Price per million tokens: 50% less for input in 2026

In its October 26, 2026 publication, Artificial Analysis compares GPT-6.1 Terra with GPT-5.6 Terra. The price per million input tokens is 50% lower, the output price is 58.3% lower, and maximum-effort cost per task falls 71.4%. The table does not change the route on its own. This compares generations, not historical readings of one model's tariff.

Key Facts / Decision Matrix: Input: $2.00 to $1.00 per million tokens. Output: $12.00 to $5.00 per million tokens. Maximum-effort cost per task: $1.40 to $0.40. Source: Artificial Analysis, published October 26, 2026; accessed November 9, 2026.

Key findings

Artificial Analysis compares prices and evaluation results for two generations in its October 26, 2026 publication, not customer bills. GPT-6.1 Terra's input price per million tokens is half the price for GPT-5.6 Terra. A test against the company's own workload remains necessary.

Artificial Analysis lists GPT-6.1 Terra input at $1.00 per million tokens, against $2.00 for GPT-5.6 Terra. Output costs $5.00 and $12.00 per million, respectively, a separate billing category. Cached input costs $0.05 versus $0.20 per million; cache writes cost $1.25 versus $2.50. At maximum effort, the source reports $0.40 versus $1.40 per task and a six-point gain in its Intelligence Index for the newer model. Price per million tokens, cost per task, and index scores use different units. None measures a company's consumption or establishes its contracted rate. A routing decision requires a task test, an acceptance criterion, and a calculation using the company's observed volumes and applicable rates.

1. Input: the newer generation has the lower rate in this comparison.
2. Output: a separate pricing category from input.
3. Cache: reads and writes are priced separately.
4. Cost per task: a result from an external evaluation.
5. Route: the decision remains conditional on task quality.

How were the numbers read?

Artificial Analysis published its comparison of GPT-6.1 Terra and GPT-5.6 Terra on October 26, 2026; it was accessed on November 9. The rates and evaluation results come from that single publication. The access date is not a second collection or a historical price series.

The October 26 analysis reports prices per million tokens and evaluation results for both models together. Input rates describe the price of a unit of consumption; maximum-effort cost per task is a result from the source's evaluation. Percentage reductions are calculated from the published values. The November 9 access date records when the publication was read, not another price observation. The publication compares different models and does not measure savings at Nexforce or any other company. It cannot establish a customer's bill or contracted rates because it does not describe that customer's workload. An enterprise must compare its applicable rates and observed consumption against the quality its task requires before changing a routing rule.

How much did token prices fall, and how did the output-to-input ratio change?

In the October 26 Artificial Analysis comparison, GPT-6.1 Terra input costs 50% less than GPT-5.6 Terra input. Output costs 58.3% less. Output remains more expensive than input, and the ratio between the two unit prices moves from six to five across the generations.

The source lists input at $1.00 per million tokens for GPT-6.1 Terra and $2.00 for GPT-5.6 Terra. Output is $5.00 and $12.00, respectively; the provider's model announcement also lists the newer model's rates. The input reduction is the difference between the two prices divided by the earlier price. The same calculation for output yields a rounded reduction of 58.3%. Dividing output by input gives five for GPT-6.1 Terra and six for GPT-5.6 Terra. That compares unit prices, not the number of tokens a task uses. The ratio is not the bill: spending depends on volumes and the rates that apply to the customer. The publication does not explain the commercial pricing decision or establish how much any customer would save by changing a route.

What do cost per task and quality results show?

In the Artificial Analysis evaluation, maximum-effort cost per task falls by 71.4% between the two models. The Intelligence Index improves in the same comparison. Those results do not guarantee performance on another workload. A team must test answers to its own task before choosing a route.

The October 26 publication reports maximum-effort cost per task of $0.40 for GPT-6.1 Terra versus $1.40 for GPT-5.6 Terra. The difference divided by the earlier cost gives a rounded 71.4% reduction for the tasks and effort considered in that evaluation. GPT-6.1 Terra gains six points in the Intelligence Index at maximum effort; at lower effort levels the gain is seven to thirteen points. The reported hallucination rate is 38% for the new model and 88% for its predecessor. The metric concerns wrong guesses when the model does not know an answer, not the error rate across every response. These results belong to an external evaluation. They neither measure the reader's workload nor guarantee the same difference in quality or cost for it. The index can help select candidates for a test, but a team must compare answers against a shared criterion and observe the cost of accepted answers.

The company's own task test determines the decision. A public index cannot replace it.

The study of orchestration and cost per task uses a different scope.

What can cache prices establish?

Artificial Analysis compares cache charges per million tokens for both generations. Cached input and cache writes have separate prices, but those rates alone do not establish a universal break-even point. A team must compare the expense with and without cache for its own workload.

The October 26 publication compares cached input prices of $0.05 per million tokens for GPT-6.1 Terra and $0.20 for GPT-5.6 Terra. Cache writes cost $1.25 and $2.50 per million, respectively. The model pricing documentation lists the newer model's write rate; each price describes a different charge. The calculated reductions are 75% for cached input and 50% for cache writes. The difference between rates alone does not establish break-even. The team must compare write and cached-input expense with the expense of the same workload without cache, using observed volumes and applicable billing rules. The publication does not provide a break-even point for each customer's workload or state how many reads pay for a write in every operation.

