 21.7× the cost per point of Terra. If you need the top score, this is the price. If you need throughput, this is not the model.

Gemini 4 Carbon (60, $7.76) is the value play of the frontier. It gives up 5 points to Fable and costs 29% as much, and it is the only model above 59 that accepts audio and video. At $4/$20 with a 95% cache discount, it is also the cheapest way to get a score of 60 or more.

Claude Opus 5.5 (59, $10.04) is beaten on both axes by Carbon — one point lower, 29% more expensive per task — at the same $4/$20 token price. Within Anthropic's own lineup it is also squeezed: Fable is 6 points higher for 2.6× the price, Sonnet is 4 points lower for essentially the same $10.02. Opus is a model you pick for reasons outside this table.

GPT-6 Astra (57, $6.24) carries the highest output price here ($50/M) and emits 54.3M output tokens, the lowest in the table. It is a specialist, not a default: Sol reaches 55 for $1.27, so two points cost 4.9×.

Muse Wonder Full (50) has the capability of a frontier model — 1M context, text/image/speech/video input — and AA reports neither a cost nor a time per task for it: both fields read Unknown. See section 5 for how to treat it.

Gemini 4 Argon (50, $9.95) is the clearest skip in the frontier: same $4/$20 list price as Carbon, 10 points lower, and 28% more per task. The identical token price buys a visibly worse bill, which is the whole argument for comparing cost per task rather than rate cards.

4. The two models most people should actually run

GPT-6.1 Terra (51, $0.96). Sub-dollar per task, 14 minutes per task — the fastest in the frontier — 1M context, $1/$5 per million tokens. It is 14 points behind Fable for 3.6% of the cost. For classification, extraction, summarisation and most agent loops, this is the default until something in your own evaluations says otherwise.

GPT-6.1 Sol (55, $1.27). Four points above Terra for 32% more. If your workload is quality-sensitive but not Fable-sensitive, this is the cheapest way to sit at 55.

And one model to stop paying for: Claude Sonnet 5.5 (55, $10.02). Sol scores the same 55 at $1.27 — 7.9× cheaper — and Sonnet's own vendor sells Fable 6 points higher for 2.6× the price. Sonnet is not beaten on capability; it is beaten on price at identical capability — for API access, the more common way to overpay.

5. Audio and video: three models, one real option

Only three of the 18 are specified by AA as accepting speech and video input: Gemini 4 Carbon, Muse Wonder Full, and Muse Wonder. Everything else is text and image; GLM-5.3 is text only, not even images.

• Gemini 4 Carbon — 60, $7.76 per task, $4/$20. Capability and price both reported by AA.
• Muse Wonder Full — 50, 1M context, same four input modalities, cost and time per task both Unknown (AA prints In – / Out – / N/A for cost; Meta says the Wonder series is reachable through its Meta Model API). We report the gap in AA's data. We do not resolve it in either direction.
• Muse Wonder — 44, $6.95 per task, $4/$20, same modalities. The variant AA prices, 6 points below Full.

So the decision is simple. If you need video or audio input and a cost figure from AA, Carbon is the only row. If you need the capability and can work without an AA cost figure, Wonder costs $6.95 at 44; Full gives 6 more points and still no cost figure.

At the same $4/$20 list price you can have 60 points with video (Carbon), 59 without (Opus), or 44 with video (Wonder) — so the capability is not what the price is buying. The model is.

We do not fill a price for Muse Wonder Full. AA's Cost field reads In – / Out – / N/A and Unknown out of 4 units for Cost; its Performance field reads Unknown out of 4 units for Performance. Unknown is not free, and it is not unavailable — it means the aggregator has no number for that cell, and we will not turn that into a word about the model.

If your provider lists video input

AA records these three models as accepting text and image; some inference providers list video for their own route to the same model — Kimi K3.1 Pro, Gemini 4 Argon and Qwen3.8 Max. The two sources record different things: AA records what the model accepts, a provider records what its endpoint accepts. Every modality cell in this table comes from AA, so all three read text-and-image here. Check your provider's model card before choosing on modality.

Argon is where this matters most: same $4/$20 list price as Carbon, 10 points lower on the Index. If video is available on your route, Carbon stops being the only priced option that takes video. It does not stop being the cheaper and stronger of the two: 60 vs 50 on the Intelligence Index, $7.76 vs $9.95 per task, 25.7 vs 30.0 minutes — a video-capable Argon closes the modality gap, not these three.

6. Cheap is not the same as economical

Divide cost per task by the index score and you get cost per index point — the only column that answers "am I overpaying for capability":

Three different answers, from one table: cheapest is Luna ($0.15), strongest is Fable (65), best value is Terra ($0.0188 per point). A page that only lists the first column would tell you to run Luna; a page that only lists the second would tell you to run Fable. Both are wrong for almost everyone.

The trap in the cheap end. GLM-5.3 looks like a budget model at 32 points and $1.40/M input — but it costs $5.01 per task and lands at $0.157 per index point, worse than Carbon (60 points) at $0.129. For API access it is beaten outright by Claude Haiku 5.5: 36 vs 32 points, $0.41 vs $5.01 per task, text and image input against text only, 14.7 minutes per task against 27.6, both at 1M context.

On cost, AA ranks GLM-5.3 #26 of 26 — last among the 26 models in its comparison class that carry a cost figure. Read that ratio carefully: the class itself holds 129 open-weight models above 150B parameters, and only 26 of them are ranked on cost, so the denominator here is models with a cost figure, not the size of the class. AA's summary calls it "particularly expensive when comparing to other open weight models of similar size", which is the same finding in words. It is not a weak model inside that class — AA ranks it #3 of 27 on intelligence — it is an expensive one. It is also verbose, at 135M output tokens against a class median of 104M. Verbosity you did not ask for is the most expensive kind.

