Is AI Getting Cheaper? 2026 API Prices, Tier by Tier
Is AI getting cheaper? Below the top tier, yes: Claude Haiku 5.5 and GPT-6 Luna cost $0.10/$0.50, and Claude Sonnet 5.5 and GPT-6.1 Sol $2/$10. But new flagships launched at $10/$50, and Gemini 4 Argon doubles on 1 Dec. 2026's API prices, tier by tier.
On this page
Is AI getting cheaper? Mostly, yes. Claude Haiku 5.5 runs our standard monthly workload for $55, a tenth of what Claude Haiku 4.5 costs. In the mid tier, Claude Sonnet 5.5 and GPT-6.1 Sol both cost $2/$10 per million input/output tokens, and GPT-6.1 Terra costs half that.
But the newest top-tier models launched at $10/$50, and Google has dated price rises for Gemini 4 Argon and two of its Flash models.
We checked every current price in this post against the vendors' own pricing pages on 4 Nov 2026. Earlier prices come from our price history, which logs each launch and price change with its date.
This is a market analysis, not a news roundup. For this month's changes, see our AI API pricing November 2026 roundup. Here we step back and ask what 2026's launches and cuts have done to a real bill.
Is AI Getting Cheaper? The Short Answer
It depends on the tier you buy. The three main tiers moved in different directions this year:
- Budget models fell hardest: Anthropic's cheapest model went from $1/$5 to $0.10/$0.50 when Claude Haiku 5.5 launched on 7 Oct 2026. OpenAI's cheapest current-generation model, GPT-6 Luna, costs the same $0.10/$0.50.
- Mid-tier models fell and met: Claude Sonnet 5.5 and GPT-6.1 Sol both cost $2/$10. Claude Sonnet 4.6, the last Sonnet before the 5 series, costs $3/$15.
- Flagships got dearer: In late April, the newest flagships from Anthropic and OpenAI were Claude Opus 4.7 at $5/$25 and GPT-5.5 at $5/$30. Claude Fable 5.1 and GPT-6 Astra now sit above them at $10/$50.
- Some prices are about to rise: Gemini 4 Argon goes from $2/$10 to $4/$20 on 1 Dec 2026. Gemini 3.8 Flash goes from $0.75/$3.75 to $1.50/$7.50 on 1 Jan 2027.
At Anthropic and OpenAI, a buyer who moves to the newest model in the same tier usually pays less per token, often much less. A buyer who always moves to the newest top model pays more. Google is the exception, as we show below.
How We Measured It
Per-token prices are hard to compare in your head, so we price one fixed job. It is 300 million input and 50 million output tokens a month, or about 100,000 requests averaging 3,000 input and 500 output tokens.
All figures are standard rates for short prompts, before batch or cache discounts. Where a vendor bills by time of day, we use the peak rate.
Every dated change is on our AI API price changes page, and our pricing history explorer charts each model month by month.
Budget Models: The Steepest Fall
The budget tier is where AI got cheaper fastest. These small, fast models handle classification, extraction, routing and simple chat, often at huge volume.
Anthropic made the biggest single move. Claude Haiku 5.5 launched on 7 Oct 2026 at $0.10/$0.50 for prompts up to 100,000 tokens, replacing Claude Haiku 4.5 at $1/$5.
On our workload, that is $55 a month against $550.
OpenAI got to the same floor in three steps. GPT-5.6 Luna went GA on 9 Jul 2026 at $1/$6, or $600 a month on our workload.
On 30 Jul, OpenAI cut it 80% to $0.20/$1.20, which is $120. Then GPT-6 Luna launched on 22 Sep at $0.10/$0.50, or $55.
The older GPT-5.4 Nano costs $0.20/$1.25, or $122.50 a month. OpenAI deprecated it on 1 Oct 2026 and names GPT-6 Luna as its replacement.
| Budget Model | Price per MTok | Our Workload a Month | Date |
|---|---|---|---|
| GPT-5.6 Luna at launch | $1/$6 | $600 | 9 Jul 2026 |
| Claude Haiku 4.5 | $1/$5 | $550 | Replaced 7 Oct 2026 |
| GPT-5.4 Nano | $0.20/$1.25 | $122.50 | Deprecated 1 Oct 2026 |
| GPT-5.6 Luna after its cut | $0.20/$1.20 | $120 | 30 Jul 2026 |
| GPT-6 Luna | $0.10/$0.50 | $55 | 22 Sep 2026 |
| Claude Haiku 5.5 | $0.10/$0.50 | $55 | 7 Oct 2026 |
Google moved the other way. Gemini 3.5 Flash-Lite, its newest Flash-Lite model, costs $0.30/$2.50, or $215 a month. That is more than the $0.25/$1.50, or $150, of Gemini 3.1 Flash-Lite, the model it supersedes.
