What is token efficiency in LLMs?

Token efficiency is how many tokens a model consumes to complete a given task, as opposed to what each token costs. A model that solves a problem in 3,000 output tokens at $50 per million costs $0.15, while a model that takes 45,000 output tokens at $3.75 per million costs $0.17 — the expensive model is cheaper. Artificial Analysis measures this directly: running its full Intelligence Index took GPT-6 Astra (high) about 16 million output tokens against roughly 123 million for Gemini 3.8 Flash (high), a 7.7x gap that erases Flash's roughly 13x token-price advantage.

Why do cheap AI models sometimes cost more?

Four compounding reasons. Verbose reasoning burns output tokens, which are typically 4 to 5 times more expensive than input tokens. Longer conversations mean more context re-read on every subsequent turn. Weaker models make more tool calls, each adding tokens and latency. And lower success rates force retries, so you pay two or three times for one completed task. A model with a 60% success rate at a sixth of the price is not cheaper than a 90% model once retry arithmetic is applied.

How do I calculate the real cost of an LLM?

Measure cost per completed task, not cost per million tokens. Run 100 representative tasks from your actual workload, record total input tokens, output tokens, cached reads, tool calls and the success rate, then divide total spend by successfully completed tasks. Include retries in the numerator and exclude failures from the denominator. Compare that single number across candidate models. Pricing pages cannot tell you this because token consumption is a property of the model and your workload together.

Quick Answer

Token Efficiency vs Token Price: Real LLM Cost 2026

Q: Does prompt caching change the token efficiency calculation?

Substantially, for agent workloads. Long-running agents re-read accumulated context on every turn, so cache reads often dominate the token bill. Claude Fable 5.1 cut cache reads to $0.25 per million in September 2026, which Anthropic estimates makes typical workloads about 25% cheaper and highly agentic ones up to roughly 45% cheaper. That is a different lever from token efficiency: one model wins by generating fewer tokens, another by making re-read tokens nearly free. Output-heavy work favours efficiency; context-heavy loops favour cheap caching.

Published: September 4, 2026

The Short Answer

The cheapest model per token is frequently the most expensive model per task. In 2026 this stopped being a subtlety and became the central fact of LLM cost management.

The measurement that proves it: running the full Artificial Analysis Intelligence Index took GPT-6 Astra (high) about 16 million output tokens. Gemini 3.8 Flash (high) needed about 123 million — a 7.7x gap that wipes out Flash’s roughly 13x advantage on token price.

Cost per completed task is the only number worth optimising. Last verified: September 4, 2026.

The Table That Breaks the Pricing Page

From the Artificial Analysis coding-agent harness, September 2026:

Agent configuration	Coding Agent Index	Cost per task
Codex · GPT-5.6 Luna (max)	57.1	$0.29
Opencode · Gemini 3.7 Flash (high)	59.6	$1.27
Codex · GPT-6 Astra (low)	62.6	$1.41
Muse Code · Muse Spark 1.3 (xhigh)	64.2	$1.72
Opencode · Gemini 3.8 Flash (high)	61.1	$2.04
Codex · GPT-6 Astra (medium)	65.1	$2.19
Codex · GPT-6 Astra (xhigh)	67.0	$3.27
Codex · GPT-5.6 Sol (max)	65.0	$5.00
Claude Code · Opus 5 (xhigh)	68.1	$8.17
Claude Code · Fable 5.1 (max)	70.4	$9.18

Read row three against row five. GPT-6 Astra costs $10 per million input tokens. Gemini 3.8 Flash costs $0.75 — thirteen times less. Yet Astra (low) delivers a higher index score (62.6 vs 61.1) at a lower cost per task ($1.41 vs $2.04).

A pricing page would have told you the opposite.

Why Verbose Models Cost More Than Their Price

Four effects compound, and only the first is obvious.

