Budgeting AI Agents on Intro Pricing Built to Double

Budgeting AI agents on introductory pricing is now a discipline of its own, because one of the most attractive workhorse rates on the market ships with a published expiry date. Gemini 3.7 Flash bills $0.75 per million input tokens and $3.75 per million output through December 31, 2026 — and Google’s own pricing page posts $1.50/$7.50 from January 1, 2027. Exactly double, on a date you can put in a calendar.

The stakes are simple: on the posted schedule, an agent fleet sized against the intro rate is a fleet that costs twice as much the week the window closes. And the calendar trap is only one of two mechanics currently separating the headline rate from the rate you actually pay — Grok 4.6 and GPT-5.6 Terra reprice an entire request the moment a prompt crosses a token threshold, no calendar required.

This playbook covers the full Gemini Flash schedule (including the detail most coverage missed: 3.6 Flash carries the identical step-up date), the one budgeting rule that makes intro pricing safe to use, the Anthropic reversal that proves step-ups are not inevitable, the threshold traps, and a trigger matrix you can hand to whoever owns the forecast.

Start with what Google actually posted. The Gemini Developer API pricing page lists Gemini 3.7 Flash at $0.75 per million input tokens and $3.75 per million output “through December 31, 2026,” with a standard rate of $1.50/$7.50 “starting January 1, 2027.” The intro rate is exactly half the standard rate Google has posted for January — 3.7 Flash’s launch-day numbers cover how that price landed against the rest of the market on August 13.

The detail most launch coverage skipped: the same page posts the identical schedule for Gemini 3.6 Flash — same $0.75/$3.75 intro figures, same December 31 window, same $1.50/$7.50 from January 1. The January doubling is not a 3.7-Flash-specific event. It is a Google-Flash-line-wide repricing date, which means teams already running Gemini 3.6 Flash as their workhorse face the same step-up whether or not they ever migrate to 3.7.

The tier structure moves in lockstep, which matters for anyone modelling batch or priority traffic. Batch is a flat 50% of Standard at every point in time: $0.375/$1.875 through December 31, then $0.75/$3.75 from January on the posted schedule — meaning batch traffic in 2027 would cost the same as standard traffic does today. Flex pricing is numerically identical to Batch on the current page, and Priority runs a constant 1.8× Standard: $1.35/$6.75 now, $2.70/$13.50 from January. Because every tier is a fixed multiple of Standard, the doubling propagates through the whole surface — there is no tier that escapes it.

One neighbouring model worth noting: Gemini 3.5 Flash-Lite sits at $0.30 input / $2.50 output with no introductory language and no expiry date anywhere in its pricing section. If you need a Google-side rate you can forecast without a step-up scenario, that is currently the stable one.

The rule this whole playbook hangs on: size the fleet against the posted standard rate, and treat the intro window as margin. If the January step-up executes, your forecast already covers it. If it gets cancelled — a real possibility, as the next section shows — you keep the difference as found budget. The only losing move is the common one: sizing agent volume against $0.75/$3.75 and discovering in January that the same traffic costs twice as much.

To make that concrete, here is a blended-cost comparison. One explicit assumption first: the table assumes a 4:1 input:output token split — a round number chosen for illustration, not a measured average. No vendor publishes a “typical” agent workload ratio, and agent fleets vary enormously (tool-heavy loops skew input-heavy; generation-heavy ones do not). Re-run the arithmetic with your own ratio before using any of these blended figures.

Two honest readings of that table. First, the doubling is real money at fleet scale: at the illustrative 4:1 split, a fleet processing one billion total tokens a month — a hypothetical volume, chosen only to make the arithmetic visible — pays roughly $1,350 a month at the intro rate and $2,700 at the posted standard rate. That is about $16,200 a year of difference riding on one calendar date.

Second — and this is the caveat that keeps the framing honest — even the post-doubling rate is not expensive by current standards. At $1.50/$7.50, 3.7 Flash’s posted 2027 price still undercuts GPT-5.6 Terra’s current $2/$12 on both legs and Sonnet 5’s permanent $2/$10 on both legs. The trap is not that the standard rate is bad; it is that a forecast built on the intro rate silently assumes the discount lasts forever. Whether the model itself earns a place in your stack is a separate question — how it actually performs against Sonnet 5 and GPT-5.6 Terra is its own analysis.

Here is why this post does not tell you the January doubling is certain: the most prominent scheduled step-up of the summer just failed to execute. Claude Sonnet 5 launched on June 30, 2026 at $2 per million input / $10 per million output, explicitly framed as introductory pricing through August 31 — with a posted increase to $3/$15 slated for September 1. On August 11, three weeks before it would have taken effect, Anthropic called it off. Its platform pricing documentation now records the reversal in plain terms.

