AI API Pricing September 2026: What Changed This Month
OpenAI launched GPT-6 Astra at $10/$50, Claude Fable 5.1 cut cache reads by 75%, and DeepSeek's peak-hour pricing went live. Every AI API price change in September 2026, verified against vendor pricing pages.
On this page
AI API pricing moved in both directions in September 2026. OpenAI launched GPT-6 Astra at $10/$50 and cut GPT-5.6 Sol to a promotional $4/$20. Anthropic's Claude Fable 5.1 held its sticker price but cut cache reads by 75%. Google halved its newest Flash rates until year end, and DeepSeek's price rise and peak-hour billing went live.
Two of those changes do not show up in a rate-card comparison at all. The cost of running Claude's top model fell without its price moving, and DeepSeek's bill now depends on what time of day your job runs. Here is every material change since our August 2026 roundup, with rates as they stand on 15 September 2026.
AI API Pricing in September 2026: The Short Version
- OpenAI launched GPT-6 Astra on 3 September at $10/$50, double the old GPT-5.6 Sol rate and the new ceiling for its general model line.
- GPT-5.6 Sol is on a promotional $4/$20 since 21 August, down from $5/$30, and OpenAI commits to it only through 21 November 2026.
- Claude Fable 5.1 replaced Fable 5 on 1 September at the same $10/$50, with cache reads cut from $1.00 to $0.25 per million tokens.
- Claude Sonnet 5 stayed at $2/$10. The increase to $3/$15 that was once scheduled for 1 September did not happen.
- DeepSeek raised prices on 16 August and now bills peak and off-peak rates, with off-peak at half of peak.
- Google shipped Gemini 3.8 Flash on 2 September at an introductory $0.75/$3.75, half its 2027 list price.
- xAI's Grok 4.6 arrived on 12 August at Grok 4.5's token rates, but with cached input at $0.50 instead of $0.30.
Every AI API Price Change Since Our August Roundup
All figures are US dollars per million tokens, input then output, at standard rates.
| Date | Provider | What changed | Rate now |
|---|---|---|---|
| 12 Aug | xAI | Grok 4.6 launches as flagship | $2.00 / $6.00 |
| 13 Aug | Gemini 3.6 Flash halved, 3.7 Flash GA at the same rate | $0.75 / $3.75 | |
| 16 Aug | DeepSeek | Price rise plus peak and off-peak billing begins | V4 Pro $1.32 / $3.96 peak |
| 21 Aug | OpenAI | GPT-5.6 Sol cut to a promotional rate from $5.00 / $30.00 | $4.00 / $20.00 |
| 1 Sep | Anthropic | Claude Fable 5.1 replaces Fable 5, cache reads cut 75% | $10.00 / $50.00 |
| 1 Sep | Anthropic | The scheduled Claude Sonnet 5 increase does not take effect | $2.00 / $10.00 |
| 2 Sep | Gemini 3.8 Flash GA on introductory pricing | $0.75 / $3.75 | |
| 2 Sep | Meta | Muse Spark 1.3 ships at an unchanged rate | $1.25 / $4.25 |
| 3 Sep | OpenAI | GPT-6 Astra launches, on Azure the same day | $10.00 / $50.00 |
| 10 Sep | DeepSeek | V4.1 Flash replaces V4 Flash and cuts Flash rates | $0.30 / $1.20 peak |
How We Verified These Prices
Every rate above was read from the provider's own pricing page or API documentation on 15 September 2026, not from a summary or a third-party tracker. Where a change had no documentation entry, such as the Gemini 3.6 Flash cut, we dated it from archived copies of the same vendor page and say so. We treat four things as a price change: the list rate, the cache and batch multipliers, the long-context threshold, and time-of-day billing. Our pricing history explorer holds the month-by-month record behind every figure.
New at the Top: GPT-6 Astra Sets OpenAI's Premium Tier
GPT-6 Astra arrived on 3 September at $10.00 input and $50.00 output per million tokens, described by OpenAI as its most capable model. It carries a 1,050,000-token context window and 128,000 maximum output tokens.
Three details matter more than the headline rate. Cached input bills at $1.00 and cache writes at $12.50, which is 1.25 times the input rate. Above 272,000 tokens the price doubles to $20/$75 for the whole request, not just the tokens over the line. Batch and Flex both run at 50% of standard, so a batched Astra job costs $5.00/$25.00. Current rates are on OpenAI's API pricing page and our OpenAI pricing page.
Astra is also the first OpenAI model in a while to launch above the tier it tops rather than at the same price. When Sol replaced GPT-5.5 in July, it inherited $5/$30. Astra doubles that.