The new floor also has a catch. Haiku 5.5's low rate stops at 100,000 tokens, and GPT-6 Luna's at 272K. For the full budget ranking, see our cheapest AI API rankings.
Mid-Tier Models: Rivals Met at $2/$10, Then OpenAI Went Lower
The mid tier is the clearest case of rival prices meeting in one place. It is also where many teams run their everyday production work.
Claude Sonnet 4.6 costs $3/$15, or $1,650 a month on our workload. We tracked its successor, Sonnet 5, at $2/$10 from 1 Jul 2026.
On 10 Aug, Anthropic made that launch rate permanent and cancelled a planned rise to $3/$15. Sonnet 5.5 kept the same $2/$10 when it launched on 28 Sep. That is $1,100 a month, a third less than Sonnet 4.6.
OpenAI's mid tier came down in steps. GPT-5.6 Terra went GA on 9 Jul at $2.50/$15, or $1,500 a month. OpenAI cut it 20% to $2/$12, or $1,200, on 30 Jul.
GPT-6.1 Sol then arrived on 29 Sep at $2/$10, the same as Sonnet 5.5.
Google's flagship joined them for a while. Gemini 4 Argon went GA on 13 Oct at an introductory $2/$10, so until 30 Nov three labs sell a model for exactly $1,100 a month on our workload.
GPT-6.1 Terra Halves the Middle Again
OpenAI then went lower. GPT-6.1 Terra launched on 26 Oct 2026 at $1/$5, half GPT-6.1 Sol's price. On our workload it costs $550 a month, less than half the $1,200 that GPT-5.6 Terra still costs.
Google's Flash Line Went Up
Google's Flash models sit below the $2/$10 group for now. Gemini 3.6 Flash launched on 21 Jul at $1.50/$7.50, or $825 a month.
On 13 Aug, Google halved it to an introductory $0.75/$3.75, or $412.50. Gemini 3.8 Flash launched on 2 Sep at the same introductory rate.
Both return to $1.50/$7.50 on 1 Jan 2027. Even at the introductory rate, they cost more than the older Gemini 2.5 Flash at $0.30/$2.50. Google's Flash line has moved up while the other labs' mid tiers fell.
Our Gemini API pricing page lists every Google rate.
Flagship Models: The Ceiling Went Up
So is AI getting cheaper at the very top? No. The labs added a new, pricier tier above the old one.
In late April, the newest flagships were Claude Opus 4.7 at $5/$25, which is $2,750 a month on our workload, and GPT-5.5 at $5/$30, or $3,000.
Then Anthropic launched Claude Fable 5 on 9 Jun 2026 at $10/$50, a new tier above Opus. OpenAI followed with GPT-6 Astra on 3 Sep at the same $10/$50. Both run our workload for $5,500 a month.
Claude Fable 5.1 replaced Fable 5 on 1 Sep at the same price. The most expensive model we track is still GPT-5.5 Pro, at $30/$180.
The tier just below the top did get cheaper. Claude Opus 5.5 launched on 22 Sep at $4/$20, 20% below Opus 5's $5/$25. That takes the Opus workload from $2,750 to $2,200.
GPT-5.6 Sol has been on a promotional $4/$20 since 21 Aug, down from its $5/$30 list price.
| Top-Tier Model | Price per MTok | Our Workload a Month | Launched |
|---|---|---|---|
| Claude Opus 4.7 | $5/$25 | $2,750 | 16 Apr 2026 |
| GPT-5.5 | $5/$30 | $3,000 | 23 Apr 2026 |
| Claude Fable 5 | $10/$50 | $5,500 | 9 Jun 2026 |
| GPT-6 Astra | $10/$50 | $5,500 | 3 Sep 2026 |
| Claude Opus 5.5 | $4/$20 | $2,200 | 22 Sep 2026 |
Google is heading up too. Gemini 4 Argon, its flagship reasoning model, costs $2/$10 until 30 Nov and $4/$20 from 1 Dec. Our Gemini 4 Argon pricing guide covers what you get for that.