That verdict stops at the API. GLM-5.3 is open weights — 753B total parameters, 40B active, weights on Hugging Face — so a reader whose plan is to self-host sits outside this comparison, and for them none of the numbers above are the deciding ones.

And the trap at the other end. Qwen3.8 Max and Grok 4.7 both sit near $0.22–0.29 per point — five to fifteen times Terra — which is where a model has to earn its place with something not in this table.

Two numbers that decide your bill and never appear in a rate card

Time per task. One evaluation run takes 47.5 minutes on Fable 5.2 and 14.0 on Terra — a 3.4× difference in wall-clock time for 14 index points. The fastest rows in the table are GPT-6 Luna and Mistral Large 4 at 11.6 minutes, both below 30 points. If your workload is latency-bound rather than quality-bound, the frontier is not where you should be shopping at all.

Output length. One evaluation run on Claude Sonnet 5.5 emits 335M output tokens — the highest of the 18 rows here, and 6.2× Astra's 54.3M. Within its own comparison class AA ranks it last on verbosity, at #73 of 73. That is why Sonnet bills at $10.02 on a $2/$10 rate card while Astra bills at $6.24 on a $10/$50 card: the rate card says Astra is five times more expensive per token, and the bill says the opposite. Verbosity is a cost multiplier you cannot see until you measure cost per task.

Neither number is in the table because both belong to the evaluation run rather than to your workload. They are still the two most common reasons a model that looked cheap in a spreadsheet turns out expensive in production.

7. Three ways to misread this table

1. Don't compare scores across indices. LLM Stats scores Fable 5.2 at 61.1; Artificial Analysis scores it 65. Both are "correct" on their own scale. Mixing a score from one index with a price from another is how tables end up internally inconsistent. Every number here comes from one index at one version.

2. Don't mix reasoning tiers. Opus 5.5 scores 59 at max effort and 35 at low — a 24-point swing inside one model. Carbon scores 60 at high and 54 at low. A row that does not name its tier is not a data point, it is an average of several.

3. Don't price by tokens alone. Gemini 4 Carbon and Muse Wonder have identical list prices ($4/$20, 95% cache) and different costs per task — $7.76 vs $6.95. Carbon also emits 205M output tokens against Wonder's 113M, so the gap runs the direction the rate cards cannot show. Token price tells you the rate; only cost per task tells you the bill.

On context windows: 16 of the 18 models offer 1M. It is a qualifying condition, not a differentiator, and it is the single most over-quoted number in AI model comparisons this year.

8. Method, sources and update policy

Every value in the table is retrieved programmatically, not transcribed:

• Intelligence Index, reasoning tier, cost per task, token price — Artificial Analysis, index version v5.0.1, retrieved November 12, 2026. All values taken at each model's highest reasoning effort.
• Input modalities and context window — each model's own Artificial Analysis page, technical specifications section.
• Existence check — every model name is verified against two independent sources before it enters the table. Models announced but not listed are excluded.

What we exclude, and why. Two kinds of entries never reach the table: models that appear in one aggregator but cannot be confirmed in a second source, and models marked unreleased or preview-only. The first rule exists because one widely-cited leaderboard currently lists a model family that neither Artificial Analysis nor the OpenRouter catalogue carries — copying it would have put a non-existent product on this page. The second exists because an unreleased model is a press release, not a purchasing option. A missing cost is not grounds for exclusion — Full stays, with the gap labelled rather than filled.

This page is rebuilt weekly. If a model's index moves by 2 points or more, or a new model enters the top 50, the table and the Updated date change together. If nothing changes, the date still moves — a timestamp that never moves is how comparison sites die.

Data as of 2026-11-12 · Source: Artificial Analysis Intelligence Index v5.0.1
(highest reasoning effort per model). Prices and cost per task: the same AA model pages.
Muse Wonder Full: AA's Cost field reads "In – / Out – / N/A" ("Unknown out of 4 units for
Cost"), and its Performance field reads "Unknown out of 4 units for Performance". Two empty
cells in AA — a gap in the source, not an omission of ours, and not a statement about the
model. Meta states the Wonder series is reachable via Meta Model API.
Updated weekly.

FAQ

Is Fable 5.2 worth 27× Terra? Only if you need the top 4 points of the index. Per index point it is 21.7× more expensive.

Which model is best for coding? Not answerable from this table — the index is a composite. See best model for coding.

Why isn't my model listed? The 18 rows are the highest-scoring models with verifiable data. A model that cannot be confirmed in two sources does not enter the table; a model whose cost AA does not report enters it with that cell marked N/A, not filled in.

Is the newest model the best one? Not necessarily. GPT-6 Astra scores 57 and sits below Gemini 4 Carbon at 60, and below Claude Opus 5.5 at 59 — both older. Generation number is a release note, not a measurement.

Why no "top 10" numbering? Because the denominator moves — and not only over time. On GLM-5.3's page AA prints four different populations for one model: 129 models in the class, 26 ranked on cost, 25 on time and 27 on verbosity. A rank without its population is a number with no meaning. The order of the table is the ranking.

Related: Claude Opus 5.5 vs GPT-6.1 · Gemini 4 Carbon vs Argon · The cheapest way to run 1M tokens · Best model for coding · AI model comparison (full table) · How we test