What Happened to GPT-5.6 Pricing This Month
On 21 August, three weeks after cutting Luna and Terra, OpenAI cut Sol as well. The flagship moved from $5.00/$30.00 to $4.00/$20.00, a 20% reduction on input and 33% on output, with cached input at $0.40.
This one comes with a clock. OpenAI describes the rate as promotional and available "at least through November 21, 2026". That is a floor, not an end date, so the rate may run longer. It may also revert. If you are modelling 2027 spend on Sol, model $5/$30 and treat the current rate as a windfall.
The rest of the family is unchanged: Terra at $2.00/$12.00 and Luna at $0.20/$1.20. Luna remains the cheapest current-generation OpenAI model, though the older GPT-5 nano is cheaper still at $0.05/$0.40 until its snapshot shuts down on 11 December 2026. The full tier list is on our ChatGPT and OpenAI API pricing guide.
Same Sticker, Smaller Bill: Claude Fable 5.1 Cuts Cache Reads
Anthropic released Claude Fable 5.1 on 1 September at $10.00/$50.00, exactly what Fable 5 charged. A rate-card comparison would record no change at all.
The change is in caching. Every other Claude model bills a cache read at 10% of its input rate. Fable 5.1 bills it at 2.5%, which is $0.25 per million tokens against Fable 5's $1.00. Anthropic's pricing documentation states the exception plainly, and it applies to Claude Mythos 5.1 as well.
That matters for the workloads people actually run on a top-tier model: long system prompts, shared documents, repeated codebases. On a job where 80% of input comes from cache, the discount is worth roughly 5% of the monthly bill. Useful rather than dramatic, and it arrives without a migration.
Cache writes still cost 1.25 times input for five minutes and 2 times for an hour, so caching only pays once content is read back. Fable 5 stays available at $10/$50 as a legacy model, and the rest of the line is unchanged: Opus 5 at $5/$25, Sonnet 5 at $2/$10 and Haiku 4.5 at $1/$5. See the Claude API pricing overview or our explainer on what the Fable tier is for.
Claude Sonnet 5: The 1 September Increase Did Not Happen
Last month we reported that Anthropic had cancelled the Sonnet 5 rise. The deadline has now passed without incident. Sonnet 5 bills $2/$10, and Anthropic's documentation still states that the increase to $3/$15 will not occur.
This is worth confirming because stale advice outlived the deadline. Comparisons written in early August told readers to act before 31 August. If anything in your own planning still carries that date, delete it. Sonnet 5 undercuts GPT-5.4 on both input and output, as our GPT vs Claude comparison sets out tier by tier.
DeepSeek's Increase Is Live, and So Is Peak-Hour Pricing
In August we reported that DeepSeek had warned of a significant rise with no number and no date, and that surge pricing was rumoured. Both are now real.
On 16 August, DeepSeek raised prices and introduced peak and off-peak billing. Its pricing documentation states that off-peak rates are half of peak rates. Then on 10 September, V4.1 Flash replaced V4 Flash under the deepseek-flash name and cut the Flash rate by about a third.
Where that leaves the two models, per million tokens:
- V4.1 Flash: $0.30/$1.20 at peak, $0.15/$0.60 off-peak. Cache hits cost $0.006 and $0.003.
- V4 Pro: $1.32/$3.96 at peak, $0.66/$1.98 off-peak. Cache hits cost $0.044 and $0.022.
Against the $0.14/$0.28 that Flash charged in August, the peak rate is more than double and the off-peak rate is close to flat on input. Pro is the harsher move: it was $0.435/$0.87 and now costs three times that at peak. V4 Pro was also scheduled for retirement on 14 September, which DeepSeek reversed, so it stays available at unchanged billing. Our DeepSeek pricing page lists the peak rate, because a table that quotes only the cheaper number understates what you might pay.
DeepSeek Peak Hours in US Time Zones
Peak runs 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday. Every other hour, including all weekend, is off-peak. That is 35 hours out of 168, or about a fifth of the week.
Converted while US daylight saving time is in effect, until 1 November, peak lands in the American night:
- Eastern: 21:00 to midnight Sunday through Thursday, and 02:00 to 06:00 Monday through Friday.
- Pacific: 18:00 to 21:00 Sunday through Thursday, and 23:00 to 03:00.
A US team working normal hours is off-peak all day and pays half rate without doing anything. A team in Asia, or one running overnight batch jobs on a US schedule, is the one exposed. Check when your scheduled jobs actually fire before assuming either.
Google Gemini 3.8 Flash Arrives at Half Price
Google shipped Gemini 3.8 Flash on 2 September at $0.75/$3.75, with cached input at $0.075. It describes it as its most intelligent Flash model, with a 1,048,576-token context window.