Cost per Million Tokens Is Not the Whole Bill
The cost per million tokens is only part of the story. Three details can move a real bill more than a price cut does.
Newer Claude Models Count More Tokens
Anthropic's pricing documentation says Claude 4.7 and later models use a newer tokenizer that "produces approximately 30% more tokens for the same text". Claude Sonnet 4.6 and earlier models use the previous one.
So moving from Sonnet 4.6 to Sonnet 5.5 cuts the price per token by a third, but the same text becomes about 30% more tokens. Our $1,100 job would then cost about $1,430, still about 13% below Sonnet 4.6's $1,650.
The budget move survives it easily. With 30% more tokens, Haiku 5.5 would cost about $71.50 for the job, still about 87% less than Haiku 4.5. Our guide to what a token is in AI explains why counts differ.
Long Prompts Cost More
Several models bill the whole request at a higher rate once a prompt passes a size. GPT-6 Luna rises from $0.10/$0.50 to $0.20/$0.75 above 272K tokens, and GPT-6.1 Sol from $2/$10 to $4/$15.
Claude Haiku 5.5 moves to $0.50/$2.50 above 100,000 tokens.
Claude Sonnet 5.5 has no such band, because Anthropic bills its full 1M-token window at standard rates. For retrieval jobs that send long documents, that can matter more than the headline price.
Caching Got Much Cheaper
Cache reads, the rate for repeated input served from a cache, fell faster than list prices.
On 1 Sep, Claude Fable 5.1 launched with cache reads cut 75% to $0.25 per million tokens. On 7 Oct, Anthropic halved Sonnet 5.5's cache reads from $0.20 to $0.10. GPT-6.1 Sol also charges $0.10, which is 5% of its input price.
If much of your input repeats, such as a long system prompt, the cache rate can matter more than the list rate. See our Claude API pricing page and OpenAI API pricing page for every cache rate.
Every AI API Price Increase on the Calendar
Falling prices are not guaranteed. Several of today's lowest rates are introductory or promotional, and the vendors have said when they end or how long they run. These are the dated AI API price increase risks for the models we track:
| Date | Model | Price Now | Price After | Our Workload a Month |
|---|---|---|---|---|
| Promotion runs at least through 21 Nov 2026 | GPT-5.6 Sol | $4/$20 | $5/$30 list price | $2,200, then $3,000 |
| 1 Dec 2026 | Gemini 4 Argon | $2/$10 | $4/$20 | $1,100, then $2,200 |
| 1 Jan 2027 | Gemini 3.8 Flash and 3.6 Flash | $0.75/$3.75 | $1.50/$7.50 | $412.50, then $825 |
OpenAI's API pricing page says GPT-5.6 Sol's promotional pricing "is available at least through November 21, 2026". That is a floor, not an end date.
Google's Gemini API pricing page prints both prices with their dates. For Argon's input it reads "$2.00 through November 30, 2026" and "$4.00 starting December 1, 2026".
A dated rise can still be cancelled. Claude Sonnet 5 was due to rise to $3/$15 on 1 Sep 2026, until Anthropic made its $2/$10 rate permanent on 10 Aug. Budget at the higher price anyway, and treat a cancellation as a bonus.
DeepSeek's Peak-Hour Price Rise
The biggest price rise we've recorded this year came from DeepSeek. On 16 Aug 2026, it raised prices and started peak and off-peak billing. At peak, V4 Flash went from $0.14/$0.28 to $0.44/$1.32, and V4 Pro from $0.435/$0.87 to $1.32/$3.96.
On our workload, V4 Pro went from $174 to $594 a month at peak. DeepSeek bills half the peak rate off-peak, so the same job costs $297 at off-peak hours.
DeepSeek then eased part of the Flash rise. On 10 Sep, V4.1 Flash replaced V4 Flash at $0.30/$1.20 at peak, or $150 a month, and $75 off-peak.
Peak runs from 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, outside Chinese public holidays. Our DeepSeek API pricing page has the details.