Read as a precedent, this cuts both ways. It proves posted step-ups are real enough to plan around — Anthropic published one, with dates and figures, and clearly intended it. And it proves they are soft enough to reverse — most plausibly under competitive pressure, in a month when Google priced a capable Flash model at $0.75 input. A budget that assumed Sonnet 5 would cost $3/$15 from September now has margin; a budget that assumed $2/$10 forever happened to be right, but only by luck.

One housekeeping note if you follow our pricing coverage: the August pricing tracker went to press on August 5 and frames Sonnet 5’s $2/$10 as a promo with a dated cliff — accurate when written, superseded by the August 11 reversal. The story moved; that is rather the point of budgeting this way. Treat every posted schedule, including Gemini’s, as the currently posted plan — an input to a forecast, not a fact about the future.

Gemini’s trap is calendar-shaped: the rate changes on a date. The second trap is threshold-shaped: the rate changes when a single request gets big enough — and it reprices the whole request, not just the overflow. Two vendors run this mechanic today, with different numbers, and the difference matters.

Grok 4.6 bills $2/$6 per million with cached input at $0.50 for prompts under 200K tokens. At or above 200K, the rate becomes $4/$12 with cached input at $1.00 — and xAI’s pricing docs are explicit that “requests whose prompt reaches the listed token threshold are billed at the higher rate for all tokens in the request.” We covered Grok 4.6’s 200K-token whole-request repricing in depth at launch.

GPT-5.6 Terra runs the same structure with different parameters: $2/$12 per million for prompts up to 272K input tokens, and per OpenAI’s Terra model documentation, “Prompts with >272K input tokens are priced at 2x input and 1.5x output for the full request.” Keep the two distinct: Grok’s threshold is 200K with a 2× multiplier on both legs; Terra’s is 272K with 2× input but only 1.5× output. Same shape, different cliff heights.

The Luna card above is a third, quieter variant of the same lesson: the rate you see quoted is not always the rate for the surface you run. Aggregator listings have circulated Luna’s $0.10/$0.60 batch rate as if it were the standard price — OpenAI’s own table puts standard short-context at $0.20/$1.20. We hit the same surface-confusion pattern with DeepSeek V4 Flash list prices versus OpenRouter aliases in the cheap-tier repricing wave. Always price the surface you will actually call.

Most pricing trackers list rates. What a budget owner actually needs is the trigger — the specific event that moves a workload from the headline rate to the second one. This matrix puts the two mechanics side by side, plus the two entries every forecast should carry as context: the step-up that got cancelled, and the rate that gets misquoted.

The pattern worth internalising: across four vendors, the headline rate and the effective rate have formally decoupled. Google decouples them with a date, xAI and OpenAI with a token count, and aggregators do it accidentally with surface labels. The common failure mode is the same in every case — a forecast keyed to the number on the marketing page rather than to the trigger conditions in the pricing documentation. Trackers list prices; budgets need triggers.

Agent workloads are unusually cache-friendly — system prompts, tool schemas, and shared context repeat across every loop iteration — so Gemini’s caching schedule deserves its own budget line. Two separate prices exist here, in different units, and conflating them corrupts a forecast. The cache read price — what you pay when a cached token is served — is $0.075 per million tokens through December 31, doubling to $0.15 from January 1. The cache storage price — what you pay to keep tokens cached — is $0.50 per million tokens per hour through December 31, doubling to $1.00 per million per hour from January. One is per-token; the other is per-token-per-hour. They are different line items on the same page.

The genuinely useful nuance: the cache-read discount ratio is constant across the doubling. At $0.075 against a $0.75 input rate, cached reads cost 10% of base input; at $0.15 against $1.50, still exactly 10%. So while every price on the schedule doubles in absolute terms, a workload that serves a large share of its input from cache sees a smaller relative jump in blended cost than a cache-cold workload — the cached share was already riding at a tenth of the input rate and keeps that ratio. Caching does not cancel the doubling; it compresses it for the cached fraction of your tokens.

Treat that quote for what it is: a customer testimonial Google selected for its own product page, not an independent benchmark. But its shape is instructive — the cost win it describes comes partly from cache behaviour, and cache-hit rate is exactly the variable that will decide how hard the January step-up lands on any given fleet. Measuring your own hit rate now, while the intro window is open, is what turns the 2027 line of your forecast from a guess into arithmetic.

The window runs about four and a half more months. Here is how we would spend them, by workload shape.

Model choice interacts with all of this. If the workload is genuinely subagent-shaped — high volume, narrow scope — the stable-priced tier below Flash may fit better than the discounted tier above it: the purpose-built subagent tier carries no expiry language at all, and bulk-workload cost math often favours models whose list price is boring. Boring is a feature in a forecast.

Looking forward, we expect more of this, not less. Dated intro windows let vendors buy market share without permanently repricing the line, and threshold repricing lets them defend margin on the long-context traffic that costs them the most to serve. Both mechanics reward the same operational habit: pricing review as a scheduled discipline rather than a launch-week event. If your team wants help building that discipline — fleet cost models, router policies, trigger-aware budget alerts — our AI transformation engagements start with exactly this kind of spend architecture.