The catch is the calendar. That rate is introductory through 31 December 2026, and standard pricing of $1.50/$7.50 takes effect on 1 January 2027. The same applies to Gemini 3.7 Flash, which went GA on 13 August, and to Gemini 3.6 Flash, which was quietly halved to the same introductory rate in mid-August. Three Flash models now share one price and one expiry date.
Google publishes no changelog entry for the 3.6 Flash cut. We dated it from archived copies of the pricing page, which show $1.50 on 12 August and $0.75 the next day. The older Gemini 3.5 Flash still bills $1.50/$9.00. While the promotion lasts, that makes it twice the price of three newer and better models. Rates are on our Google Gemini pricing page.
Mistral, xAI, Meta and Perplexity
xAI launched Grok 4.6 on 12 August as its flagship, at Grok 4.5's $2.00/$6.00 with the same 500,000-token context. Cached input is the difference, at $0.50 against 4.5's $0.30. Neither model gets a batch discount, and both double to $4.00/$12.00 above 200,000 prompt tokens. From 21 September, X Search bills $5 per 1,000 posts and $10 per 1,000 profiles fetched, replacing $5 per 1,000 calls.
Mistral changed no prices. We did correct two of our own entries: Devstral 2 and Magistral Small both retired on 31 July 2026, and our table carried them until this sweep. Codestral's context window is 128K, not the 256K we listed. Regional inference carries a 10% uplift.
Meta released Muse Spark 1.3 on 2 September at an unchanged $1.25/$4.25, following 1.2 on 5 August, which widened access beyond US developers. Hosted Llama rates did not move.
Perplexity changed no prices, but set a deadline: Sonar is supported only until 27 September 2026, as Sonar Chat Completions becomes the Agent API. If you call a Sonar model in production, that is a two-week migration, not a pricing question.
The September 2026 Rate Card at a Glance
| Model | Input | Output | Since August |
|---|---|---|---|
| GPT-5.5 Pro | $30.00 | $180.00 | Unchanged |
| GPT-6 Astra | $10.00 | $50.00 | New |
| Claude Fable 5.1 | $10.00 | $50.00 | New, cache reads cut |
| Claude Opus 5 | $5.00 | $25.00 | Unchanged |
| GPT-5.6 Sol | $4.00 | $20.00 | Promotional, was $5.00 / $30.00 |
| GPT-5.6 Terra | $2.00 | $12.00 | Unchanged |
| Claude Sonnet 5 | $2.00 | $10.00 | Unchanged, increase cancelled |
| Grok 4.6 (to 200K) | $2.00 | $6.00 | New |
| DeepSeek V4 Pro (peak) | $1.32 | $3.96 | Up from $0.435 / $0.87 |
| Muse Spark 1.3 | $1.25 | $4.25 | Unchanged |
| Claude Haiku 4.5 | $1.00 | $5.00 | Unchanged |
| Gemini 3.8 Flash | $0.75 | $3.75 | New, introductory |
| Mistral Large 3 | $0.50 | $1.50 | Unchanged |
| DeepSeek V4.1 Flash (peak) | $0.30 | $1.20 | Up from $0.14 / $0.28 |
| GPT-5.6 Luna | $0.20 | $1.20 | Unchanged |
| Mistral Small 4 | $0.15 | $0.60 | Unchanged |
| Gemini 2.5 Flash-Lite | $0.10 | $0.40 | Unchanged |
| Llama 3.1 8B (hosted) | $0.02 | $0.04 | Unchanged |
The budget ranking reshuffled this month. DeepSeek V4.1 Flash fell out of the top five on its peak rate, and Mistral Small 4 moved up. Compare all of it side by side on our AI API comparison table.
Do Bedrock, Azure and Vertex AI Charge the Same?
Mostly, with exceptions worth checking. OpenAI states that Bedrock pricing in commercial regions matches its direct pricing, and Microsoft listed GPT-6 Astra on Foundry Global Standard at the same $10/$50 on launch day. Two caveats: Azure's US Data Zone adds 10%, at $11/$55, and Azure publishes no batch pricing for Astra, so the 50% batch discount available directly from OpenAI is not there.
OpenAI also charges a 10% uplift on regional processing endpoints for models released on or after 5 March 2026, and Mistral applies the same 10% for regional inference. Data residency is a line item, not a free preference.
