Gemini 4 Carbon got 21 days at its launch price. Every LLM API price change in November 2026, and the eight that land between November 30 and December 1.
November's launches were mostly cheaper than what they replaced. The bill that matters is the one dated December 1.
Photo by Ondrej Polak on Unsplash
Google made Gemini 4 Carbon generally available on November 10. The same changelog entry said the $2 / $10 price runs through November 30. So Google's highest-scoring model on the Artificial Analysis index came with 21 days at its launch price, and on December 1 every line of its card doubles.
Carbon isn't alone. Between the evening of November 30 and the morning of December 1, Argon doubles too, two Gemini 2.5 IDs start answering with different models, Claude Sonnet 4.5 retires, and OpenAI's GPT-5.6 Sol promotion runs out with no price published for what comes next. OpenAI also shuts down Prompts, Evals and Agent Builder on November 30, and gpt-image-1-mini and gpt-image-1.5 on December 1. That's eight changes in about a day. We went through every rate card and deprecation page we track and put the month in one place.
The short version
- Carbon and Argon cost twice as much from December 1. After that, Claude Fable 5.2 on high is 17% cheaper per task than Carbon and scores higher.
- GPT-5.6 Sol's promotion ends November 30 and nobody has printed the next price. GPT-6.1 Sol is $2 / $10 today.
- Old Gemini 2.5 IDs keep answering after December 1, from different models. Old Sonnet 4.5 calls just fail.
- Most November launches cut prices. GPT-6 Cyber, at $20 / $100, went the other way.
The 24 hours that matter
Each row is the same workload: a month of 1 billion input tokens and 150 million output, no caching, prompts under 200K. The "if you do nothing" column is the part people miss. Google reroutes old IDs, so your code keeps working and the model quietly changes. Anthropic retires them, so the code stops.
| Model | Date | If you do nothing | November | After |
|---|---|---|---|---|
| Gemini 4 Carbon | Dec 1 | price doubles | $3,500 | $7,000 |
| Gemini 4 Argon | Dec 1 | price doubles | $3,500 | $7,000 |
| GPT-5.6 Sol promo | Nov 30 | price not published | $7,000 | $9,500 if $5 / $30 returns |
gemini-2.5-pro | Dec 1 | served by 3.8 Flash | $2,750 | $1,312.50, then $2,625 Jan 1 |
gemini-2.5-flash | Dec 1 | served by 3.5 Flash-Lite | $675 | $675 |
claude-sonnet-4-5-20250929 | Nov 30 | requests fail | $5,250 | $3,500 on Sonnet 5.5 |
Carbon and Argon: $2 / $10 to $4 / $20 per million tokens, from Google's pricing page. gemini-2.5-pro at $1.25 / $10 lands on 3.8 Flash at $0.75 / $3.75, which becomes $1.50 / $7.50 on January 1. gemini-2.5-flash and 3.5 Flash-Lite are both $0.30 / $2.50. Sonnet 4.5 is $3 / $15, Sonnet 5.5 $2 / $10. Reroutes from Google's deprecations page, the Sonnet date from Anthropic's.
The two cheaper rows aren't free wins. The 2.5 Pro reroute halves the bill for exactly one month, because 3.8 Flash leaves its own introductory price on January 1, and after that you pay about what you paid on 2.5 Pro for a different model. The 2.5 Flash reroute keeps the price and lands you on Flash-Lite. We went through both in our reroute post.
The Sonnet row is the only one that fails loudly. If claude-sonnet-4-5-20250929 is still hard-coded somewhere on November 30, those requests stop working. Sonnet 5.5 is a third cheaper per token, though Anthropic's newer tokenizer counts more tokens for the same text, so the saving in practice is smaller.
Carbon at $4 / $20 loses to Fable 5.2 per task
Artificial Analysis scores Carbon at 60 on high and charges it $5.48 a task. That figure uses the $2 / $10 card. Input, cached input and output all double on December 1, so the same run costs $10.96. Here it is next to the configurations people would actually weigh it against.
