AI API Pricing October 2026: What Changed This Month

The month to October 2026 brought new models more than new prices. Claude Sonnet 5.5 and GPT-6.1 Sol launched a day apart at an identical $2/$10 per million tokens, and Grok 4.7 kept $2/$6. The biggest move was at the bottom: GPT-6 Luna costs $0.10/$0.50, a tenth of Claude Haiku 4.5.

Most of the other changes sit off the rate card. Perplexity ended support for its Sonar Chat Completions API on 27 September and moved Sonar to its Agent API, where web searches are billed separately. Google closed its Gemini 2.5 models to new developers without deprecating them. For its newest flagship, see our Gemini 4 Argon pricing guide.

DeepInfra stopped serving Llama 4 Maverick, and DeepSeek extended its half-price hours to Chinese public holidays. Two deadlines are now close: OpenAI shuts down o4-mini on 23 October, and GPT-5.6 Sol's promotional rate is only guaranteed to 21 November. Here is every change since our September 2026 pricing roundup, checked against the providers' own pages on 5 October 2026.

AI API Pricing in October 2026: The Short Version

• Claude Sonnet 5.5 launched on 28 September at $2/$10, the same price as the Sonnet 5 it replaces.
• GPT-6.1 Sol launched on 29 September at $2/$10, matching Sonnet 5.5 exactly, with cached input at $0.10 against Sonnet's $0.20.
• GPT-6 Luna launched on 22 September at $0.10/$0.50. It is now OpenAI's cheapest current-generation model and costs a tenth of Claude Haiku 4.5.
• Claude Opus 5.5 launched on 22 September at $4/$20, 20% below Opus 5 and level with GPT-5.6 Sol's promotional rate.
• Grok 4.7 replaced Grok 4.6 as xAI's flagship on 21 September at an unchanged $2/$6.
• Perplexity Sonar moved to the Agent API on 27 September, at $0.25/$2.50 per million tokens plus $2.50 per 1,000 web searches.
• GPT-5.4 Nano is deprecated, with its API shutting down on 1 April 2027, and the o4-mini API shuts down on 23 October.
• Google closed Gemini 2.5 to new developers on 18 September, without deprecating it or changing its prices.

Every AI API Price Change Since Our September Roundup

Our September roundup was verified on 15 September. This is everything that has moved since, in date order.

How We Verified These Prices

We read every rate in this post from the provider's own pricing page, model page, changelog or deprecation notice on 5 October 2026, not from a summary or a third-party tracker. Where a provider did not date a change, as with DeepSeek's holiday rule, we dated it from archived copies of the same page and say so.

As in previous months, we count four things as a price change: the list rate, the cache and batch multipliers, the long-context threshold, and time-of-day billing. Access restrictions and host retirements are not price changes, so they get their own section below. The month-by-month record behind every figure is in our pricing history explorer.

Claude Sonnet 5.5 and GPT-6.1 Sol: Same $2/$10, One Day Apart

The mid-tier is where most production traffic runs, and at the end of September it became a straight tie. Anthropic released Claude Sonnet 5.5 on 28 September. OpenAI released GPT-6.1 Sol the next day. Both cost $2 per million input tokens and $10 per million output tokens, and both halve to $1/$5 through their batch APIs.

Sonnet 5.5 Keeps Sonnet 5's Price

Anthropic priced Sonnet 5.5 exactly like Sonnet 5, according to its API pricing documentation. That means $2/$10, with cache reads at $0.20, five-minute cache writes at $2.50 and one-hour cache writes at $4. It keeps the 1M-token context window at standard rates, with 128K max output. It also uses Sonnet 5's tokenizer, so the same text produces the same token count. On cost, that makes it a like-for-like swap.

Update, 10 October 2026: Two days after this roundup, on 7 October, Anthropic cut Sonnet 5.5's cache reads to $0.10 per million tokens, matching GPT-6.1 Sol. It also launched Claude Haiku 5.5 at $0.10/$0.50. The cache comparisons below reflect prices on 5 October. See our October 2026 Claude releases roundup for the changes.

