This guide compiles published price changes from vendor pricing pages as reported by independent trackers in November 2026. Prices move often, some figures come from third-party trackers rather than vendors directly, and each one is attributed. Check the vendor’s own pricing page before you commit a budget.
AI pricing is about to move in both directions at once. On 1 December, Google doubles the price of its two newest Gemini models. A string of introductory offers ends within days of each other. At the same time, Anthropic, OpenAI and Alibaba have spent the autumn cutting prices on capable models. If your 2027 budget is built on November’s rates, some lines will be wrong by a factor of two, in either direction.
The practical answer is not to chase every change. It is to know which of your workloads sit on a model whose price is about to jump, and to move those before the invoice arrives. Here are the six changes that matter most, with dates, and what to do about each.
Table of Contents
- The AI Pricing Calendar for December and January
- Change 1: AI Pricing Doubles for Gemini 4 on 1 December
- Change 2: Gemini 3.8 Flash AI Pricing Doubles in January
- Change 3: Introductory AI Pricing Offers Are Ending
- Change 4: Tool Calls Join Your AI Pricing Bill
- Change 5: Rising GPU Rental Costs Hit Self-Hosted AI Pricing
- Change 6: Some AI Pricing Is Falling Fast
- Why Vendors Use Introductory AI Pricing
- Why Cost per Task Beats Headline AI Pricing
- A Simple AI Pricing Plan for 2027
- Who Should Act on AI Pricing Before 1 December
- FAQ About AI Pricing
- Verdict
- Sources and Further Reading

The AI Pricing Calendar for December and January
| Date | Change | Source |
|---|---|---|
| 21 November 2026 | GPT-5.6 Sol promotional rate of $4 / $20 ends | Release tracker |
| 25 November 2026 | Qwen4-Max preview model ID retires | Alibaba, via eWeek |
| 30 November 2026 | Last day of Gemini 4 introductory pricing; reported end of Qwen Token Plan new-user discounts | Google pricing page, DataCamp |
| 1 December 2026 | Gemini 4 Argon and Carbon double to $4 / $20 | Google pricing page |
| 1 January 2027 | Gemini 3.8 Flash doubles to $1.50 / $7.50 | Release tracker |
Prices are per million input and output tokens. Put these dates in a shared calendar now. Most AI pricing surprises are not secret; they are announced, and then forgotten.
Change 1: AI Pricing Doubles for Gemini 4 on 1 December
This is the biggest scheduled jump. Until 30 November, Gemini 4 Argon and Gemini 4 Carbon cost $2 per million input tokens and $10 per million output. From 1 December, they cost $4 and $20. BuildMVPFast’s analysis makes the key point: early users do not keep the introductory rate. Every paid account moves to the new price on the same day.
The doubling applies across tiers. For Carbon, according to BrainDetox’s copy of Google’s pricing table, batch and flex go from $1 and $5 to $2 and $10, priority goes from $3.60 and $18 to $7.20 and $36, and cached input goes from $0.10 to $0.20.
What to do: list every workload on either model. For each, decide whether it stays on Gemini at the new price, moves to batch or flex to halve the cost, or moves to another model. Do this before 1 December, not after the first invoice. Our LLM API cost optimisation guide covers the routing and caching patterns that make this easier.
Change 2: Gemini 3.8 Flash AI Pricing Doubles in January
Google’s cheaper workhorse is next. A November release tracker reports that Gemini 3.8 Flash doubles to $1.50 per million input tokens and $7.50 output on 1 January 2027.
Flash-tier models tend to carry high-volume, low-margin work: classification, extraction, routing and summarisation. Those workloads are exactly where a doubling hurts most, because they run millions of times. If Flash is in your pipeline, model the January cost now and test a cheaper alternative on your actual traffic.
Confirm this one on Google’s own pricing page before acting, since it comes from a tracker rather than a direct vendor announcement.
Change 3: Introductory AI Pricing Offers Are Ending
Launch discounts have become the norm, and several end in the next fortnight.
GPT-5.6 Sol. A release tracker reports that its $4 / $20 promotional rate ends on 21 November, with standard pricing from 22 November. The tracker did not state the standard rate, so check OpenAI’s pricing page.
Qwen4-Max. The preview model, available through Alibaba’s Token Plan subscription, retires on 25 November. DataCamp reports that discounted new-user prices of $3 and $5 a month on the two cheapest Token Plan tiers run through 30 November, against regular prices of $6 and $10. Pay-as-you-go API pricing for the full model is $1.50 input and $4.50 output.
The lesson is broader than these two. Whenever you adopt a new model, write down the date its introductory price ends, and set a reminder a week before. Launch pricing is a marketing tool, and it is designed to be forgotten.