What the Price Moves Mean for Your Budget
Averages hide what matters to one team. These are the moves our standard workload would see if you switched within the same family today:
- Claude Haiku 4.5 to Haiku 5.5: From $550 to $55 a month, a saving of $495. Allow for about 30% more tokens, and it is still about $71.50.
- GPT-5.6 Terra to GPT-6.1 Terra: From $1,200 to $550 a month, a saving of $650.
- Claude Opus 5 to Opus 5.5: From $2,750 to $2,200 a month, a saving of $550.
- Staying on Gemini 4 Argon: From $1,100 to $2,200 a month on 1 Dec, an extra $1,100.
Every saving on that list needs a migration. Newer models can behave differently, so test your prompts before you move production traffic. The one rise on the list needs nothing from you at all.
That is the real shape of the market in late 2026. The savings are large, but they go to buyers who keep moving, and the rises fall on buyers who stand still.
Is There an AI Price War?
Below the top tier, our data says yes. Anthropic and OpenAI now price a model identically at four levels: $0.10/$0.50, $2/$10, $4/$20 and $10/$50. At the three lower levels, both got there by bringing prices down.
But this AI price war is mostly about matching, not undercutting:
- GPT-6.1 Sol matched Sonnet 5.5: OpenAI launched it on 29 Sep at Claude Sonnet 5.5's $2/$10.
- Haiku 5.5 matched GPT-6 Luna: Anthropic launched it on 7 Oct at Luna's $0.10/$0.50, on prompts up to 100,000 tokens.
- GPT-6 Astra matched Claude's top tier: OpenAI launched it on 3 Sep at the same $10/$50 as Claude Fable 5.1.
Matching prices change how you choose. When two models cost the same per token, the decision comes down to quality, context limits, cache rates and token counts. Our AI API comparison table puts those side by side.
Frequently Asked Questions
Are AI APIs Getting Cheaper in 2026?
Yes, below the top tier. Anthropic's cheapest model and OpenAI's cheapest current-generation model both cost $0.10/$0.50. Claude Sonnet 5.5 and GPT-6.1 Sol both cost $2/$10, and GPT-6.1 Terra costs $1/$5. The newest top-tier models cost $10/$50, more than April's $5/$25 and $5/$30 flagships.
Which AI API Prices Go Up Next?
Gemini 4 Argon goes from $2/$10 to $4/$20 on 1 Dec 2026. Gemini 3.8 Flash and 3.6 Flash go from $0.75/$3.75 to $1.50/$7.50 on 1 Jan 2027. GPT-5.6 Sol's promotional $4/$20 runs at least through 21 Nov 2026, and its list price is $5/$30.
What Is the Cost per Million Tokens for a Budget Model Now?
GPT-6 Luna and Claude Haiku 5.5 both cost $0.10 per million input tokens and $0.50 per million output tokens. Haiku 5.5's rate applies to prompts up to 100,000 tokens. On the Batch API, both halve to $0.05/$0.25.
Why Is My AI Bill Going Up If Prices Are Falling?
Usually for one of four reasons. You moved to a pricier top-tier model, or a newer Claude model counts your text as more tokens. Your prompts may have crossed a long-prompt threshold, or a promotional rate may have ended.
Is DeepSeek Still a Cheap AI API?
Off-peak, yes. DeepSeek V4.1 Flash costs $0.15/$0.60 off-peak and $0.30/$1.20 at peak. That is cheaper than V4 Flash after the August rise, but dearer than the $0.14/$0.28 it cost before.
Fun Fact: A Hundredfold Gap in One Family
GPT-6 Luna and GPT-6 Astra are siblings in OpenAI's GPT-6 family. On our standard workload, Luna costs $55 a month and Astra costs $5,500, exactly a hundred times more. The model you pick moves your bill far more than any price cut this year.
So, Is AI Getting Cheaper for Your Workload?
Is AI getting cheaper? For most workloads, yes.
On Claude's budget and mid-tier models, and on OpenAI's budget models, the newest model runs the same job for less than the spring's line-up did, even after the newer Claude tokenizer. If you always move to the newest top model, or you rely on Google's introductory rates, expect to pay more.
Three habits keep the savings:
- Budget at the list price: Check whether your model is on an introductory or promotional rate, and plan for the price after it.
- Re-test in your tier: The cuts land on the newest model in each tier, not on the one you already use.
- Price your own mix: Run your token counts through our AI pricing calculator before you switch.
