LLM Routing: What to Do When the Token Price Changes

When the token price changes, the cost of a fixed LLM route changes with it, without notice. The procedure: re-read the AI model pricing table with a reading date, recalculate the fixed-route cost on the real consumption profile, compare it against the routed alternative, and set a re-evaluation cadence.

The notice comes from the data. Between the 2026-08-17 and 2026-09-07 readings on Artificial Analysis, GPT-5.6 Sol, the top-of-the-line model, cut its input price by 20% and its output price by 33%. Inside the same 21-day window, DeepSeek V4 Flash 0731, the economy category, rose 214% on input and 371% on output. Same window, opposite directions.

The premise that breaks is operational: the AI model pricing table stopped being an annual input and became an input that moves within the month. This guide runs the re-decision in five steps, with those two models as the running example, and shows where the middleware of the AI stack at scale absorbs the move without renegotiating a contract and without reintegrating code.

The full roadmap:

1. Pin down the consumption profile per model, separating input and output.
2. Read the pricing table with an annotated reading date.
3. Recalculate the fixed-route cost on the real profile.
4. Compare the fixed route against the routed route.
5. Set the re-evaluation cadence and the re-decision rule.

What you need before you re-decide the route

Route re-decision is a calculation with three inputs. Without a consumption profile, a percentage is abstract. Without a reading date, a comparison between tables is invalid. Without the current route policy, there is no baseline to measure what the change did. None of this demands a new tool, and no step depends on the model vendor.

Three items, in this order:

1. Consumption profile per model. Millions of input and output tokens per month, separated by model, read from the call trace or from the invoice. An aggregate total serves accounting, not route decisions.
2. Pricing table with a reading date. Input and output price per million tokens, the contracted tier, the source, and the day of the reading written next to the number, because a price without a date is not data, it is memory.
3. Current route policy. Which model serves which workload today, and under which rule, so the comparison in Step 4 has a baseline.

Step 1: Pin down the consumption profile per model

The consumption profile converts a percentage into money: millions of input and output tokens per model, per month, read from the call trace. With it, the repricing of the top model becomes US$400 a month on the route, US$300 from output and US$100 from input, and the economy model's becomes US$61.20, US$31.20 from output and US$30 from input, in the Step 3 example. Without the profile, a percentage is noise.

The profile lives in the gateway's audit trail of calls, in the vendor's consumption reports, or in the data team's spreadsheet. The split between input and output is invariant, because the two carry different prices and, in this window, moved in different directions: output carried the top model's 33% cut and the economy model's 371% increase, between the 2026-08-17 and 2026-09-07 readings from Artificial Analysis.

The workload mix enters too. Structured data extraction and document summarization are input-heavy workloads. Report and code generation are output-heavy. Two workloads with the same token total pay different invoices, and the profile is what shows which invoice is yours.

Step 2: Read the AI model pricing table with a reading date

The AI model pricing table is a document that moves, so a reading is valid for the day it was taken. Write down input and output price per million tokens, the contracted tier, the source, and the date. Two readings 21 days apart were enough to move both models in opposite directions.

The two readings of this guide:

Values per million tokens, at the highest tier listed on each date. Source: Artificial Analysis, readings of 2026-08-17 and 2026-09-07.

The current reading is that AI model prices only go down. This window is the refutation with two dated numbers: the top model cut price while the economy model raised it, inside 21 days.

A background trend is not an operation.

The same window carried a second movement, of scoreboard and not of price: Artificial Analysis re-based the Intelligence Index to version 4.2 on September 4, 2026. Score and price are separate readings; score-based re-decision lives in the article on the intelligence scoreboard re-base, and this guide stays on the pricing table.

In production, the dated reading is the input that feeds the routing policy; the model router pillar shows how the layer turns a reading into a rule.

Step 3: Recalculate the cost per token of the fixed route

The recalculation multiplies the consumption profile by the dated prices, twice: once with the old table, once with the new one. The difference is how much the fixed route's cost moved without a single decision from you. The example below uses the illustrative profile of 100 million input tokens and 30 million output tokens per month.

GPT-5.6 Sol, 2026-08-17 table: 100 million input tokens × US$5.00 = US$500; 30 million output tokens × US$30.00 = US$900. Total: US$1,400 a month.

GPT-5.6 Sol, 2026-09-07 table: input at US$4.00 and output at US$20.00 close the same route at US$1,000 a month: US$400 less, a 28.6% drop on the invoice, a percentage that is not printed on the table itself: the combined effect of the 20% input cut and the 33% output cut.

