What AI Pricing Did in 2026: Every Major Change, Dated and Sourced

I went to update the pricing on my own site this week and found five wrong numbers in a review I had written seven weeks earlier. Not typos. The prices had genuinely moved underneath me while the post sat there looking authoritative.

That is the state of AI pricing in 2026. It changes faster than anyone maintaining a website can keep up with, which is why almost every comparison article you read is quietly out of date, including the ones that look confident.

So this page is not another list of current prices. It is a dated record of what actually changed this year, what it cost before and after, and where I verified each number. Prices rot. Dated facts do not.

The pattern behind everything that happened

If you want one sentence for 2026, it is this: AI tools stopped selling subscriptions and started selling credits.

Not entirely, and not all at once. But the direction was consistent across every major vendor, and it followed the same shape each time. The conversational part of the product stayed flat-rate. The agentic part got metered.

The reason is not greed, or not only greed. A chat message costs a provider a fairly predictable amount. An agent that runs for twenty minutes, reads a hundred pages, calls six external tools, and retries twice costs something wildly different, and it varies enormously per run. Flat pricing works when usage clusters around an average. It breaks when a handful of users can each generate a hundred times the cost of a normal one.

Credits move that variance off the vendor's margin and onto your bill. Whether that feels fair depends entirely on which side of the average you sit on.

Keep that pattern in mind, because it explains almost every entry below.

The 2026 changelog

Dated, with what changed and what it cost before and after.

May 2025: Notion retires the standalone AI add-on

Before: any Notion plan plus roughly $10 per member per month for AI. After: full AI access only on Business, at $20 per member per month.

I am including a 2025 change because its consequences landed through 2026 and because it is the single most misreported price on the internet right now. Search for Notion AI pricing today and you will still find articles, including four that were on my own site until this week, quoting a $10 add-on that has not existed for over a year.

Nothing was announced as a price increase. Every published number technically still existed. But the combination that made the product affordable for a solo user, a cheap base plan with AI bolted on, was removed. If you were on Plus at $10 paying about $10 for AI, your path forward was a $20 per member Business plan you probably did not otherwise need.

This is the quiet version of a price rise, and it is worth learning to spot because it happened repeatedly this year. Full detail in my Notion AI review.

February 24, 2026: Notion announces Custom Agents

Cost at launch: free on Business and Enterprise.

Agents that run on a schedule or trigger, read your workspace, and reach out to Slack, Mail, Calendar, Figma, Linear, or a custom MCP server to complete multi-step work unattended. Business and Enterprise only.

Free to use through May 3. Remember that date.

March 18, 2026: Perplexity makes Comet free

Before: paywalled. After: free on iOS, Android, Windows, and Mac.

Comet is Perplexity's own browser, not a Chrome extension, and a lot of writing about it still gets that wrong. Dropping the paywall moved a genuinely useful assistant from a paid perk to a free one, which is the opposite direction from everything else on this list and worth noting for that reason alone. Covered in the Perplexity review.

April 1, 2026: ElevenLabs ships its biggest update of the year

API version 41.0 brought agent workflow controls, video-to-music generation, a Scribe v2 speech-to-text upgrade, and a voice library expansion adding over 10,000 voices through an IBM partnership.

No headline price change, but the plan contents shifted around this period, which is how my own review ended up with five stale figures. Sometimes the price stays and the product moves.

April 2026: Notion cuts agent running costs

Notion reduced the cost of running agents by a reported 35 to 50 percent and expanded connector access. If you evaluated agent workflows early in 2026 and concluded they were too expensive, that conclusion was priced against a different rate.

May 4, 2026: Notion Custom Agents go metered

Before: free. After: $10 per 1,000 Notion credits.

The free period ended on May 3. From May 4, Custom Agents consume credits sold as an add-on to Business and Enterprise.

The mechanics matter more than the rate. Credits reset monthly and do not roll over. Consumption scales with how much data an agent reads, how many tools it calls, how many steps it takes, and which model it uses. Admins get warnings at 80 and 100 percent, and agents pause automatically when the balance hits zero. Notion's own documentation puts example agents at roughly 30 to 60 runs per 1,000 credits, which works out to about 17 to 33 cents per run.

Downgrading the workspace to Free or Plus switches your existing Custom Agents off entirely, which is worth knowing before you spend a weekend building them.

June 1, 2026: GitHub Copilot moves to usage-based billing

Before: Copilot Pro at $10 per month, effectively unlimited. After: Copilot Pro at $10 per month including $15 of AI credits.

This is the clearest example of the pattern on the whole list, because the price did not move at all and the product changed underneath it.

