Loading
Loading
Intelligence has a price, and the price is convex. This page divides each frontier model's Artificial Analysis (AA) Intelligence Index score by its blended API price, weighting input 80% and output 20%. The last 2 index points cost 2.0× the money: Claude Fable 5.1 against Claude Opus 5 on blended price. The chart below is the same claim as the table: up and to the left is better, and the amber line is the Pareto frontier.
AA Intelligence Index and first-party list prices, retrieved 13 Sep 2026. Scores are mode-specific and measured at the effort level named in the table. Latency is a third axis this chart does not have: both OpenAI and Anthropic now sell a Fast tier at twice the standard rate.
Unverified. Our frontier is the Pareto set of the 30 models we hold, drawn from the 200 Artificial Analysis evaluates, and Artificial Analysis's own published line has not been pulled into the registry yet. Until it is, read the 6 names on the frontier above as optimal within our sample rather than optimal outright: a model we do not track could dominate any of them. Artificial Analysis's scatter is the comparison set.
Cost bases differ: cost per Intelligence Index task (not blended $/Mtok — orderings agree, units do not). Membership is compared, not coordinates.
The chart above prices tokens. This one prices the job: what Artificial Analysis pays to run one model through one Intelligence Index task, which charges a model for how many tokens it spends thinking as well as for the rate on each. The two disagree, and the disagreement is the point. Claude Fable 5.1 and GPT-6 Astra carry an identical $10 / $50 list price. Per task they differ 2.4×, $6.12 against $2.57, because Fable 5.1 emits roughly 160M output tokens across the suite where the median peer emits 79M. On the chart above, Astra is dominated and falls off the frontier entirely; here it sits on it. Where the question is what work costs rather than what tokens cost, this is the axis to read.
AA Intelligence Index and first-party list prices, retrieved 13 Sep 2026. Scores are mode-specific and measured at the effort level named in the table. Latency is a third axis this chart does not have: both OpenAI and Anthropic now sell a Fast tier at twice the standard rate.
Cost is not the only thing a frontier can be drawn against, and it is not always the binding one. Below, the horizontal axis is throughput once generation starts; further right is faster. The frontier here has almost no members in common with the one above, which is the finding: paying more does not buy speed, and the models that top the index are among the slowest on the page.
AA Intelligence Index and first-party list prices, retrieved 13 Sep 2026. Scores are mode-specific and measured at the effort level named in the table. Latency is a third axis this chart does not have: both OpenAI and Anthropic now sell a Fast tier at twice the standard rate.
Time to first token, log scale, lower is better. For a reasoning model this is mostly think time rather than network latency, which is why it spreads across three orders of magnitude while price spreads across two. Read it as the axis that decides whether a model can sit behind an interactive product at all.
AA Intelligence Index and first-party list prices, retrieved 13 Sep 2026. Scores are mode-specific and measured at the effort level named in the table. Latency is a third axis this chart does not have: both OpenAI and Anthropic now sell a Fast tier at twice the standard rate.
Holding every model Artificial Analysis scores would not make this chart better: a frontier is decided by the handful of rows along its edge, and the rest only add ink. What costs us is a missing lab, because that is the one error a frontier drawn over a subsample can make. So the labs worth tracking are listed rather than assumed, and the unpriced ones are named.
Named flagships are leads for the next pull, not data. Nothing enters the table above without a score and a price from the model's own source page.
Click any column to sort.
| GLM-5.3-Flashopen weights rank #3 of 113 open-weight; MIT licence, 320B total / 18B active | 42 | 0.15 | 0.50 | 0.22 | 190.9 | 109 | 2.42 |
| Qwen3.8-Flash-Nextopen weights rank #5 of 113 open-weight; Qwen Community License 1.0. Price verified on Alibaba Cloud (International region) | 40 | 0.15 | 0.47 | 0.21 | 186.9 | 50 | 2.81 |
| GPT-5.6 Luna scored at max · rank #4 of 177 in its cost cohort; cheapest cost per task here, open or closed | 38 | 0.20 | 1.20 | 0.40 | 95.0 | 120 | 121.85 |
| DeepSeek V4.1 Flashopen weights scored at max · rank #6 of 113 open-weight; MIT, 552B total / 16B active. Peak-hour price, off-peak is $0.15 / $0.60 | 40 | 0.30 | 1.20 | 0.48 | 83.3 | 228 | 1.17 |
| MiniMax M3open weights rank #16 of 113 open-weight; MiniMax Community Licence | 30 | 0.30 | 1.20 | 0.48 | 62.5 | 109 | 1.27 |