What September Does to the Same Monthly Workload
We used a standard shape in August: 100,000 requests a month, averaging 3,000 input and 500 output tokens. That is 300 million input tokens and 50 million output tokens. Running the same job through this month's rates:
- GPT-6 Astra or Claude Fable 5.1: $5,500
- Claude Opus 5: $2,750
- GPT-5.6 Sol: $2,200, against $3,000 at its list price
- GPT-5.6 Terra: $1,200
- Claude Sonnet 5: $1,100
- Grok 4.6: $900
- Claude Haiku 4.5: $550
- Gemini 3.8 Flash: $412, rising to $825 on 1 January
- GPT-5.6 Luna: $120
- DeepSeek V4.1 Flash: $150 at peak, $75 off-peak, about $91 if traffic is spread evenly across the week
Two things stand out. That DeepSeek job cost $56 in August, so the cheap lane got roughly 60% more expensive even after the September cut, and it now varies by clock. And at the top, the choice between Astra and Fable 5.1 is not about the rate card at all: the bills are identical at $5,500. Add 80% cache hits and Fable 5.1 lands near $3,160 against Astra's $3,340, before cache writes.
Model your own numbers in the pricing calculator, or work per head with the cost per seat tool.
August's Watch List: What We Expected and What Happened
Last month we said to watch three things.
Would DeepSeek attach a number and a date to its increase? Yes, within a week. The rise took effect on 16 August, and the peak windows reported in August turned out to be accurate, with the detail everyone missed being that weekends are off-peak entirely. Time-of-day pricing has not spread to any other provider we track.
Would Anthropic's permanent promotion set a pattern? Half right. Promotional pricing spread quickly, with OpenAI's Sol cut and Google's three Flash models. But both come with end dates attached, which is the opposite of what Anthropic did. Sonnet 5 remains the only promotion made permanent.
Would the mid-tier cluster hold? It held on list price. Terra, Gemini 3.1 Pro and Sonnet 5 all sit where they did. The pressure came from outside the cluster: Sol's promotion dropped into the range from above, and Gemini's Flash models undercut it from below.
Fun Fact: The 40x Discount Applies to Last Year's Model
Researchers at Epoch AI measured how fast the price of a fixed level of capability falls. The price to reach GPT-4's performance on a set of PhD-level science questions fell by about 40x per year. Across performance milestones, the decline ranged from 9x to 900x per year, with a median of 50x. They add a caveat: the fastest drops happened in the most recent year, so it is less clear those will persist. September is that caveat in miniature. Yesterday's capability keeps getting cheaper, while the new top tier launched at $10/$50 and the cheapest major provider raised its prices.
What to Watch in October 2026
Four dates. Perplexity ends Sonar support on 27 September. OpenAI shuts down the o4-mini API on 23 October, with Azure following on 19 November. GPT-5.6 Sol's promotional rate has a floor of 21 November, and Google's three Flash models revert on 1 January 2027.
The broader question is whether promotional pricing becomes the norm. Four of the models in the table above bill at a rate with an expiry date. A rate card where the advertised number is temporary is harder to budget against than one that simply falls. It also puts the work back on the buyer, who has to track when each clock runs out. Our price changes page shows when a tracked model moves.
Frequently Asked Questions
Did Any AI API Prices Go Down in September 2026?
Yes. DeepSeek cut V4.1 Flash by about a third on 10 September, though it remains well above its August rate. GPT-5.6 Sol's promotional $4/$20 started on 21 August and still applies. Google's newest Flash models bill at half their 2027 list price until 31 December. Claude Fable 5.1 cut cache reads by 75% without changing its headline rate.
How Does GPT-6 Astra Pricing Compare With Claude Fable 5.1?
They are identical on the rate card: $10 input and $50 output per million tokens. The differences are in the mechanics. Fable 5.1 bills cache reads at $0.25 against Astra's $1.00, and Anthropic charges standard rates across its full 1M-token window, where Astra doubles to $20/$75 above 272,000 tokens. Astra offers a 50% batch discount, and it is not available on Azure.
Is DeepSeek Still the Cheapest Major AI API?
No. At its peak rate of $0.30/$1.20, V4.1 Flash is more expensive than GPT-5.6 Luna at $0.20/$1.20 and Gemini 2.5 Flash-Lite at $0.10/$0.40. Off-peak, at $0.15/$0.60, it is competitive again. Whether DeepSeek is cheap now depends on when your jobs run.
When Does Gemini 3.8 Flash Introductory Pricing End?
On 31 December 2026. Standard pricing of $1.50/$7.50 takes effect on 1 January 2027, which doubles the rate. The same date applies to Gemini 3.7 Flash and Gemini 3.6 Flash.
Compare current rates for all 10 providers side by side, or see how each model got here in the pricing history explorer. Our cost optimisation FAQ covers the levers that cut a bill without changing model.