Code completion and next-edit suggestions remain unlimited on paid plans and consume no credits. Chat, agent mode, code review, and Copilot CLI now draw from a monthly credit allowance. GitHub states the conversion plainly: 1 AI credit equals $0.01.

Current allowances are $15 monthly on Pro, $70 on Pro+, and $200 on Max. When you exhaust them you can wait for the next cycle, switch to cheaper models, or enable paid usage against a budget you set.

If you read a guide claiming Copilot's $10 plan is unlimited, it was written before June 1. Mine was, until this week.

July 9, 2026: ElevenLabs removes two legacy models

eleven_monolingual_v1 and eleven_multilingual_v1 were deprecated and removed. The migration path is eleven_multilingual_v2 or a current model.

This only affects you if you were calling those models directly through the API or had them saved in a project preset. Web app users will not have noticed. I include it because deprecations are a real cost that never appears in a pricing comparison: your bill did not change, but if you built on those models your afternoon did.

Through 2026: Perplexity adds a Max tier and credits

Perplexity now runs Free, Education Pro at $10, Pro at $20, and Max at $200 per month, with Enterprise Pro at $40 per seat.

Credits arrived here too, but with a distinction other vendors did not make: standard search and Ask consume no credits at all. Only the compute-heavy agentic features draw from the pool. Pro includes a one-time 4,000 credit bonus rather than a monthly allowance. Max includes 10,000 per month plus a 35,000 one-time bonus, and carries a default $200 monthly spending cap that can be raised to $5,000.

Leaving the core product uncapped while metering the expensive edges is, in my view, the most honest version of this transition that anyone shipped this year.

Through 2026: Cursor runs the model everyone else moved to

Cursor deserves a place on this list for getting there first and for showing what the destination looks like once the transition settles.

Its tiers are denominated in credit pools rather than usage limits. Pro at $20 a month carries a $20 monthly credit pool, Pro+ at $60 carries $60, and Ultra at $200 carries $400, the last of which is better value per dollar because of bundled overage discounts. Auto mode, which routes requests to cheaper models, consumes no credits at all and stays unlimited on every tier. Credits burn only when you manually select a frontier model, at roughly $0.04 to $0.20 per request depending on model and context size. MAX mode adds a 20 percent surcharge.

I have run this daily and the structure works better than it reads. Keeping Auto mode as the default for casual questions and reserving frontier models for genuinely hard problems means the credit pool lasts, and a $20 pool covers something like 300 to 500 frontier requests. Full breakdown in my Cursor review.

The reason it matters here is that Cursor is the proof that metered pricing is not automatically worse for the customer. It is worse when the vendor meters everything. It is fine, and arguably fairer, when the cheap path stays free and only the expensive path costs.

Through 2026: ChatGPT's tiers multiply

OpenAI went from a simple three-tier structure to something considerably more complicated. Alongside Free and Plus at $20, third-party sources now consistently report a Go tier at $8, two separate Pro tiers at $100 and $200, and Business pricing around $25 to $30 per user.

I could not confirm any of this from OpenAI directly, for reasons covered in the next section, so treat the specifics as unconfirmed. The direction is not in doubt though. A tier at $8 below Plus and two above it is market segmentation: pull in price-sensitive users at the bottom, capture heavy users at the top, and leave the familiar $20 anchor untouched in the middle so the headline never appears to move.

Worth knowing if your mental model of ChatGPT is still the 2023 version where there was Free and there was Plus. My ChatGPT Plus review covers whether the middle tier still earns its place.

Through 2026: Claude splits monthly and annual

Claude's pricing is worth reading carefully because the headline number depends on how you pay. Pro is $17 per month on an annual subscription billed at $200 up front, or $20 if you pay monthly. Max starts from $100. Team is $20 per seat per month billed annually or $25 monthly for standard seats, with a premium seat tier at $100 and $125.

No dramatic change here, and I include it as a caution rather than an event. If you see Claude Pro quoted at $17 somewhere and at $20 somewhere else, both are right and neither article told you which billing cycle it meant. That ambiguity runs through most pricing comparisons on the internet.

What did not change

Worth stating plainly, because a list of changes creates the impression that everything moved.

The $20 anchor held everywhere. ChatGPT Plus, Claude Pro on monthly billing, Perplexity Pro, and Cursor Pro all sat at $20 through the entire year, as they have since 2023. That number has become a convention rather than a calculation, and vendors appear reluctant to be the first to break it.

Free tiers also survived, and in one case improved. Every major tool on this list still has one. Perplexity's got better when Comet went free in March. The prediction that free AI would disappear once the funding tightened did not happen in 2026.