On code, it is not quite like-for-like. Anthropic lists five breaking changes for applications moving from Sonnet 5, including requests that disable thinking or force a specific tool, which now return an error. Sonnet 5 stays available at $2/$10, the rate Anthropic made permanent in August. Our Claude API pricing guide has the full Anthropic line-up.

GPT-6.1 Sol Replaces GPT-6 Sol After One Week

OpenAI took a less direct route to the same price. It launched GPT-6 Sol at $2/$10 on 22 September, alongside GPT-6 Luna, then released GPT-6.1 Sol on 29 September at the same list price. OpenAI describes GPT-6.1 Sol as near-Astra performance for complex work at a lower cost. GPT-6 Sol is still listed on OpenAI's pricing page, but the newer model has the better cache rate.

Both carry a 1.05M-token context window with 128K max output. As with every current GPT model, the rates rise for long prompts: above 272K input tokens, GPT-6.1 Sol bills $4/$15 for the whole request. Our OpenAI API pricing guide covers every GPT tier.

Where the Tie Breaks: Cached Input at $0.10 vs $0.20

The only line on the standard rate card where the two differ is cached input. GPT-6.1 Sol bills cache hits at $0.10 per million tokens, 5% of its input rate and half of the $0.20 that both GPT-6 Sol and Sonnet 5.5 charge. Cache writes cost the same on both, at $2.50.

On our standard workload with 80% of input served from cache, that difference is worth about $24 a month: $644 on GPT-6.1 Sol against $668 on Sonnet 5.5, before cache writes. Output tokens dominate the bill, so the cache gap matters less than it looks.

The bigger structural difference is long prompts. Sonnet 5.5 bills its full 1M window at standard rates, while GPT-6.1 Sol's input price doubles above 272K tokens. For retrieval-heavy jobs that routinely send very long contexts, that flips the comparison. Our GPT vs Claude pricing comparison breaks down every tier.

GPT-6 Luna Sets OpenAI's New Floor at $0.10/$0.50

GPT-6 Luna launched on 22 September at $0.10 per million input tokens and $0.50 per million output tokens, with cached input at $0.01. That makes it OpenAI's cheapest current-generation model, at half the input price of GPT-5.6 Luna ($0.20/$1.20) less than two months after that model's own 80% cut. It has the same 1.05M-token context window, 128K max output and 50% batch discount, which takes it to $0.05/$0.25.

GPT-6 Luna vs Claude Haiku 4.5: A Tenth of the Price

Claude Haiku 4.5 costs $1/$5, so GPT-6 Luna is exactly a tenth of its price on both input and output. Our standard workload costs $550 a month on Haiku 4.5 and $55 on GPT-6 Luna. Luna also carries a 1.05M-token window against Haiku's 200K. Haiku's case now rests on output quality for the tasks you run, not on price, and Anthropic has not released a new Haiku since 4.5.

GPT-6 Luna also matches Gemini 2.5 Flash-Lite's $0.10 input rate. With Gemini 2.5 now closed to new developers, Luna is the cheapest current-generation first-party model on our rate card that a new project can use. OpenAI's older GPT-5 nano is cheaper still at $0.05/$0.40, but its snapshot shuts down on 11 December. See the full ranking of the cheapest AI APIs.

GPT-5.4 Nano Is Deprecated, and Luna Is the Cheaper Replacement

OpenAI deprecated GPT-5.4 Nano on 1 October, along with GPT-5.1 and GPT-5.3-Codex, and will remove it from the API on 1 April 2027, according to its deprecations page. Its named replacement is GPT-6 Luna, which is cheaper on both counts. GPT-5.4 Nano costs $0.20/$1.25, so the standard workload drops from $122.50 to $55 a month, a 55% saving, with more than twice the context.

