AI API pricing this week: GPT-6.2 Astra cuts 20%, Qwen4-Max at $1.50 and Gemini 4 Neon at $3/$15

Between November 13 and November 20, 2026, the official price lists of OpenAI, Alibaba Cloud, Google, DeepSeek and Amazon Bedrock changed in 21 verifiable places. The largest move came from OpenAI, which launched GPT-6.2 Astra on November 17 at $8 input and $40 output per 1M tokens, 20% below the $10 and $50 that GPT-6 Astra still lists, and cut the cached-input rate from $1.00 to $0.40 per 1M tokens on the same day.
Alibaba Cloud followed on November 16 with Qwen4-Max, priced at $1.50 input and $4.50 output per 1M tokens in its Singapore (International) region, against $2.00 and $6.00 for Qwen3.8-Max. Google added two paid-only models to the Gemini API: Gemini 4 Neon on November 18 at $3.00 input, $0.15 cached input and $15.00 output per 1M tokens, and Genie 4 on November 19, billed at about $0.10 per second of generated video with audio.
OpenAI also opened discounted two-hour and eight-hour background windows in the Responses API (30% and 40% below Standard rates), added GPT-Live-1.1 to its API with a $0.02-per-minute standby rate, and halved the Ultrafast premium on GPT-6.2 Sol until February 28, 2027. DeepSeek published prices for new search, fetch, OCR and voice APIs on November 13 and then stopped billing model-triggered search and fetch tool calls on November 18. Prices as listed on the official pricing pages and developer documentation on November 20, 2026, in USD per 1M tokens unless another unit is stated.
··········
ALL API PRICE AND TERMS CHANGES THIS WEEK.
The table lists every change verified on the providers' own pages on November 20. Where a model is new, the "Before" column names the predecessor it replaces or says that no earlier price existed.
........
Provider | Model | Item | Before | After | Effective date |
OpenAI | GPT-6.2 Astra (vs GPT-6 Astra) | Input | $10.00 | $8.00 | Nov 17 |
OpenAI | GPT-6.2 Astra (vs GPT-6 Astra) | Cached input | $1.00 | $0.40 | Nov 17 |
OpenAI | GPT-6.2 Astra (vs GPT-6 Astra) | 30-minute cache write | $12.50 | $10.00 | Nov 17 |
OpenAI | GPT-6.2 Astra (vs GPT-6 Astra) | Output | $50.00 | $40.00 | Nov 17 |
OpenAI | GPT-6.2 Astra (vs GPT-6 Astra) | Batch input / output | $5.00 / $25.00 | $4.00 / $20.00 | Nov 17 |
OpenAI | GPT-6.2 Sol (vs GPT-6.1 Sol) | Ultrafast multiplier | 6x Standard | 3x to Feb 28, 2027, then 6x | Nov 17 |
OpenAI | GPT-6.2 Astra, GPT-6.2 Sol, GPT-6.1 Terra, GPT-6 Luna | 2-hour background window | Standard rate | 30% below Standard | Nov 17 |
OpenAI | GPT-6.2 Astra, GPT-6.2 Sol, GPT-6.1 Terra, GPT-6 Luna | 8-hour background window | Standard rate | 40% below Standard | Nov 17 |
OpenAI | GPT-Live-1.1 | Active session | Not in API | $0.05 per minute | Nov 19 |
OpenAI | GPT-Live-1 and GPT-Live-1.1 | Standby (WebRTC, inference paused) | Idle billed at $0.05 per minute | $0.02 per minute | Nov 19 |
Alibaba | Qwen4-Max (vs Qwen3.8-Max), Singapore | Input / output | $2.00 / $6.00 | $1.50 / $4.50 | Nov 16 |
Alibaba | Qwen4-Max, Singapore | Cache hit | Discount, no rate listed | $0.075 | Nov 16 |
Alibaba | Qwen4-Max (vs Qwen3.8-Max), Global | Input / output | $1.650 / $4.951 | $1.413 / $4.24 | Nov 16 |
Gemini 4 Neon | Input / cached / output | New model | $3.00 / $0.15 / $15.00 | Nov 18 | |
Genie 4 (genie-4) | Video output with audio | Not in API | About $0.10 per second | Nov 19 | |
DeepSeek | Search API and Fetch API | Per 1,000 successful requests | Not offered | $0.75 | Nov 13 |
DeepSeek | File parsing and OCR API | Per 1,000 pages | Not offered | $0.225 | Nov 13 |
DeepSeek | Speech to text (ASR) | Per hour of audio | Not offered | $0.116 list, $0.058 limited-time | Nov 13 |
DeepSeek | Text to speech (TTS) | Per 10,000 characters | Not offered | $0.45 list, $0.225 limited-time | Nov 13 |
DeepSeek | Model-triggered web search and fetch | Tool-call fee | Billed per call | Not billed, results count as input tokens | Nov 18 |
Amazon Bedrock | GPT-6.2 Astra | Ultrafast mode | Not on Bedrock | Promotional 3x Standard to Feb 28, 2027 | Nov 17 |
........