- GPT-6.1 Sol maxAA score 55$1.27
- Gemini 4 Hydrogen highAA score 51$1.87
- Claude Fable 5.2 mediumAA score 56$3.80
- Gemini 4 Carbon high, until Nov 30AA score 60$5.48
- Claude Fable 5.2 highAA score 62$9.08
- Gemini 4 Argon high, from Dec 1AA score 50$9.94
- Claude Opus 5.5 maxAA score 59$10.04
- Gemini 4 Carbon high, from Dec 1AA score 60$10.96
Cost per Intelligence Index task, v5.0.1, read November 14, 2026 (Carbon, Argon, leaderboard). Highlighted bars are AA's figure times two, our arithmetic, not an AA number.
From December 1, Claude Fable 5.2 on high scores about two points more than Carbon and costs 17% less per task, $9.08 against $10.96. Opus 5.5 on max sits one point under Carbon for $0.92 less. Carbon has a list price identical to Opus 5.5, and it writes enough tokens per task that the identical card doesn't make it the cheaper run.
Google built Carbon for security work first, and pitched it on security benchmarks. If that's why you picked it, the December price is the price of that. If you picked it because it topped the index for $5.48 before Fable 5.2 arrived on November 9, that reason is already gone, and the price advantage goes with the month. Argon is worse off: $9.94 a task for a 50, when Gemini 4 Hydrogen scores 51 for $1.87.
GPT-5.6 Sol now has an end date, and still no next price
When OpenAI cut GPT-5.6 Sol to $4 / $20 in August, the model page said the rate was available "at least through November 21". That's a floor, not an expiry, and we said so at the time. The page now reads "available until November 30, 2026". The August changelog entry still has the old wording.
So there's a date and no number after it. The last list price before the cut was $5 / $30. If that comes back, our 1B / 150M month goes from $7,000 to $9,500. OpenAI may also set something else, or extend again, and we'd rather tell you we don't know than guess.
The move we'd make is the same either way. GPT-6.1 Sol is $2 / $10 with no end date, half the promotional 5.6 rate, and it's the model OpenAI's own deprecations page now names for gpt-5 and o3 on December 11. The same page points gpt-5-mini at GPT-6.1 Terra ($1 / $5), a far cheaper target than the GPT-5.6 Terra it named when we wrote October's roundup.
What launched in November
| Date | What | Compared with | New |
|---|---|---|---|
| Nov 3 | Tencent Hy4, 50% off to Dec 31 | ~$0.83 / $2.50 | ~$0.42 / $1.25 |
| ~Nov 4 | Kimi K3.1 Pro weights; cheapest output host | Moonshot $2 / $10 | Decart $1.22 / $6.10 |
| Nov 5 | GPT-6 Cyber | GPT-5.6 Cyber $12.50 / $75 | $20 / $100 |
| Nov 9 | Claude Fable 5.2 and Mythos 5.2 | 5.1 at $10 / $50 | $8 / $40 |
| Nov 9 | Meta Muse Ember | Haiku 5.5 $0.10 / $0.50 | $0.10 / $0.50 |
| Nov | Gemini Omni 1.2 Flash video output | 1.1 Flash | $12, 30% lower |
| Nov 10 | Gemini 4 Carbon GA | Opus 5.5 $4 / $20 | $2 / $10 to Nov 30 |
| Nov 11 | Qwen4-Max Preview | pay-as-you-go | Token Plan only |
| Nov 12 | Gemini 4 Hydrogen | 3.8 Flash Jan rate | $1.50 / $7.50 |
| Nov 13 | DeepSeek speech to text | new | $0.116 / hour |
Input / output per million tokens unless marked. Read November 14, 2026. Hy4's yuan rates (¥6 / ¥18) are converted, and its sale date comes from Vercel's changelog, not Tencent. Kimi hosts from OpenRouter; sources put the weights release on November 4 or 5. Linked rows have their own post.
Read down the last column and the month looks generous. Anthropic cut its top card 20%. Google's new Flash replacement costs less per task than the model it replaces, even at twice the per-token rate. Meta matched Haiku's price at a smaller per-task bill. Decart sells Kimi K3.1 Pro output 39% under Moonshot.
