Four dates in November 2026 will change what your AI stack costs, what it is allowed to do, or whether it works at all. None of them is a rumor. Each one comes from a vendor’s own pricing page, deprecation notice, or policy announcement — and if you use Gemini, GPT, or Claude through an API, at least one of them almost certainly touches something you run.
Here is the calendar, up front, so you can plan the month instead of discovering it:
| Date | What happens | Who it hits |
|---|---|---|
| November 12, 2026 | Anthropic’s new Usage Policy takes effect | Anyone building with Claude, especially in hiring, finance, elections, or hardware |
| November 21, 2026 | GPT-5.6 Sol’s promotional price is guaranteed “at least through” this date — and nothing is published after it | OpenAI API users on GPT-5.6 Sol |
| November 30, 2026 | Gemini 4 Argon’s introductory price ends; standard rates double | Gemini API users on Argon |
| November 30, 2026 | Claude Sonnet 4.5 retires on Anthropic’s platforms; requests to it fail | API users still calling claude-sonnet-4-5-20250929 |
Two deadlines share November 30, which makes the last week of the month the crunch point. The rest of this article gives you each deadline, the exact numbers, a worked cost example, and a week-by-week plan.
A quick note on verification: every price and date below was checked against vendor pages in this run. Where sources disagree — and they do, on one Sonnet 4.5 date — I say so plainly.
November 12: Anthropic’s new Usage Policy takes effect
Anthropic published its annual Usage Policy refresh on October 8, 2026, and the new version takes effect November 12. The company says most of the changes clarify existing rules rather than rewriting them, reflecting a year in which Claude took on longer, more independent work. Six changes matter for small businesses and developers.
Deceptive campaigns, consolidated
Previously scattered rules about fake accounts, fabricated news sites, and influence-operation infrastructure now sit in a single section: “Do Not Engage in Deceptive Campaigns or Artificial Activity.” It covers deceptive activity of any kind, political or commercial — obscuring who is behind a message, amplifying content through fake accounts or posts, and building the tools and infrastructure for running influence campaigns. If your product touches social posting, outreach automation, or content generation at scale, read this section first. The consolidation is not a loosening; it is a wider net with fewer gaps.
Elections, reframed around deception
The elections section is renamed “Do Not Undermine Democratic Processes” and narrowed to what it actually prohibits: using Claude to deceive voters or disrupt elections — spreading false information about candidates or how to vote, impersonating candidates or election officials, or suppressing turnout. The old blanket prohibition on personalized voter and campaign targeting is gone. That is a real change for legitimate civic work: nonprofits writing voter information in other languages, or election officials sending ballot-cure notices, are no longer caught by a blanket ban. Targeting that relies on deception or misuse of voters’ personal data is still prohibited under the deceptive-campaign, surveillance, and privacy sections. The line moved from “no targeting” to “no deception,” and that distinction is the whole story.
Weapons prohibitions now reach software
The weapons section now explicitly covers “the software and components that make weapons work,” including guidance and control software, arming drones and other autonomous vehicles, and targeting algorithms. Anthropic says it observed attempts to use its models for exactly this. If you work anywhere near defense-adjacent automation or drone software, this is no longer a gray area you can reason your way around — it will require explicit written authorization from Anthropic, not a permissive system prompt.
Surveillance and law enforcement, sharpened
Claude cannot be used to decide or recommend who to investigate, arrest, or charge, and it cannot be used to build or improve surveillance tools or to track people without their consent, whether in real time or through analysis of previously collected data. Uses that remain explicitly permitted include fraud monitoring, investigative journalism, court-ordered evidence review, and tracking people have agreed to. If your app touches hiring, tenant screening, or fraud review, the next item matters more.
High-risk use cases get a sharper requirement
Candidate screening, loan pricing, and claim decisions are now treated as high-risk. That means a qualified human reviewer with the authority to change or modify the model’s recommendation, plus clear disclosure to the person on the receiving end that AI was used to produce it. This is not a suggestion buried in an appendix — it is the requirement that will decide whether your hiring or lending workflow is compliant on November 12. If your pipeline auto-approves or auto-rejects without a human who can overrule it, you have eleven days from November 1 to add that human.
Hardware gets a stop-switch rule
Any autonomous physical system that could cause injury now requires a qualified human operator who can observe and stop it, and the equipment must hold a safe state if Claude disconnects mid-task. This is the change with the most engineering work behind it: a stop-switch and a fail-safe state become design requirements for Claude-driven agents running lab equipment, drones, or other machinery. Claude for Work and Enterprise customers get a separate compliance addendum covering the hardware and surveillance provisions before the effective date.