Not every forced migration saves money. OpenAI's named replacement for o4-mini, which shuts down on 23 October, is GPT-5.6 Terra. On the same workload, that move takes the bill from $550 to $1,200, while GPT-6 Luna would cost $55. Test more than one replacement before the deadline, and use our migration guides for the code changes.

Claude Opus 5.5 Meets GPT-5.6 Sol at $4/$20, but Only One Rate Is Permanent

Anthropic's Opus 5.5 arrived on 22 September at $4/$20, 20% below Opus 5's $5/$25, and we covered it in detail in our Claude Opus 5.5 guide. That price lands exactly on GPT-5.6 Sol's promotional $4/$20. The difference is how long each rate lasts.

Opus 5.5's $4/$20 is its list price. Sol's is a promotion that OpenAI says runs at least through 21 November 2026. That is a floor, not an end date, and after it Sol could return to its $5/$30 list price. Opus 5.5 also charges less for cache reads, at $0.20 per million tokens against Sol's $0.40.

Both now face pressure from below. GPT-6.1 Sol costs half as much at $2/$10, and OpenAI pitches it close to GPT-6 Astra on capability. Teams paying for Sol on its promotional rate should test GPT-6.1 Sol before 21 November rather than after.

Grok 4.7: A New xAI Flagship at Grok 4.6's $2/$6

xAI released Grok 4.7 on 21 September as its frontier model for coding, agentic tasks and knowledge work. The price did not move: $2 input, $6 output and $0.50 cached input per million tokens, exactly Grok 4.6's rates, rising to $4/$12 for requests that reach 200K prompt tokens. The context window stays at 500K tokens, and there is still no batch discount.

That gives xAI three flagships in three months at one price: Grok 4.5 in July, 4.6 in August and 4.7 in September. Only the cache rate changed along the way, from $0.30 on 4.5 to $0.50 on 4.6 and 4.7. xAI's lower prices sit on the older Grok 4.3 and Grok 4.20, at $1.25/$2.50 with a 1M window and a 20% batch discount. Full rates are on our xAI Grok pricing page.

Perplexity Sonar Moves to the Agent API

The deadline in our September post arrived on schedule. Perplexity ended support for Sonar Chat Completions on 27 September, as its Sonar migration guide confirms. Synchronous and streaming Sonar requests keep working, because Perplexity is converting them into Agent API requests model by model. Asynchronous Sonar requests are no longer supported.

What Sonar Costs Now: $0.25/$2.50 Plus $2.50 per 1,000 Searches

On the Agent API, Perplexity's own Sonar model costs $0.25 per million input tokens and $2.50 per million output tokens, with cached input at $0.0625. Search is no longer bundled into a per-request fee. Each web search call is billed as a tool, at $2.50 per 1,000 calls, or $1 per 1,000 with Fast Search. The old Sonar rate was $1/$1 plus $5 to $12 per 1,000 requests, depending on search context size.

On our standard workload with one search per request, Sonar now costs about $450 a month: $200 in tokens and $250 in searches. Under the old pricing, the same job cost $850 to $1,550. The search fee is now the larger part of the bill, so the number of searches per request matters more than the token rates.

Where Sonar Pro, Reasoning Pro and Deep Research Went

The other three Sonar models have no Agent API equivalents with their own rate cards. Perplexity suggests presets as starting points instead: sonar and sonar-pro map to the fast preset, sonar-reasoning-pro to low, and sonar-deep-research to high.

Presets are configurations, not models. Each one runs an underlying model from providers such as OpenAI and Anthropic, billed at that model's token rate plus tool fees, and the model behind a preset can change. We now track Sonar alone on our Perplexity pricing page, because the models that power the presets already appear on our rate card.

Access Changes That Move Your Bill Without a Price Change

Three changes this month did not touch a list price, but they can still change what you pay or which models you can use.