So the accurate summary is not that AI got more expensive. It is that the $20 stayed still while what sits behind it was quietly renegotiated.

How to read an AI pricing page without getting caught

Four habits, learned the hard way while assembling this.

Check which billing cycle the page is showing you. Pricing pages frequently default to the annual rate because it is the lower number, and comparison articles copy it as though it were monthly. Claude at $17 and Claude at $20 are the same plan. Perplexity Pro at $200 a year and at $20 a month are the same plan. If an article quotes one number with no cycle attached, it does not know either.

Look for the usage allowance behind the subscription. This is where the cost now hides. Copilot Pro at $10 sounds identical to what it was in May, and the $15 credit allowance behind it is the entire story. Ask what happens when you exhaust the included usage: does it stop, throttle, queue, or bill you.

Find out whether unused allowance rolls over. Almost universally it does not. Notion credits expire monthly. Perplexity credits do not roll over. ElevenLabs characters never did. This means buying a large buffer to be safe is the wrong instinct, because the buffer evaporates. Buy close to measured usage and top up.

Confirm the price from your own country. Covered below, and it caught me out during the research for this very article.

Current prices, with sourcing

Every row below carries where I checked it and when. Verify before you buy. This table starts aging the day it publishes and I would rather tell you that than pretend otherwise.

Two of those rows are marked unverified for a specific reason. OpenAI and Midjourney both returned HTTP 403 when I tried to read their pricing pages, twice each, from two different URLs. The figures I have for them come from third-party aggregators that broadly agree with each other, which is weaker evidence than a vendor's own page.

Most comparison articles would print those numbers without the caveat. I would rather you know which half of my table is load-bearing.

For the record, the third-party consensus on ChatGPT is Free, Go at $8, Plus at $20, Pro at $100 and $200, and Business around $25 to $30 per user. On Midjourney it is Basic $10, Standard $30, Pro $60, and Mega $120, with roughly 20 percent off annual billing. Treat both as probably right rather than confirmed.

The regional pricing nobody mentions

Something happened while I was researching this that belongs in every pricing article and appears in almost none.

I opened Grammarly's plans page and was served Indian pricing in rupees, on a plan the page itself described as exclusive to India, at roughly ₹392 per member per month on annual billing and ₹1,000 billed monthly, available only through local payment methods.

I did not ask for that. I got it because I am on an Indian connection.

This matters more than it sounds. Regional pricing is not simply the US price converted at the day's exchange rate. It is sometimes a different plan structure entirely, with different features, different billing methods, and different tier names. So when a comparison article tells you a tool costs $12 a month, the honest version of that sentence is that it costs $12 a month for someone browsing from the United States.

If you are outside the US, check the vendor's page yourself from your own connection before trusting any comparison, including this one. My table above reflects what a vendor served me, and I am in India.

What this means for your budget

Four things I would take away from the year.

Stop treating AI tools as fixed monthly costs. The useful mental model now is cloud infrastructure: a base fee plus a usage line you have to watch. If your tool has an agentic feature, it probably has a variable component behind it.

Read what the subscription includes, not what it costs. Copilot at $10 was a different product on May 31 than it was on June 2 at the same price. The sticker is no longer a reliable signal.

Set spending caps where vendors offer them. Perplexity Max defaults to a $200 monthly cap. Copilot lets you set a budget for paid overage. Notion pauses agents at zero rather than billing through. Use these. They exist because the alternative is a surprise.

Measure before you build. Every horror story about agent costs I have read this year has the same shape: somebody built five agents in week one and read the bill in week five. Build one, run it a full cycle, read the actual consumption, then extrapolate. If you are weighing tools by price, my under $20 per month guide covers what still fits a small budget after all of this.

How I verified this

Because you should ask that of anything claiming to be a pricing reference.

Every figure marked official came from the vendor's own pricing or documentation page, read on the date in the table. Where a vendor's page displayed both monthly and annual rates, I recorded both, because pricing pages frequently default to the annual figure and comparison articles frequently copy it as if it were the monthly one.

Where I could not reach an official page, the row says so. I did not fill those gaps with confident-sounding numbers from a competitor's blog and present them as verified. Two rows in that table are weaker than the others and they are labelled.

Dated changes came from vendor announcements and documentation where available. Where a fact came only from third-party reporting, I have said so in the entry.

I will not be revising this page monthly, and I want to be upfront about that rather than let it quietly rot while looking current. It is a record of 2026, dated on purpose. The changelog entries stay true forever. The snapshot table does not, which is exactly why every row carries the day I checked it.

If you are reading this well after July 2026, use the changelog and ignore the table.