The clause everyone talked about — and its narrow scope
Yes, the new policy bans “sustained and needless abusive or cruel behavior” toward Claude itself — but not ordinary frustration, pushback, dark creative writing, or model testing. It drew the headlines; for most businesses it is the least operationally important change here. Do not let it distract you from the high-risk and hardware items above.
What to do this week: if you use Claude in hiring, lending, insurance claims, elections-adjacent work, surveillance-adjacent features, or physical hardware, read the full policy on Anthropic’s announcement page and diff it against your system prompts and product documentation. Most teams will find one prompt or one workflow that needs a human-review step added.
November 21: GPT-5.6 Sol’s promo floor — a guarantee with no ceiling
OpenAI’s GPT-5.6 Sol, the flagship model in the GPT-5.6 family, currently costs $4 per million input tokens and $20 per million output tokens on short context, with cached input at $0.40. Before the promotion, announced August 21, 2026, the list price was $5 input and $30 output — so the promo is a 20% cut on input and a 33% cut on output. The catch is in the wording on OpenAI’s own model page: the promotional pricing is available “at least through November 21, 2026.”
Read that carefully, because it is a floor, not a ceiling, and not an end date either. “At least through” means OpenAI guarantees the rate until November 21. It does not say what happens on November 22. OpenAI has published no post-promotional rate at all. The promo could extend. It could end exactly on that date, snapping back to $5/$30 or to some new rate nobody has seen. Any number you assume for November 22 is your own invention, and any budget built on $4/$20 past November 21 is built on a guess.
Three details that change the real bill:
- Long context costs more. Prompts over 272,000 input tokens are priced at 2x input and 1.5x output for the entire request — $8 input and $30 output on the promo rate. If your app occasionally sends very long prompts, a handful of them can dominate a month’s bill.
- The siblings are the escape hatch. GPT-6.1 Sol runs $2/$10, Terra is cheaper still, and Luna handles focused, high-volume work at a fraction of flagship prices. If a job does not need the flagship, it should not be paying the flagship rate — promo or not.
- Subscriptions did not move. ChatGPT Plus, Pro, and Business prices are unchanged. The promo affects API callers and token-billed Work and Codex usage, not chat subscribers.
What to do this week: if Sol is your default model, test one cheaper sibling on your most common task before November 21 and record cost per completed task, not cost per token. If the cheaper model passes your quality bar, route that task to it now — then November 22 becomes a non-event instead of a budget surprise. Verify the promo line on OpenAI’s GPT-5.6 Sol model page rather than trusting any secondary roundup, including this one.
November 30, part one: Gemini 4 Argon’s price doubles
Google released Gemini 4 Argon, its new flagship, on October 13, 2026, and attached an end date to its introductory API price from day one: $2 per million input tokens and $10 per million output tokens through November 30, then $4 and $20 from December 1. Every tier doubles at the same moment — Batch and Flex go from $1/$5 to $2/$10, Priority from $3.60/$18 to $7.20/$36, and implicit cache hits from $0.10 to $0.20 per million tokens, a 95% discount on cached input that survives the doubling.
Two facts frame the decision. First, there is no free tier for Argon — it is paid API only, so everyone using it is already paying. Second, output prices include thinking tokens, so the bill covers the model’s reasoning as well as the answer you see. A model that “thinks” for a long time before answering is not free thinking; it is metered thinking.
Here is the worked example the trend briefs promised, with real arithmetic. Take a workload of 10 million input tokens and 2 million output tokens in a month:
| Input cost | Output cost | Monthly total | |
|---|---|---|---|
| Argon through Nov 30 | 10 × $2 = $20 | 2 × $10 = $20 | $40 |
| Argon from Dec 1 | 10 × $4 = $40 | 2 × $20 = $40 | $80 |
| Argon Batch/Flex through Nov 30 | 10 × $1 = $10 | 2 × $5 = $10 | $20 |
| Argon Batch/Flex from Dec 1 | 10 × $2 = $20 | 2 × $10 = $20 | $40 |
The doubling is exactly that: a doubling, on both sides of the bill. The Batch and Flex escape hatch is real — half price on both sides of the change — but only if your workload tolerates asynchronous processing. If it cannot, budget on the December numbers now, not the November ones. And note the coincidence that matters for comparison shopping: from December 1, Argon at $4/$20 costs exactly what Claude Opus 5.5 costs and what GPT-5.6 Sol’s promo costs. The three flagships converge on the same headline rate, which means the decision stops being about price per token and starts being about which model does your specific job best.