Gemini 2.5 Models Are Closed to New Developers, Not Deprecated

On 18 September, Google limited Gemini 2.5 Pro, 2.5 Flash and 2.5 Flash-Lite to developers who have used them before. Google says the models are not deprecated and will keep serving existing users until further notice, and its Gemini API deprecations page still shows no shutdown date for any of the three. New projects are steered to Gemini 3.5 Flash-Lite ($0.30/$2.50) or Gemini 3.8 Flash ($0.75/$3.75 on introductory pricing until 31 December).

The practical effect is at the budget end. Gemini 2.5 Flash-Lite, at $0.10/$0.40, has been the cheapest major first-party API on our rate card since it reached general availability. It still is for teams that already use it. Anyone starting fresh should plan around GPT-6 Luna or Gemini 3.5 Flash-Lite instead.

Vertex AI customers face a separate, earlier deadline: Google's Vertex AI lifecycle page lists 20 October 2026 as the retirement date for the 2.5 models there. Our Google Gemini pricing page has every current rate.

DeepSeek Off-Peak Now Covers Chinese Public Holidays in Full

DeepSeek's peak-hour pricing, which started on 16 August, gained an exemption in September. Peak still runs from 01:00 to 04:00 and from 06:00 to 10:00 UTC, Monday to Friday, at $0.30/$1.20 for V4.1 Flash and $1.32/$3.96 for V4 Pro.

DeepSeek's pricing page now excludes Chinese public holidays from peak, so those days bill at the half-price off-peak rate around the clock, just like weekends. DeepSeek did not announce the change. Archived copies of the page show it added the rule between 17 and 22 September.

For work that can wait, China's holiday calendar is now part of the scheduling decision. Our DeepSeek pricing page has the full peak and off-peak rates.

Llama 4 Maverick Leaves DeepInfra

Meta's Llama models are open-weight, so their prices depend on who hosts them. DeepInfra, which we use as the reference host for Llama rates, retired Llama 4 Maverick on 1 October. Novita still serves it with the full 1M-token context at $0.27/$0.85, up from DeepInfra's $0.20/$0.80, so Maverick's tracked price rises by about a third on input.

The other Llama rates we track are unchanged on DeepInfra: Llama 3.1 8B at $0.02/$0.04, Llama 4 Scout at $0.10/$0.30 and Llama 3.3 70B at $0.10/$0.32. Meta's own Muse Spark API stays at $1.25/$4.25 on its standard tier. See our Meta Llama pricing page for every hosted rate.

The October 2026 Rate Card at a Glance

All prices are USD per million tokens at standard rates, before batch discounts.

Compare every tracked model side by side on our AI API comparison table.

What October Does to the Same Monthly Workload

We use the same shape every month: 100,000 requests averaging 3,000 input and 500 output tokens, which is 300 million input tokens and 50 million output tokens. At October's rates, that job costs:

• GPT-6 Astra or Claude Fable 5.1: $5,500
• Claude Opus 5.5 or GPT-5.6 Sol: $2,200, with Sol rising to $3,000 if its promotion ends
• GPT-5.6 Terra: $1,200
• Claude Sonnet 5.5 or GPT-6.1 Sol: $1,100
• Grok 4.7: $900
• Claude Haiku 4.5: $550
• Perplexity Sonar: About $450 with one web search per request
• Gemini 3.8 Flash: $412, rising to $825 on 1 January
• DeepSeek V4.1 Flash: $150 at peak, $75 off-peak, and about $91 spread evenly across a normal week
• GPT-5.6 Luna: $120
• GPT-6 Luna: $55

The month's story is in the last line. In September, the cheapest current OpenAI model ran this job for $120. GPT-6 Luna does it for $55.

The mid-tier did not get cheaper on its existing list prices, but a buyer paying $2,200 for GPT-5.6 Sol's promotional rate can now run OpenAI's newest mid-tier model for half that. With 80% cache hits, GPT-6.1 Sol lands at about $644 and Sonnet 5.5 at about $668, before cache writes.