Output does not settle that calculation. Reuse frequency must enter the comparison.

The analysis of caching and routing examines a different cost attribution.

Why doesn't the bill follow the comparison, and how should the route be reviewed?

A lower rate does not change a route or prove a lower bill. The October 26 Artificial Analysis comparison is a reason to review the options. Teams should check applicable rates and test the same task against common acceptance criteria. Quality determines which candidates remain viable; observed cost then separates those that meet the requirement.

The publication provides prices per million tokens and a maximum-effort cost per task. It does not describe a customer's bill or change its routing rule. To review a route, the team should record the task and current rule, check the applicable rates, and test alternatives against the same acceptance criterion. The operational comparison uses observed input, output, and cache volumes rather than treating the price of one category as total spending. If a model fails the task's quality criterion, a lower rate does not justify switching, even if the alternative has a higher price. Among acceptable answers, observed cost can support the choice provided the record preserves consumption and the rates used in the evaluation. The publication does not isolate the causes of the cost-per-task difference and does not permit attributing it to one component.

The rate cannot choose the route. The answer must satisfy the task criterion.

The real cost of an AI application separates rates from spending. The session cost analysis uses a different unit from the task.

What do these numbers change in practice?

The Artificial Analysis comparison calls for a review of the routing rule, not an automatic model switch. Engineering needs to identify the task and evaluate the answer. The budget owner needs to connect consumption to the applicable rate. Nexforce Router supports monitoring without changing the published prices.

The analysis compares models from different generations. It does not prove savings in a company's operation or replace an evaluation of the task sent through the route. Before changing the rule, the team should define its acceptance criterion, compare answers, and record the rates and consumption observed for each alternative. Nexforce Router supports spending limits by key, agent, or project and centralized observability through logs, metrics, tracing, alerts, and dashboards. Those features support consumption monitoring. The source does not credit Router with measuring the models or changing published prices: it evaluates models, not customer routing configurations. If the answer fails the task criterion, the team keeps the route and investigates before considering a change.

Among alternatives meeting the required quality, observed cost supports the choice and should be recorded alongside the criterion used to accept the answer. The next decision rests on the test, not the price table. Afterward, the team monitors consumption and quality and reviews the route when a material change in the options warrants another test.

Frequently asked questions

The October 26, 2026 Artificial Analysis comparison covers GPT-6.1 Terra and GPT-5.6 Terra. Prices per million tokens and maximum-effort cost per task use different units. The results do not establish a market-wide decline or savings on a customer's bill.

Did the price of every model fall by 50%?

No. That generalization is not supported by the comparison. The 50% reduction refers to the price of input per million tokens for GPT-6.1 Terra, with GPT-5.6 Terra as the baseline in the Artificial Analysis publication. It does not establish a reduction for every model.

Does the 71.4% reduction apply to every task?

No. This percentage comes from one reference workload. The calculation uses maximum-effort cost per task reported by Artificial Analysis for the compared models. An enterprise task needs its own evaluation using applicable rates and observed consumption. Its acceptance criterion must match the company's workload.

Do the published prices establish when caching pays off?

Not without the volumes used by the task. Artificial Analysis reports prices for cached input and cache writes. Break-even depends on volumes and applicable billing rules. The team must compare the cost with cache against the cost of the same workload without reuse.

Does the source's index replace a company test?

No. The company needs to test its own workload. The six-point Intelligence Index gain at maximum effort belongs to the Artificial Analysis evaluation. Before changing the route, the team needs to assess answers to its own task, observe cost, and define an acceptance criterion.

References and further reading

The source of the prices is the Artificial Analysis publication dated October 26, 2026, accessed November 9. The linked internal studies cover cost per task, application, cache, and session under their own scopes. Each study uses its own unit. Source attribution and units prevent readers from confusing their results.

Artificial Analysis published the comparison between GPT-6.1 Terra and GPT-5.6 Terra on October 26, 2026; this article accessed it on November 9. A citation of rates should retain the per-million-token unit, while a citation of cost per task should identify the maximum effort used in the source's evaluation. The access date is not another price observation. Internal articles on applications, cache, sessions, and orchestration use their own scopes and do not replace the source of this comparison. Nexforce did not run the benchmark described by Artificial Analysis. None of the values measures a particular company's bill.

How should the results be cited and when should the route be reviewed?

Attribute the prices and evaluation to Artificial Analysis and include the October 26, 2026 publication date. Identify the models and the unit of each indicator. Route review is an operational decision: the team compares applicable rates and quality on its own workload. The company sets the cadence.

The October 26 publication compares GPT-6.1 Terra and GPT-5.6 Terra. The price reduction describes models from different generations; the maximum-effort reduction in cost per task belongs to the published evaluation, not each company's operation. Neither is a measurement of an enterprise bill. To record a routing decision, the team identifies the task, applicable rates, and observed consumption. It also records the quality criterion used to accept the answer. A quarterly review is an operational recommendation, not a finding measured by Artificial Analysis. The team should review earlier when a material change in the options warrants a new test. The record preserves the reason for keeping or changing the route.