DeepSeek V4 Flash 0731, 2026-08-17 table: 100 million input tokens × US$0.14 = US$14; 30 million output tokens × US$0.28 = US$8.40. Total: US$22.40.

DeepSeek V4 Flash 0731, 2026-09-07 table: input at US$0.44 and output at US$1.32 close the same route at US$83.60 a month: US$61.20 more, a 273% increase on the invoice.

Price source: Artificial Analysis, readings of 2026-08-17 and 2026-09-07. Profile: illustrative.

The math exposes the error this guide fights. The repricing of the top model becomes US$400 a month on the route, US$300 from output and US$100 from input, and the economy model's becomes US$61.20, US$31.20 from output and US$30 from input: the larger percentage belongs to the smaller invoice. The secular fall of the token price stays real and long-term, but within three weeks the direction that matters is the direction of your profile, and it does not follow the market trend.

Step 4: Compare the fixed route with the routed route

The comparison turns the recalculation into a decision. The fixed route costs what the new table says, and the cost wins every time the pricing table moves. The routed route hands the re-decision to a layer that re-evaluates models on price and performance in real time. What is left for you is governance, not reading the table.

The table compares categories, and it deserves an honest reading. The routed route does not decide on its own which task tolerates the economy model: the policy is yours, with rules per key and spend caps. The layer executes. Where quality demands the top model, the rule pins that path and routing respects the rule, because the cost default is the wrong decision exactly where latency or quality are hard constraints. The routing policy in production exists to mark those paths before cost decides them.

The route choice looks like an engineering decision and gets taken as one, in the code; at the end of the month it shows up on the invoice as a margin decision.

The Nexforce Router is that layer: an LLM gateway with more than 300 models behind a single OpenAI-compatible API, smart routing by cost, performance, latency, and context, real-time model ranking by price and performance, automatic failover with configurable fallback, and model swaps without reintegration. Nexforce claims up to 50% savings on the cost per token, a number the company itself declares as a positioning ceiling, not as a verified result: the math on your route, done in Step 3, is what decides.

Before moving the route, the same prompt runs in parallel on the candidate models, and the decision comes from a measured result, not from a guess. The static economic argument for routing was made in the article on AI model cost; the comparison above adds time: the fixed route re-prices when the pricing table moves, and the routed route re-decides. The ruler of the LLM benchmark for CFOs, cost before score, applies line by line: the pricing table is its raw material.

Step 5: Set the cadence and the re-decision rule

Cadence is what stops the fixed route from expiring in silence again. Pin down three things: how often the pricing table gets read, the variation threshold that forces a recalculation, and the rule that decides between keeping the fixed route and moving to the routed route. A 21-day window was enough to move prices 33% and 371%.

Frequency: a dated read of the pricing table once a month, with the history kept, because comparing today's table against the one from six months ago without the intermediate reads hides the moment the route expired.

Threshold: a 10% variation in the fixed route's monthly cost forces a recalculation. Below that, noise. AI model price changes are recurring, and the trigger is the monthly cost, not the percentage on the table: this guide's window showed 371% costing less, in dollars, than 28.6%.

Rule: a path with a firm quality or latency requirement stays on a monitored fixed route; volume without a quality policy moves to the routed route. None of the three demands renegotiating a contract or reintegrating code. With the routing layer at scale in place, re-decision becomes policy at the gateway, and the pricing table goes back to being an input, not an event.

How to confirm the decision paid off

Verification is comparing the effective cost per token before and after, on the same consumption profile, with the trace of every call. Without a per-call trace, the comparison is opinion. With a trace, the route's effective cost per token shows up the following month, and the route decision becomes an auditable number.

The Nexforce Router delivers the trace verification asks for: every call auditable, central observability with logs, metrics, and dashboards, and savings and performance analytics per model. Three numbers close the verification. The month's effective cost per token, the invoice divided by the total tokens served. The share of calls that fell to the fallback, which reveals how many times the route's primary model was unavailable and how many times the fallback held the operation without anyone opening a ticket. Consumption against the spend cap, which shows whether governance held the month.

Verification of this kind is one instance of a wider discipline: the guide to reducing LLM inference costs maps the other levers, and this cadence feeds them with dated numbers.

The final number is one number.