| DeepSeek V4-Flashopen weights scored at max · rank #8 of 113 open-weight; retired 10 Sep 2026, name now routes to V4.1 Flash. Repriced UP from $0.14 / $0.28 | 35 | 0.44 | 1.32 | 0.62 | 56.8 | 213 | 0.98 |
| Gemini 3.8 Flash scored at high · rank #28 of 200; introductory price through 31 Dec 2026, $1.50 / $7.50 from 1 Jan 2027 per Google | 41 | 0.75 | 3.75 | 1.35 | 30.4 | 278 | 14.91 |
| Gemini 3.7 Flash scored at high · rank #37 of 200; same introductory price and 1 Jan 2027 doubling as 3.8 Flash | 39 | 0.75 | 3.75 | 1.35 | 28.9 | 288 | 10.71 |
| Muse Spark 1.3 scored at max · rank #13 of 200; list price unmoved across 1.1, 1.2 and 1.3 | 48 | 1.25 | 4.25 | 1.85 | 25.9 | 237 | 23.48 |
| GLM-5.3open weights scored at max · rank #1 of 113 open-weight; bespoke GLM-5.3 licence, commercial use with restrictions | 45 | 1.40 | 4.40 | 2.00 | 22.5 | 72 | 3.20 |
| Muse Spark 1.2 scored at xhigh · rank #33 of 200; standard tier, contributor tier is $0.10 / $0.20 in exchange for training rights | 40 | 1.25 | 4.25 | 1.85 | 21.6 | 166 | 15.21 |
| DeepSeek V4-Proopen weights scored at max · rank #7 of 113 open-weight; 0813 build, MIT, 1.6T total / 49B active. Peak-hour price, off-peak is half | 36 | 1.32 | 3.96 | 1.85 | 19.5 | 78 | 1.71 |
| Muse Spark 1.1 scored at xhigh · rank #50 of 200; deprecated, superseded by 1.2 and 1.3 at this same unmoved price | 34 | 1.25 | 4.25 | 1.85 | 18.4 | n/r | n/r |
| GLM-5.2open weights scored at max · rank #9 of 113 open-weight; MIT, 753B total / 40B active. Reverted to $1.40 / $4.40 from $1.35 / $4.29 | 34 | 1.40 | 4.40 | 2.00 | 17.0 | 76 | 7.39 |
| Grok 4.6 scored at high · rank #20 of 200; under-200K tier, a prompt of 200K or more reprices the whole request to $4.00 / $12.00 | 44 | 2.00 | 6.00 | 2.80 | 15.7 | 60 | 34.76 |
| Qwen3.8 Max rank #30 of 200; closed, unlike the Qwen models Alibaba opens below the flagship | 40 | 2.00 | 6.00 | 2.80 | 14.3 | 40 | 2.77 |
| Grok 4.5 scored at high · rank #38 of 200; under-200K tier, a prompt of 200K or more reprices the whole request to $4.00 / $12.00 | 39 | 2.00 | 6.00 | 2.80 | 13.9 | 62 | 6.73 |
| Gemini 3.5 Flash scored at high · rank #57 of 200; deprecated, Google points to 3.6 Flash and above | 33 | 1.50 | 9.00 | 3.00 | 11.0 | 220 | 15.42 |
| Claude Sonnet 5 scored at max · rank #41 of 200; $2 / $10 made standard, the scheduled 1 Sep rise to $3 / $15 was cancelled | 38 | 2.00 | 10.00 | 3.60 | 10.6 | 76 | 209.43 |
| GPT-5.6 Terra scored at max · rank #26 of 200 | 42 | 2.00 | 12.00 | 4.00 | 10.5 | 114 | 164.65 |
| Qwen3.7 Max rank #69 of 200; deprecated, Alibaba points to Qwen3.8 Max | 30 | 2.50 | 7.50 | 3.50 | 8.6 | 188 | 2.55 |
| Kimi K3open weights scored at max · rank #2 of 113 open-weight; Kimi K3 License | 44 | 3.00 | 15.00 | 5.40 | 8.1 | 37 | 4.52 |
| Gemini 3.1 Pro rank #68 of 200; still Preview. Prompts over 200K bill $4.00 / $18.00 | 30 | 2.00 | 12.00 | 4.00 | 7.5 | 115 | 25.29 |
| GPT-5.6 Sol scored at max · rank #14 of 200; cut from $5 / $30, promotional at least through 21 Nov 2026 per OpenAI | 47 | 4.00 | 20.00 | 7.20 | 6.5 | 58 | 133.98 |
| Claude Opus 5 scored at max · rank #7 of 200; 140M output tokens across the suite | 51 | 5.00 | 25.00 | 9.00 | 5.7 | 53 | 52.98 |
| Claude Opus 4.8 scored at max · rank #27 of 200; deprecated, Anthropic points to Opus 5 | 42 | 5.00 | 25.00 | 9.00 | 4.7 | 56 | 23.23 |
| GPT-5.5 scored at xhigh · rank #40 of 200; deprecated, OpenAI points to GPT-5.6 Sol. Context is 922K, not 1M | 39 | 5.00 | 30.00 | 10.00 | 3.9 | 96 | 52.06 |
| Claude Fable 5.1 scored at max · rank #1 of 200, tied on score with GPT-6 Astra; 190M output tokens across the suite vs 60M for Astra | 53 | 10.00 | 50.00 | 18.00 | 2.9 | 67 | 212.01 |
| GPT-6 Astra scored at max · rank #3 of 200, tied on score with Fable 5.1; standard tier, the long-context band is $20.00 / $75.00 | 53 | 10.00 | 50.00 | 18.00 | 2.9 | 60 | 294.93 |
| Claude Fable 5 scored at max · rank #8 of 200; superseded by Fable 5.1 (1 Sep 2026) at the same list price | 50 | 10.00 | 50.00 | 18.00 | 2.8 | 65 | 83.53 |
Blended cost per 1M tokens = 0.8 × input price + 0.2 × output price, the mix of a retrieval-heavy workload that reads far more than it writes. Scores are the AA Intelligence Index at the reasoning effort shown against each model, where AA names one; prices are first-party API list prices. The ratio rewards cheapness, so compare models within a capability tier rather than across tiers · a model that fails the task is expensive at any price.
GLM-5.3, Kimi K3, GLM-5.3-Flash, Qwen3.8-Flash-Next, DeepSeek V4.1 Flash, DeepSeek V4-Pro, DeepSeek V4-Flash, GLM-5.2 and MiniMax M3 are open-weight models; the rest are closed.
Source: artificialanalysis.ai per-model pages, retrieved 13 Sep 2026. Prices and scores move; the retrieval date is the claim.