1. Output tokens are the expensive ones. Most 2026 providers charge output at 4–5x input. Claude Fable 5.1 and GPT-6 Astra both run $10 input / $50 output. A model that reasons at length is spending your money in the most expensive currency available.

2. Long conversations re-read themselves. Every additional turn drags the accumulated context back through the input side. A verbose model does not just produce more output once — it inflates the input cost of every subsequent turn in the same session.

3. Weaker models make more tool calls. Each call is tokens in, tokens out, plus latency and an error surface. Meta measured Muse Spark 1.3 using roughly 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2 for equivalent work — that reduction is worth more than a headline price cut of similar size.

4. Retries multiply everything. This is the one that dominates. On a task class where a cheap model succeeds 60% of the time and a frontier model succeeds 90%, you pay for roughly 1.67 attempts versus 1.11. That is a 50% cost multiplier applied to the cheap model, before counting the latency and orchestration complexity of failure handling.

Typical per-task output-token consumption on the AA Intelligence Index, September 2026:

Model	Output tokens per task
GPT-6 Astra	2,200 – 14,000 (by effort)
Claude Fable 5.1	14,500 – 45,000
Gemini 3.8 Flash	~48,000

The Other Lever: Cheap Cache Reads

Token efficiency is one way to win. Making re-read tokens nearly free is another, and it targets a different workload shape.

Claude Fable 5.1 cut cache reads to $0.25 per million in September 2026 — a 75% reduction from Fable 5 — while keeping the same $10/$50 headline. GPT-6 Astra cached input costs $1 per million, with writes around $12.50. Muse Spark 1.3 sits near $0.15.

Anthropic estimates the cache change alone makes typical workloads about 25% cheaper and highly agentic ones up to roughly 45% cheaper. For a long research sweep or a codebase-wide agent run, cache reads frequently make up the majority of the token bill.

So the two strategies compete on workload shape:

Output-heavy generation — drafting, synthesis, long-form reasoning — favours the token-efficient model.
Context-heavy agent loops — repeated passes over a large stable context — favours the cheap-cache model.

Measure your own input:output:cached ratio before assuming which applies.

How to Measure It Properly

Pricing pages cannot answer this because token consumption is a property of the model and your workload together. You have to measure.

Step 1 — Build a representative task set. A hundred tasks drawn from real production traffic, not curated examples. Include the hard tail; that is where models diverge.

Step 2 — Instrument everything. Per task, record: input tokens, output tokens, cached read tokens, cache write tokens, number of tool calls, wall-clock latency, and a binary success flag against your own acceptance criteria.

Step 3 — Compute cost per completed task. Total spend across all attempts, divided by tasks that succeeded. Retries count in the numerator; failures do not count in the denominator. This single number is your answer.

Step 4 — Sweep the effort dial. Models with low→max reasoning levels behave like several different models. GPT-6 Astra ranges from index 56.7 at 2,200 output tokens per task up to 61.1 at 14,000. Test at least three levels; the best cost-per-task point is often not the highest one.

Step 5 — Re-run after every model change. A “free upgrade” at identical per-token pricing can still raise your bill, because newer reasoning-tuned models frequently emit more output tokens for the same job.

The Practical Rules

Never compare models on a pricing page. It measures one of four cost drivers.

Never assume a same-price upgrade is free. Verify output-token volume after any model swap.

Route, do not pick. Classify each request and send bounded high-volume work to a cheap tier while reserving a frontier model for the hard tail. Teams doing this consistently land the majority of volume on the cheap tier while quality-sensitive work still reaches the top — reported per-report savings around 80% in some agent stacks.

Keep the model identifier behind a config flag. Tier swaps should be a deploy, not a refactor. This also protects you against scheduled repricing — Gemini 3.8 Flash’s $0.75/$3.75 is introductory and doubles to $1.50/$7.50 on January 1, 2027.

Budget at the post-promotional price. Introductory and promotional rates are the norm in 2026, not the exception.