What to do this week: run your representative workloads on Argon before December 1 and record cost per completed task. If the task tolerates Batch, move it there. If it does not, re-run your cost model on the December rates and decide before the bill surprises you — not after.
November 30, part two: Claude Sonnet 4.5 retires — and the date is messier than it looks
Anthropic deprecated Claude Sonnet 4.5 on September 30, 2026, giving 61 days’ notice — just over its own 60-day minimum for publicly released models. The recommended replacement is Claude Sonnet 5.5 (claude-sonnet-5-5). After the retirement date, requests to claude-sonnet-4-5-20250929 fail outright, with no grace period.
The price change is good news, with an asterisk
Sonnet 5.5 costs $2 input, $10 output, and $0.10 cache read per million tokens, against $3, $15, and $0.30 for Sonnet 4.5. On the same 10-million-input, 2-million-output workload: Sonnet 4.5 costs $30 + $30 = $60, while Sonnet 5.5 costs $20 + $20 = $40. That is a genuine saving — but smaller than the headline suggests, because Sonnet 5.5’s newer tokenizer produces roughly 30% more tokens for the same text. If your 10 million input tokens become 13 million and your 2 million output tokens become 2.6 million, the real Sonnet 5.5 bill is $26 + $26 = $52, not $40. Still cheaper than $60, but not the two-thirds price the headline implies. Measure with your own prompts, not with the rate card alone.
The migration is not a one-line change
Swapping the model string is the easy part. The breaking changes are where migrations actually fail:
| What Sonnet 4.5 accepted | What Sonnet 5.5 does |
|---|---|
Non-default temperature, top_p, top_k |
Returns a 400 error under API version 2026-10-21 |
| Thinking budgets | Returns a 400 error |
| Assistant prefill | Returns a 400 error |
| Forced tool choice | Returns a 400 error |
| Thinking disabled | Not available; lowest setting is thinking: between_tools |
| First response block assumed to be text | Thinking is on by default, so the first block may be a thinking block |
That last row is the silent failure: code that reads content[0] as text can break without throwing an error at all, because the first block is now a thinking block. Audit every place your code parses a response, not just the places that set parameters.
The date conflict you need to know about
Here is where I owe you honesty instead of a clean story. Anthropic’s documentation lists November 30, 2026 as the retirement date. But two public GitHub issues filed after the deprecation notice report that the customer email gives November 24, 2026 at 9 a.m. Pacific, recommends Sonnet 5 rather than Sonnet 5.5, and warns of reduced availability and rate errors starting October 30. Those reports are secondhand — nobody outside Anthropic has published the email itself — and nothing on Anthropic’s documentation pages confirms the earlier dates. The safe plan is to treat November 24 as your hard cutoff and ignore the November 30 date for planning purposes. If the docs are right, you finish a week early. If the email reports are right, you finish on time. Planning to November 30 and discovering the email was right is the only way to lose.
The platform caveat
The November 30 date applies only to platforms Anthropic operates: the Claude API, Claude Platform on AWS, Claude Models on Google Cloud (Anthropic-managed), and Microsoft Foundry. Amazon Bedrock sets its own schedule — reporting puts Bedrock’s Sonnet 4.5 end-of-life at April 8, 2027 — and Google-managed Cloud endpoints have not published a date at all. If you call Sonnet 4.5 through Bedrock or Google-managed endpoints, check that provider’s lifecycle page before assuming any date in this article applies to you.
The migration sequence that works
Anthropic’s own recommended sequence, which matches what migration guides converge on: first, open the Usage page in Claude Console and export the CSV broken down by API key and model — that finds every key still calling the retiring model without reading scripts by hand. Second, let Claude Code do the mechanical first pass with its bundled migration skill (/claude-api migrate this project to claude-sonnet-5-5), which swaps the model ID and the breaking settings and leaves a checklist of things to check by hand. Third, run your evaluation suite against Sonnet 5.5 on a staging branch, send a small share of production traffic to it, and finish the cutover by November 23 at the latest — keeping a week of buffer before even the earliest reported cutoff. If you only use Claude through the consumer chat app, none of this applies to you: this retirement is about the API model string, not the assistant you talk to.
For the broader price-war context these deadlines sit inside, see our earlier analysis of the October AI API price war for small businesses — and for the Claude pricing picture specifically, our Claude Haiku 5.5 price-cut breakdown.