Five common failures in route re-decision

The failures below are the five that the August and September 2026 window exposes hardest. Each has a one-line fix, and all of them share the same root: deciding the route with the wrong datum, a percentage without a profile, or a reading without a date.

Reading a percentage without a consumption profile. The repricing of GPT-5.6 Sol becomes US$400 a month on the Step 3 profile, US$300 from output and US$100 from input, and the repricing of DeepSeek V4 Flash 0731 becomes US$61.20, US$31.20 from output and US$30 from input. The percentage without the profile inverts the priority of the queue. Fix: no table number enters a decision without passing through the real profile.

Comparing readings without a date. Without a reading date, a comparison between two tables compares nothing: every number is valid for the day it was read, and the day is gone. Fix: source and date written next to every price, as in the 2026-08-17 and 2026-09-07 readings from Artificial Analysis.

Looking only at the input line. On the top model, output is the expensive line of the invoice; on this window's economy model, input takes over the bill. Both movements of this window were larger on output, -33% and +371% against -20% and +214% on input, and whoever read only the first line underestimated the bill. Fix: input and output in the recalculation, always.

Swapping models on price and losing the task. The economy model rose 371% and is still cheap in absolute value; the route question is never which model costs less per token, it is which one completes the task at the lowest cost per result. Fix: compare inside the same delivery quality, with the same prompt run in parallel before migrating.

Treating an index re-base as a price change. The Intelligence Index v4.2, re-based on September 4, 2026, moved the score scoreboard, not the pricing table. Whoever confuses the two readings re-decides the route over a movement that never touched cost. Fix: scoreboard and price as separate readings, each with its own date.

Token price and LLM routing: frequently asked questions

Why do AI model prices change?

Competitive repricing between vendors, compute cost that moves, and the restructuring of tiers and long-context pricing are the most common engines. No table comes with advance notice: the two readings of this window, 2026-08-17 and 2026-09-07 on Artificial Analysis, captured a cut of up to 33% and an increase of up to 371% with no explanatory note recorded on the price pages consulted. The price is the vendor's decision, and the defense is a reading cadence, not guesswork.

How much does the cost change when the token price changes?

It depends on the consumption profile, and the answer for each case: on the illustrative profile of 100 million input tokens and 30 million output tokens per month, the cut at the top model is worth US$400 a month and the increase at the economy model is worth US$61.20. A profile three times larger multiplies both values by three. A percentage without a profile does not answer. Ever.

Is it worth keeping a fixed route on one model?

It depends on the path, and the answer for each case: a path with a firm quality or latency requirement stays on a fixed route, monitored with the Step 5 cadence, because predictable behavior is worth more than the cost difference. Volume without a quality policy moves to the routed route, which absorbs the price change without rework.

What is LLM routing?

It is the layer that picks the model for each call on cost, performance, latency, and context, keeping the model ranking current and the failover ready. The model router as middleware is the pillar that details the full architecture, and the route decision becomes policy there, not an event on your calendar.

Referências e Leitura Complementar

• Artificial Analysis. GPT-5.6 Sol model page, with the input and output prices per million tokens used in every calculation in this guide, at the 2026-08-17 and 2026-09-07 readings: https://artificialanalysis.ai/models/gpt-5-6-sol
• Artificial Analysis. Intelligence Index v4.2, the intelligence scoreboard re-base announced on 2026-09-04: https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v4-2
• Artificial Analysis. Model portal and prices per million tokens, readings of 2026-08-17 and 2026-09-07, model DeepSeek V4 Flash 0731: https://artificialanalysis.ai

Take this into the next pricing table read

The AI model pricing table will keep moving, in both directions, without notice. What stays: the consumption profile pinned down, the dated reading, the fixed-route recalculation, the comparison against the routed route, and the re-evaluation cadence. Routing executes the re-decision as policy; the decision to keep, migrate, or monitor stays yours.

Two dated numbers open and close this article: in the same three-week window, between the 2026-08-17 and 2026-09-07 readings from Artificial Analysis, the top model became US$400 a month cheaper on the worked profile, and the economy model became US$61.20 more expensive. The fixed route that never re-evaluates pays for both, and the one that re-evaluates turns the pricing table back into an input. The architecture reference lives in the model router reference guide, and the layer that executes the re-evaluation is described on the Nexforce Router page.

The pricing table will change again. That part does not depend on you. Arriving at the next change with a consumption profile, a reading date, and a cadence does, and that is where the decision beats the surprise.