GPT-6.2 Sol keeps the $2.00 input, $0.10 cached input and $10.00 output of GPT-6.1 Sol, so its only price change is the Ultrafast multiplier. Perplexity's Decider v1.3 models, released on November 12, are left out of the count because their $0.02 and $0.01 input rates are unchanged and the official pricing page still lists the v1.2 model names.
··········
GPT-6.2 ASTRA API PRICING: 20% LOWER TOKEN RATES AND A 60% CHEAPER CACHE.
OpenAI's pricing page lists GPT-6.2 Astra at $8.00 input, $0.40 cached input, $10.00 for a 30-minute cache write and $40.00 output per 1M tokens, with a 12-hour cache write at $20.00. GPT-6 Astra remains on the same page at $10.00, $1.00, $12.50 and $50.00, so the new model is 20% cheaper on fresh input, cache writes and output, and 60% cheaper on cache reads. Batch rates fall from $5.00 and $25.00 to $4.00 and $20.00. The full launch details, including the ChatGPT and Codex rollout, are in our article on the GPT-6.2 Astra and Sol launch, and the model-to-model differences are covered in GPT-6.2 Astra vs GPT-6 Astra.
The cache cut does more for agent and chatbot workloads than the headline 20%. A support bot that sends 10,000 requests a day, each with a 2,000-token cached system prompt, 1,000 fresh input tokens and 400 output tokens, costs about $9,600 a month on GPT-6 Astra and about $7,440 on GPT-6.2 Astra, a reduction of 22.5%. A coding agent session of 40 turns with 50,000 input tokens per turn, 90% of them served from cache, and 1,500 output tokens per turn falls from $6.80 to $4.72, about 31% less, because cached tokens dominate the bill. Cache-write charges are left out of both estimates.
GPT-6.2 Sol, the lower-cost model of the release, keeps GPT-6.1 Sol's token prices. Its Ultrafast mode costs 3 times the Standard rate through February 28, 2027 and 6 times from March 1, 2027 (Pacific time), against the 6x multiplier that GPT-6.1 Sol carries today. At the promotional rate, Ultrafast on GPT-6.2 Sol works out to $6.00 input and $30.00 output per 1M tokens. OpenAI's background guide also states that, starting with GPT-6.2 Astra and Sol, new models launch without Flex processing.
··········
OPENAI RESPONSES API BACKGROUND WINDOWS: 30% AND 40% OFF FOR SLOWER JOBS.
A November 17 changelog entry put two discounted completion windows for background requests in the Responses API into public beta. The `2h` window bills model tokens 30% below Standard and the `8h` window 40% below Standard, while the default `soonest` window carries no discount. The discount applies to GPT-6.2 Astra, GPT-6.2 Sol, GPT-6.1 Terra and GPT-6 Luna, requires Standard processing (`service_tier` set to `default`), and covers every eligible request, follow-ups included. Our report on the Responses API background windows explains the mechanics and the rate limits.
Applied to the list prices, GPT-6.2 Astra drops to $5.60 input and $28.00 output in the two-hour window and to $4.80 and $24.00 in the eight-hour window; GPT-6.2 Sol drops to $1.40 and $7.00, then $1.20 and $6.00. These are derived figures, since OpenAI publishes only the percentages. The Batch API at 50% off stays cheaper for work that can wait up to its own completion window, so the new tiers suit jobs that need results within a working morning.
··········
GPT-LIVE-1.1 IN THE OPENAI API: $0.05 PER ACTIVE MINUTE, $0.02 IN STANDBY.
OpenAI added GPT-Live-1.1 to the Live sessions endpoint on November 19, with improved pronunciation, accent consistency and multilingual transcription. Both GPT-Live-1 and GPT-Live-1.1 are billed per second at $0.05 per minute of active session, and a new standby mode for WebRTC sessions pauses inference during silence and bills it at $0.02 per minute. For a contact center running 1,000 calls a day, each eight minutes long with two minutes of hold or silence, monthly cost falls from about $12,000 to about $10,200 if the idle minutes move to standby. The details are in our article on GPT-Live-1.1 and standby mode.