The combined bill: one month, three models
Put the three price deadlines together for a small team running all three models at the same modest scale — 10 million input and 2 million output tokens per model, per month:
| Model | November bill | December bill | Change |
|---|---|---|---|
| Gemini 4 Argon (Standard) | $40 | $80 | +$40 |
| GPT-5.6 Sol (promo, short context) | $40 | Unknown — no post-promo rate published | Unbudgetable |
| Claude Sonnet 4.5 → 5.5 | $60 | ~$52 (with the 30% tokenizer effect) | −$8 |
| Total | $140 | $132 + unknown | Sol decides the total |
Argon’s doubling is predictable, Sonnet’s migration saves a little and costs a week of engineering, and Sol’s unknown post-promo rate is the only line that can genuinely surprise you — which is why testing a cheaper sibling before November 21 is the highest-value move in this article.
Your November action plan, week by week
Week of November 2: Export your Claude Console usage CSV and find every Sonnet 4.5 call. Read Anthropic’s usage-policy diff against your system prompts. List every workload running Argon or Sol, with its monthly token volume.
Week of November 9: Anthropic’s policy takes effect November 12 — have the human-review step in place for high-risk Claude workflows before then, not after. Start Sonnet 5.5 evals on staging. Test a cheaper OpenAI sibling on your most common Sol task.
Week of November 16: Sol’s promo floor arrives November 21. Decide your fallback model for each Sol workload now, while the promo still holds. Move batch-tolerant Argon work to Batch or Flex.
Week of November 23: Finish the Sonnet 4.5 cutover, keeping buffer before the November 24 reported date. Re-run your cost model on December rates for Argon and confirm your December budget.
November 30: Argon’s intro price ends and Sonnet 4.5 retires on Anthropic’s platforms. If you did the work above, this is a quiet day. If you did not, it is the day your app starts failing or your bill doubles. The difference is three weeks of ordinary, unglamorous preparation.
One last note: December does not let up — OpenAI’s schedule already lists GPT-5 models retiring December 11 and GPT-4.1 models on January 14, 2027. The teams that handle this churn well keep one place in the codebase where the model name lives, one usage export they actually read, and one cost-per-task number they trust more than any rate card.
Frequently Asked Questions
What are the AI price deadlines in November 2026?
Four: Anthropic’s new Usage Policy takes effect November 12; GPT-5.6 Sol’s promotional $4/$20 pricing is guaranteed at least through November 21, with no post-promo rate published; Gemini 4 Argon’s introductory $2/$10 pricing ends November 30, doubling to $4/$20 on December 1; and Claude Sonnet 4.5 retires November 30 on Anthropic’s platforms, with requests to it failing after that date.
Does GPT-5.6 Sol’s price go up on November 22?
Nobody outside OpenAI knows. The published wording is “at least through November 21, 2026” — a floor, not an end date. The promo could extend, or it could end and revert to the previous $5/$30 list price or a new rate. Budget November 22 onward as unknown, and test a cheaper sibling before the floor date.
Is Claude Sonnet 4.5 really retiring November 30?
Anthropic’s documentation lists November 30, 2026, but customer-email reports cited in public GitHub issues give November 24 at 9 a.m. Pacific and warn of reduced availability from October 30. Plan to November 24 to be safe. The date applies to Anthropic-operated platforms only; Amazon Bedrock reportedly runs Sonnet 4.5 until April 8, 2027, and Google-managed endpoints have published no date.
Is Sonnet 5.5 a drop-in replacement for Sonnet 4.5?
No. It costs less per token ($2/$10/$0.10 vs. $3/$15/$0.30), but its newer tokenizer produces roughly 30% more tokens for the same text, thinking is on by default, and non-default temperature, top_p, top_k, thinking budgets, assistant prefill, and forced tool choice all return 400 errors. Audit response parsing too — the first block may now be a thinking block, which can break code silently.
How much will Gemini 4 Argon cost in December 2026?
Standard pricing doubles to $4 per million input tokens and $20 per million output tokens from December 1, with cached input at $0.20. Batch and Flex are half the standard rate ($2/$10). There is no free tier for Argon, and output prices include thinking tokens.
What does Anthropic’s November 12 usage policy change for small businesses?
Most for most teams: little. The operational changes are in high-risk use cases (hiring, loan pricing, claim decisions now need a qualified human reviewer plus AI-use disclosure), hardware (a qualified operator and fail-safe state for physical systems), and the consolidated deceptive-campaigns rule. The model-abuse clause covers only extreme, repeated cruelty — not ordinary frustration, creative writing, or testing.