LLM vIPD

Intelligence per Dollar — nine models against Fable 5

Model economics · August 2026

Contents show

How much intelligence
does a dollar actually buy?

Nine models scored against Claude Fable 5. Two metrics: one prices tokens,
one prices finished work. They disagree, and the disagreement is the interesting part.

IPD — intelligence per dollar

Index ÷ blended $/M tokens

Blended at 3:1 input-to-output. Sticker-price economy. Rewards cheap tokens and ignores how many of them a model burns.

vIPD — verbosity-adjusted

Index ÷ (tokens per task × output $/M)

Prices one finished Intelligence Index task using measured output consumption. A model that rambles pays for it here.

The index itself

Artificial Analysis v4.1, max effort

Nine evaluations averaged to one 0–100 score: agentic work, terminal coding, long-context retrieval, graduate science, hallucination rate.

Model	Index	Out tok /task	Output $/M	Cost /task	IPD	vIPD	vs Fable
DeepSeek V4 Flash 0731 Talks twice as much as anything here and still wins by an order of magnitude. Verbosity is free when output runs 28¢ a million.	50	~78K	$0.28	$0.022	286	2,296	61×
Inkling The one model that improves when you stop counting tokens and start counting tasks. Frugality is the product, not the score.	41	25K	$4.68	$0.117	16	350	9.4×
GLM-5.2 Holds roughly the same value under either metric. MIT weights, no unpleasant surprises, and the cheapest way to run a long coding agent.	51	43K	$4.40	$0.189	24	270	7.2×
Qwen3.7 Max Four times Fable’s value and three times its speed at 201 tokens/sec. Verbose, but at $7.50 output that costs less than it does upstream.	46	~37K	$7.50	$0.278	12	166	4.4×
Kimi K3 Best open-weight capability on the board, but it spends the most words of any frontier model getting there — which eats a third of the discount.	57	~48K	$15.00	$0.722	10	79	2.1×
GPT-5.6 Sol The concise one at the top. Fewest tokens of any frontier model here, which is the only thing making $30 output survivable.	59	~26K	$30.00	$0.778	5	76	2.0×
Claude Opus 5 Highest index measured, at half Fable’s token price. It thinks harder to get there, so the saving is real but smaller than the sticker suggests.	61	~37K	$25.00	$0.926	6	66	1.8×
Claude Fable 5 Now second on capability and last on value. Its cost problem is entirely sticker price — it is one of the more concise models measured here.	60	~32K	$50.00	$1.611	3	37	1.0×
Token consumption not yet published — IPD only
MiniMax M3 Cheapest credible route to a 1M window with video in. Sparse attention suggests it lands well on vIPD, but nobody has measured it.	44	—	$1.20	—	84	—	28×

The frontier collapses

Opus 5, GPT-5.6 Sol and Kimi K3 land within 0.3× of each other despite indices of 61, 59 and 57 and output prices spanning $15 to $30. Verbosity cancels the price differences almost exactly. At the top of the market you are choosing a vendor, not a value tier.

Where the tokens/task come from

Artificial Analysis publishes Index totals, not per-task figures, for six of these. Those are divided by ~2,700 tasks, inferred from AA’s own medians. Inkling and GLM-5.2 are published per-task directly. Ratios are solid; absolute cost-per-task carries that inference.

Output-only, so the top is exaggerated

Input and cache are excluded. AA puts Opus 5 at 26% below Fable on all-in cost per task; output-only makes it 42%. Expect the real frontier spread to be tighter than shown — which reinforces the point above rather than undercutting it.

Two version caveats

Qwen3.7 Max launched to coverage citing 56.6 and a #5 placing; AA’s v4.1 rescoring puts it at 46. Fable 5’s 60 is the max-effort setting with Opus 4.8 fallback, so its lead narrows on cybersecurity and biology-adjacent evaluations.