··········
QWEN4-MAX API PRICING: ALIBABA CUTS ITS FLAGSHIP RATE BY 25%.
Alibaba Cloud's Model Studio pricing page, last updated November 20, lists Qwen4-Max at $1.50 input and $4.50 output per 1M tokens in the Singapore (International) region, with a single price tier up to 1M input tokens, a 1M-token free quota valid for 90 days, cache creation at $1.875 (5 minutes) or $3.00 (1 hour) and cache hits at $0.075. Qwen3.8-Max is still listed at $2.00 and $6.00 in Singapore and in the US (Virginia) region. In the Global deployment regions, Qwen4-Max costs $1.413 input and $4.24 output, with Batch File jobs at $0.707 and $2.12, against $1.650 and $4.951 for Qwen3.8-Max.
On an uncached summarization run of 1,000 documents of 100,000 tokens each, with 1,500 output tokens per document, the Singapore bill falls from about $209 on Qwen3.8-Max to about $157 on Qwen4-Max. The launch is covered in our article on Qwen4-Max pricing and the generational comparison in Qwen4-Max vs Qwen3.8-Max.
··········
[[ADPLUS_300x250_1]]
GEMINI 4 NEON AND GENIE 4: GOOGLE'S TWO NEW PAID-ONLY GEMINI API MODELS.
Google's developer page for Gemini 4 Neon lists $3.00 input, $0.15 cached input (implicit caching only) and $15.00 output per 1M tokens, flat across the 1M-token context window, on the paid tier of the Gemini API only. Batch, Flex and Priority prices were not published on November 20, and the main Gemini API pricing table did not yet show a Neon row. Gemini 4 Carbon, the model Neon sits above, is billed at $2.00 and $10.00 until November 30 and at $4.00 and $20.00 from December 1, 2026, which makes Neon 25% cheaper than Carbon's December rate. Our Gemini 4 Neon launch article and Gemini 4 Neon vs GPT-6.2 Sol give the detail.
Genie 4, Google's interactive world model, reached the Gemini API through the Interactions API on November 19. The Genie developer documentation states that it requires a paid-tier key and bills at 240 video output tokens per frame, which is 5,760 tokens per second at 24 frames per second or about $0.10 per second with stereo audio included. That works out to roughly $6 per minute of generated world, or about $60 for a ten-minute demo session. The documentation also warns that closing a browser tab does not stop server-side generation and that sessions must be ended explicitly to stop billing. See our article on Genie 4 in the Gemini API.
··········
DEEPSEEK API APPLICATION PRICES AND FREE MODEL-TRIGGERED SEARCH.
DeepSeek's November 13 changelog entry opened web search, URL reading, file input and model safety on `deepseek-flash`, together with standalone Search, Fetch, OCR, speech-to-text and text-to-speech APIs. The pricing page lists the Search API and the Fetch API at $0.75 per 1,000 successful requests, file parsing and OCR at $0.225 per 1,000 pages, speech to text at $0.116 per hour of audio and text to speech at $0.45 per 10,000 characters, with both voice services currently at a limited-time 50% discount ($0.058 and $0.225) and no end date given. Our DeepSeek API application pricing article covers the launch.
Five days later the terms changed again. From 12:00 Beijing time on November 18, server-side web search and fetch calls that the model triggers on its own are no longer billed as tool calls; their results are counted as model input tokens instead. The standalone Search and Fetch APIs keep the $0.75 rate. DeepSeek's token prices also vary by time of day: `deepseek-v4-pro` costs $1.32 input, $0.044 cache hit and $3.96 output per 1M tokens at peak hours (01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday) and half that off-peak, while `deepseek-flash` costs $0.30, $0.006 and $1.20 at peak.
··········
AMAZON BEDROCK ADDS GPT-6.2 ASTRA AND GPT-6.2 SOL WITH AN ULTRAFAST PROMOTION.
AWS announced on November 17 that GPT-6.2 Astra and GPT-6.2 Sol are generally available on Amazon Bedrock. The Bedrock pricing page carries a promotional note for GPT-6.2 Astra in Ultrafast mode at 3 times the Standard rate through February 28, 2027, matching OpenAI's own promotion, but it did not show Standard per-token rates for either model on November 20. Teams buying through AWS can therefore plan on the Ultrafast multiplier while checking the Bedrock console for the base rate in their region.