DeepSeek's November 13 additions are worth a look if you run voice. Speech to text is $0.116 an hour, $0.058 during a promotion with no published end date; text to speech is $0.45 per 10,000 characters ($0.225 on promotion); OCR and file parsing are $0.225 per 1,000 pages. All three are on DeepSeek's pricing page, not in the changelog.
Two you can't buy, and one with no price
GPT-6 Cyber is the priciest model in the GPT-6 family, and you need Daybreak Red approval to call it: $20 in, $2 cached, $100 out, and $40 / $4 / $150 once a prompt passes 272K tokens. That is exactly twice GPT-6 Astra on input and output, and 60% more input than GPT-5.6 Cyber. Cache writes are $25 for the default 30 minutes or $50 for 12 hours.
Claude Mythos 5.2, Anthropic's verified-only security model, launched four days later at $8 / $40, the same card as Fable 5.2. GPT-6 Cyber costs 2.5 times as much on both lines. Neither publishes a shared benchmark with the other, so we can't say what the extra money buys.
Alibaba's Qwen4-Max Preview went the other way. It has no per-token price at all. The preview page says it's on the Token Plan subscription only, from November 11 until 24:00 Beijing time on November 25, and then it's retired. If you were searching for its API price, that's the answer: there isn't one yet.
The rest of the calendar
Nov 19
Vertex AI retires gemini-3.6-flash and points it at 3.8 Flash at the same price. On the Gemini API 3.6 Flash has no shutdown date.
Nov 25
Qwen4-Max Preview ends.
Dec 3
gemini-omni-1.1-flash is served as Omni 1.2 Flash, whose video output is 30% cheaper.
Dec 11
gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3 and o3-pro shut down.
Dec 18
Gemini 3.1 Flash TTS preview and the 2.5 Flash and 2.5 Pro preview TTS models shut down.
Dec 31
Tencent Hy4's half-price billing ends at 23:59:59 Beijing time.
Jan 1, 2027
Gemini 3.8 Flash and 3.6 Flash go from $0.75 / $3.75 to $1.50 / $7.50. Gemini Robotics ER 2 preview doubles from $1 / $5 to $2 / $10.
From Google, Vertex AI and OpenAI deprecation pages, read November 14, 2026.
January 1 still carries more money than December 1, because far more traffic runs on Flash than on Carbon or Argon. Gemini 4 Hydrogen already charges the January Flash rate with half the cached-input price, which is why we'd test it before the new year rather than after.
What we'd do this week
Search your code and configs for gemini-4-carbon, gemini-4-argon, gpt-5.6-sol, gemini-2.5-pro, gemini-2.5-flash and claude-sonnet-4-5. Every hit is a line whose price or model changes in two weeks.
For Carbon and Argon, run your own eval against Fable 5.2 and Hydrogen this month, while Carbon is still cheap enough that the comparison is cheap too. For GPT-5.6 Sol, move to GPT-6.1 Sol now rather than waiting to see what December's price is. Of everything on this list, that one costs you nothing to do early.
Plug your own token counts into the cost calculator, which already applies dated price changes, or compare every model above on the pricing table.
Sources
- Gemini API pricing and changelog - Carbon GA date, Carbon, Argon, Hydrogen, 3.8 Flash and Robotics ER 2 rates before and after
- Gemini API deprecations and Vertex AI model versions - 2.5 Pro, 2.5 Flash and Omni reroutes, TTS shutdowns, Vertex 3.6 Flash retirement
- OpenAI GPT-5.6 Sol, GPT-6 Cyber and deprecations - promotion end date, Cyber card, December shutdowns and named replacements
- Claude pricing and model deprecations - Fable 5.2 and Mythos 5.2 rates, Sonnet 4.5 retirement
- DeepSeek pricing, Qwen4-Max Preview and Kimi K3.1 Pro on OpenRouter
- Artificial Analysis leaderboard - index scores and cost per task