Change 4: Tool Calls Join Your AI Pricing Bill
Token prices get the headlines, but tool calls are quietly becoming a significant cost for agent workloads.
Google Search grounding for Gemini includes 5,000 free requests a month, shared across Gemini 3 and newer models, then costs $14 per 1,000 requests, with a daily cap of 1,500. DeepSeek resumed billing for its Search and Fetch tools on 31 October at $0.75 per 1,000 calls, according to a release tracker.
For an agent that searches the web several times per task, those charges can rival the token bill. Track tool calls separately from tokens in your cost reporting, and cache search results where your use case allows it. Our guide to AI browser agents covers how different agents use web tools.
Change 5: Rising GPU Rental Costs Hit Self-Hosted AI Pricing
If you run your own models in the cloud, the trend is up. AIMultiple’s cloud GPU index reports that the on-demand median for the latest-generation group of GPUs rose from $2.12 per GPU-hour in October 2024 to $5.45 in November 2026. The November on-demand median for an H100 was $3.92 per GPU-hour, with the A100 at $1.94.
Consumer hardware is no escape. TechSpot’s fourth-quarter survey found graphics cards averaging 44% above their launch prices across ten countries, driven by the memory shortage. Running models locally has become more expensive to start, even if it is cheap to run once you own the hardware.
Consumer-class GPUs remain the bargain in the cloud: GetDeploying tracks a median of about $0.50 an hour for an RTX 4090 across 18 providers. For many inference workloads, a quantised open model on a rented consumer GPU still beats an API on cost. Our guide to LLM inference servers explains how to serve one.
Change 6: Some AI Pricing Is Falling Fast
The other direction matters just as much. Several capable models got cheaper this autumn.
| Model | Change | Current price per 1M tokens |
|---|---|---|
| Claude Fable 5.2 | 20% cut from Fable 5.1 | $8 / $40, cache reads $0.20 |
| GPT-6.1 Sol | Cached input half of GPT-6 Sol’s rate | $2 / $10, cached input $0.10 |
| Kimi K3.1 Pro | About a third cheaper than Kimi K3 | $2 / $10, cache hits $0.20 |
| Qwen4-Max | Cached input down 70% from the previous generation | $1.50 / $4.50, cached $0.075 |
If you are still on an older model in the same family, these are free savings. Fable 5.2 replaces Fable 5.1 at a lower price; GPT-6.1 Sol replaces GPT-6 Sol with cheaper caching. Test the newer model on your workload and switch if quality holds. Our Claude Fable 5.1 guide covers the older model’s breaking changes if you are migrating.
Why Vendors Use Introductory AI Pricing
It is worth understanding why so many launches now come with a temporary price, because the pattern is not going away.
A new model has to win developers away from whatever they already use. Switching takes effort: new prompts, new tests, sometimes new code. A low launch price lowers that barrier. Once a team has rebuilt its prompts and tests around a model, the cost of switching again is higher, so many will stay even after the price rises. That is the bet behind every introductory rate.
There is a capacity angle too. Early in a model’s life, providers often want lots of real traffic to tune serving, find bugs and build case studies. Cheap pricing brings that traffic. Later, when demand is established, prices rise to reflect what the model actually costs to serve.
None of this is sinister, and it is often a genuine bargain while it lasts. But it means the launch price of a model tells you very little about its long-term cost. Budget on the regular rate from day one, treat the introductory period as a cheap trial, and decide whether the model is worth its real price before the discount ends.
The reverse pattern, permanent price cuts on newer versions, usually comes from efficiency gains: better hardware, better serving software and smaller or more efficient models. Those cuts tend to stick. When you see one, it is usually safe to plan around it, which is why the falling AI pricing in change six deserves as much attention as the rises.
Why Cost per Task Beats Headline AI Pricing
Every number above is a price per token. What you actually pay depends on how many tokens a model uses to finish a task, and that varies enormously.
Two examples from this month make the point. Kimi K3.1 Pro cut its token price by a third, but used about 14% more output tokens per task than its predecessor, so its cost per task on the Artificial Analysis index barely moved. Qwen4-Max is cheap per token but verbose: on Vals’s index it cost $9.33 to complete a test that GPT-6.1 Sol finished for $3.61, despite Sol’s higher per-token price.
Thinking models make this worse, because reasoning tokens usually bill as output. A model that thinks twice as long at the same price costs twice as much per task. When a price doubles, as Gemini’s does on 1 December, a heavy-thinking model’s cost per task doubles with it.