··········
WHAT 2-3 TYPICAL WORKLOADS NOW COST ACROSS THE MAIN MODELS.
The comparison uses Standard list prices verified on November 20 and three workloads. The support chatbot handles 10,000 requests a day for 30 days, each with 2,000 cached system-prompt tokens, 1,000 fresh input tokens and 400 output tokens. The coding agent session runs 40 turns of 50,000 input tokens, 90% cached, plus 1,500 output tokens per turn. The summarization job processes 1,000 documents of 100,000 uncached input tokens with 1,500 output tokens each. Cache reads are billed at each provider's cache-hit rate, cache-write charges are left out, and DeepSeek V4 Pro is priced at its peak rate.
........
Model | Input / cached / output | Support chatbot, per month | Coding agent, per session | Summarization, per 1,000 docs |
GPT-6.2 Astra | $8.00 / $0.40 / $40.00 | $7,440 | $4.72 | $860 |
Claude Opus 5.5 | $4.00 / $0.20 / $20.00 | $3,720 | $2.36 | $430 |
Gemini 4 Neon | $3.00 / $0.15 / $15.00 | $2,790 | $1.77 | $322 |
GPT-6.2 Sol | $2.00 / $0.10 / $10.00 | $1,860 | $1.18 | $215 |
Claude Sonnet 5.5 | $2.00 / $0.10 / $10.00 | $1,860 | $1.18 | $215 |
Grok 4.7 | $2.00 / $0.50 / $6.00 | $1,620 | $1.66 | $209 |
Gemini 4 Hydrogen | $1.50 / $0.075 / $7.50 | $1,395 | $0.89 | $161 |
Qwen4-Max (Singapore) | $1.50 / $0.075 / $4.50 | $1,035 | $0.70 | $157 |
DeepSeek V4 Pro (peak) | $1.32 / $0.044 / $3.96 | $898 | $0.58 | $138 |
........
The ranking shifts with the workload. Grok 4.7 beats Claude Sonnet 5.5 and GPT-6.2 Sol on the chatbot and the summarization job because its output rate is $6.00, yet it is the most expensive of the three on the coding agent, where its $0.50 cache-hit rate is five times theirs. Qwen4-Max and DeepSeek V4 Pro are the cheapest on all three workloads, and DeepSeek's off-peak rates cut its figures in half, to about $449, $0.29 and $69.
Discount tiers change the totals for patient workloads. Run through OpenAI's eight-hour background window, the summarization job costs about $516 on GPT-6.2 Astra and $129 on GPT-6.2 Sol, and through the Batch API about $430 and $108. Claude's Batch API also halves its rates, to $2.00 and $10.00 for Opus 5.5 and $1.00 and $5.00 for Sonnet 5.5. Google had not published Batch prices for Gemini 4 Neon on November 20, and xAI's models page lists no batch rate for Grok 4.7. Prompts above 200,000 tokens raise Grok 4.7 to $4.00 input and $12.00 output, so the summarization figures assume documents stay under that threshold.
··········
AI API PRICES IN THE WEEK OF NOVEMBER 16: WHAT BUYERS CAN COUNT ON.
The week's changes push frontier pricing down in two places. OpenAI's top model now costs $8 and $40 instead of $10 and $50, with a cache rate low enough to cut agent bills by about 30%, and Alibaba's flagship dropped by a quarter at the same time. Google's Gemini 4 Neon enters between Carbon and the cheaper Hydrogen at a price that undercuts Carbon's December rate, and DeepSeek removed a per-call fee from agentic search. The cost table shows that a team moving a cache-heavy agent from GPT-6.2 Astra to Claude Opus 5.5 or Gemini 4 Neon halves or more than halves its bill, while Qwen4-Max and DeepSeek V4 Pro cost a fraction of either for teams that accept their regions and peak-hour rules.
Several figures are still incomplete. Google has not yet put Gemini 4 Neon or Genie 4 into its main pricing table and has not published Neon's Batch rate, Amazon Bedrock lists no Standard rate for GPT-6.2, and DeepSeek gives no end date for its 50% voice discount. Two scheduled increases also land soon: Gemini 4 Carbon and Gemini 4 Argon double to $4.00 and $20.00 on December 1, and GPT-6.2 Sol's Ultrafast mode returns to 6x Standard on March 1, 2027.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org