Model your own numbers in the AI pricing calculator, or read our cost optimisation FAQs for ways to cut the bill without switching models.

September's Watch List: What We Expected and What Happened

Last month we flagged four dates and one open question.

Perplexity's Sonar deadline on 27 September: It arrived on schedule. The replacement pricing is lower on input and higher on output, and search is now billed separately.

The o4-mini shutdown on 23 October: Still on track. OpenAI's deprecations page still lists 23 October, and Azure's own retirement follows on 19 November.

GPT-5.6 Sol's promotional floor on 21 November: Still pending. OpenAI's wording has not changed: the rate runs "at least through" 21 November, with no extension or end date announced.

Google's Flash pricing on 1 January: Unchanged. Gemini 3.6, 3.7 and 3.8 Flash stay at $0.75/$3.75 until 31 December.

Would promotional pricing become the norm? Not this month. None of the new models launched on a promotional rate. Instead, the labs competed with lower list prices: GPT-6 Luna at half GPT-5.6 Luna's input price, GPT-6.1 Sol at half GPT-5.6 Sol's promotional rate, and Opus 5.5 at 20% below Opus 5. That is easier to budget against, because a list price has no expiry date.

What to Watch in November 2026

Two OpenAI dates come first. The o4-mini API goes on 23 October, and GPT-5.6 Sol's promotional floor passes on 21 November. Our next roundup will show whether OpenAI extends that rate, ends it or leaves it running.

Azure retires o4-mini on 19 November. Claude Haiku 4.5's retirement commitment runs to "not sooner than" 15 October, so watch for a deprecation notice, though Anthropic gives at least 60 days' warning before retiring a model. Further out, OpenAI's GPT-5 nano snapshot shuts down on 11 December and Google's Flash introductory pricing ends on 31 December.

The broader question is whether the GPT-6 family keeps pulling prices down. Two of its three tiers launched at half the price, or less, of the GPT-5.6 models they sit beside. Our price changes page shows when a tracked model moves.

Frequently Asked Questions

Did Any AI API Prices Go Down in October 2026?

Not on any existing model from the major labs. The cuts came through new models: GPT-6 Luna at $0.10/$0.50, GPT-6.1 Sol at $2/$10 and Claude Opus 5.5 at $4/$20, each cheaper than the model it sits beside or replaces. Perplexity's Sonar input price fell from $1 to $0.25, but its output price rose to $2.50 and web searches are now billed separately. The only straight increase we track is Llama 4 Maverick, after its reference host retired it.

Is GPT-6.1 Sol Cheaper Than Claude Sonnet 5.5?

On list price they are identical, at $2/$10 per million tokens, and both cost $1/$5 through batch. GPT-6.1 Sol is cheaper on cached input, at $0.10 against $0.20. Claude Sonnet 5.5 is cheaper on prompts above 272K input tokens, where GPT-6.1 Sol's rates rise to $4/$15 and Sonnet 5.5's stay at $2/$10.

What Is the Cheapest OpenAI API Model in October 2026?

GPT-6 Luna, at $0.10/$0.50 per million tokens, or $0.05/$0.25 through the Batch API. The older GPT-5 nano is cheaper at $0.05/$0.40, but its snapshot shuts down on 11 December 2026.

How Much Does Perplexity Sonar Cost on the Agent API?

$0.25 per million input tokens and $2.50 per million output tokens, plus $2.50 per 1,000 web search calls, or $1 per 1,000 with Fast Search. There is no longer a per-request fee.

When Does the o4-mini API Shut Down?

OpenAI shuts down the o4-mini API on 23 October 2026, and Azure retires its deployments on 19 November 2026. OpenAI names GPT-5.6 Terra as the replacement, which costs more than o4-mini. GPT-6 Luna is far cheaper and worth testing as well.

Compare current rates for all 10 providers side by side, or see how each model got here in the pricing history explorer. Our September roundup has last month's changes.