So measure cost per completed task on your own workload, not price per million tokens on a vendor’s page. Run a fixed set of real tasks through each candidate model, record total spend and quality, and compare those two numbers. That is the only AI pricing comparison that predicts your invoice.
A Simple AI Pricing Plan for 2027
Here is a lightweight process that keeps AI costs predictable without a dedicated finance team.
Inventory. List every model you call, which product or workflow uses it, and roughly how many tokens and tool calls it consumes per month. Most teams discover at least one forgotten workload.
Date-stamp prices. Next to each model, record its current price, whether that price is introductory and when it ends. The calendar above is a starting point.
Benchmark quarterly. Keep a set of twenty to fifty real tasks with known good answers. Every quarter, run them through your current model and two alternatives, and record cost per task and quality. Switch when a cheaper option matches quality.
Use the cheap tiers. Batch and flex modes are typically half price. Anything that does not need an answer within seconds belongs there.
Cache aggressively. Cached input is 90% to 97.5% cheaper than fresh input across the models above. Long, repeated system prompts and documents should always hit the cache.
Keep a fallback. Have at least one alternative model tested and ready for each critical workload. When a price jumps or a provider has an outage, switching should take minutes.
Who Should Act on AI Pricing Before 1 December
| If you are | Start with | Why | Link |
|---|---|---|---|
| On Gemini 4 Argon or Carbon | Move non-urgent work to batch or flex, re-test alternatives | Standard rates double on 1 December | BuildMVPFast analysis |
| On Gemini 3.8 Flash | Model the January cost and test a cheaper option | Reported to double on 1 January | November tracker |
| On Claude Fable 5.1 or GPT-6 Sol | Test the newer version | Same family, lower price or cheaper caching | Fable 5.1 vs 5.2 |
| Running agents with web search | Track tool calls separately from tokens | Grounding and search calls now bill per thousand | Gemini pricing details |
| Self-hosting in the cloud | Compare consumer GPU rental with data-centre GPUs | Latest-generation medians have more than doubled since 2024 | AIMultiple GPU index |
Related on this site: for more context, see Claude Haiku 5.5, Copilot AI Credits Explained and GPT-6 Luna.
FAQ About AI Pricing
Which AI models are getting more expensive?
Gemini 4 Argon and Gemini 4 Carbon double from $2 / $10 to $4 / $20 per million tokens on 1 December 2026. A tracker reports Gemini 3.8 Flash doubling to $1.50 / $7.50 on 1 January 2027.
Do existing users keep the introductory Gemini price?
No. Every paid account moves to the new rate on the same day.
Which AI models got cheaper recently?
Claude Fable 5.2 is 20% cheaper than Fable 5.1, GPT-6.1 Sol halved cached input versus GPT-6 Sol, Kimi K3.1 Pro is about a third cheaper than Kimi K3, and Qwen4-Max cut cached input by 70%.
Is price per token the best way to compare AI pricing?
No. Cost per completed task is better, because models use very different numbers of tokens, especially when they think before answering.
How can I cut AI costs quickly?
Use batch or flex modes for non-urgent work, cache repeated prompts, move simple tasks to cheaper models and switch to newer, cheaper versions within the same model family.
Are GPU rental prices going up?
For the newest data-centre GPUs, yes. AIMultiple reports the on-demand median for the latest generation rose from $2.12 to $5.45 per GPU-hour between October 2024 and November 2026. Consumer GPUs such as the RTX 4090 remain much cheaper to rent.
Verdict
The AI pricing story for the end of 2026 is not “prices go up” or “prices go down”. It is that both happen at once, on specific dates, and the teams that save money are the ones that track those dates. Move Gemini 4 workloads to cheaper tiers or alternatives before 1 December, upgrade to cheaper newer versions where they exist, and start tracking tool calls as their own cost.
The caveat is that several dates here come from trackers rather than vendors, and prices can change again with little notice. Use this list as a prompt to check, not as a final answer.
Your next step: today, open your billing dashboards, list every model you pay for, and mark any that are on an introductory rate. For each one, write down the end date and the model you would switch to. That single list will prevent most December surprises.
Sources and Further Reading
- BuildMVPFast on the Gemini price change, for how the introductory rates expire for every account
- BrainDetox on Gemini 4 Carbon, for the full before-and-after pricing table and grounding charges
- GGUF Discovery’s November dispatch, for the deadline calendar and tool billing changes
- DataCamp on Qwen4-Max, for Token Plan pricing and the cost-per-task comparison
- AIMultiple’s cloud GPU index, for GPU rental price trends
On this site, read LLM API cost optimisation for practical savings, LLM inference servers compared if you self-host, and AI models compared to find alternatives. For a recent model breakdown, see GPT-6 Astra.

