AI Trends (November 2026)

AI Trends (November 2026): price hikes, open weights, and user-billed agents

🔎 The month when prices stopped falling

November 2026 marks the end of quiet pricing. In early November, no fewer than 32 API price changes were recorded within just a few days — 23 increases versus 9 decreases — and 10 new models landed in the catalog simultaneously. The era of ever-falling costs has come to a close, giving way to fragmented, tiered pricing.

At the same time, 13 models were released by 13 different organizations, 8 of them with open weights. AI agents, for their part, reached general availability at Salesforce, Adobe, and Snowflake — with usage multiplied 21-fold in three months for Agentforce Coworker alone.

Finally, a regulatory deadline is approaching: on December 2, 2026, machine-readable labeling of synthetic content becomes mandatory for systems already on the market. This monthly edition separates the noise from the signals. Here's what really matters this month.

Key Takeaways

• API prices are rising: 32 changes in early November (23 increases, 9 decreases), and promotional pricing expires at the end of November — GPT-5.6 Sol goes from $4/$20 to $5/$30.
• Open weights become the majority: 8 of the 13 models released in November are open weights (GPT OSS 120B, ERNIE-5.5-Flash-Base, Muse Spark, lightweight FLUX 3).
• Agents move into paid production: Agentforce Coworker at $50/user/month with 21x usage; Adobe and Snowflake roll out their coworkers.
• Routing is industrializing: Copilot Fusion automatically routes between models — Fusion Max rivals Claude Code at roughly half the cost per task.
• Images: the price war rages on: ChatGPT Images 2.5 dominates (Elo 1222) but Nano Banana 2.1 is only 46 points behind, at half the price.
• AI Act: mark your calendars for December 2, 2026 — mandatory machine-readable labeling, a new ban, and penalties of up to €15M or 3% of global revenue.

Recommended Tools

The median of the cheapest paid plan stands at $15/month across 772 tools analyzed by AurionX (November 2026). The market has clearly segmented — here is our current selection.

💰 API Pricing: Across-the-board hikes, surgical cuts

Yes, prices are going up: 23 increases versus 9 decreases were recorded in early November, along with 10 new models added to the catalog, according to AI Calculator Pro. But reducing November to a blanket price hike would mean missing the real story: pricing is splintering into tiers.

The examples come one after another, and they all tell the same story. The GPT-5.6 Sol promo ($4/$20 per million tokens) ends in late November, moving up to $5/$30. Gemini 4 Argon enjoys an introductory price of $2/$10 through November 30, before moving to $4/$20, according to LLM Data Hub. Translation: teaser rates are there to hook your workloads, then the bill catches up.

On the other side, a few well-targeted cuts. GPT-6.1 Terra, dedicated to computer use, costs $1/$5 per million tokens with a 1.05M context — 50% cheaper than GPT-5.6 Terra. Z.ai's GLM 5.3 FlashX Legal drops to $0.37/1M. And the open-weight GPT OSS 20B and 120B start between $0.004 and $0.015/1M.

The high end, meanwhile, floats: GPT-6 Cyber, reserved for organizations in the Daybreak Red program, charges $20/$100 per million tokens, according to DataStudios. Claude Opus 5.5 lists at $4/$20 standard, $8/$40 fast mode, and $2/$10 batch — with Opus 4.8's fast mode disappearing on November 30.

Sources: AI Calculator Pro, DataStudios, LLM Data Hub and Singularity.Kiwi — November 2026 prices; check the vendors' websites.

Deprecations to put on your calendar

MiniMax is retiring M2.5 and M2.7 on November 18 (traffic routed to M3.1 Flash). Claude Sonnet 4.5 leaves the Anthropic API on November 30 in favor of claude-sonnet-5-5. Mistral is deprecating Medium 3.5 and Large 2512, with removal scheduled for April 30, 2027. And on December 1, Gemini 2.5 Pro and Flash requests will automatically route to gemini-3.8-flash and gemini-3.5-flash-lite — billed at the new models' rates, with breaking changes to expect, according to Singularity.Kiwi.

One last cost line to watch: DeepSeek now charges $0.75 per 1,000 requests for its Search and Fetch APIs, according to TokenBill. If your agents do a lot of searching, it shows up on the bill fast.

This pricing pressure didn't start in October: we were already tracking it in AI Trends (June 2026). The difference is that it's now structured in tiers — and your budget needs to account for it. My advice: move your workloads to batch whenever possible ($2/$10 on Opus 5.5, i.e., half the price) and don't build any forecasts on a promo rate.

📦 Open weights: the majority of the market, for the first time

Eight of the thirteen models released in November 2026 are open weights, according to AIFOD Realtime. The question is no longer whether to pay attention to them, but which one to deploy.

The movement is driven by the big players. OpenAI releases GPT OSS 20B and GPT OSS 120B, available from $0.004 to $0.015/1M depending on the provider. Baidu opens up ERNIE-5.5-Flash-Base. Moonshot publishes the weights of Kimi K3.1 Pro as early as November 2, Tencent finalizes Hy4 on the 4th, and LongCat-2.5, Sarvam Prism, Ant Ling-3.2-flash and Kakao Kanana-2 round out the wave. Google even opens up two Gemma models: gemma-4-31b-it and gemma-4-26b-a4b-it.

On the Western side, three releases particularly catch my attention, according to Fellow.news:

• Meta Muse Spark: an open multimodal MoE with 1M tokens of context — the closed-model argument, in open form.
• LFM2.5-24B-A2B (Liquid AI): 24B parameters with 2B active, designed for low-cost on-device agents.
• MiMo-V2.6-Pro (Xiaomi): omnimodal open-weights at $0.43/$0.87 — a price-to-scope ratio that's hard to beat.

On the image front, FLUX 3 from Black Forest Labs offers a lightweight self-hostable version, according to Digitiz. Enough to run a generation pipeline without sending a single image to a third party.

Why it matters: sovereignty, predictable costs, no surprise deprecations. When a provider can pull a model overnight — see Claude Sonnet 4.5 on November 30 — hosting your own weights becomes an insurance policy.

On the infra side, a VPS from Hostinger is enough for compact models like LFM2.5 or GPT OSS 20B; the 120B ones require beefier GPU hardware. Our selection of free AI tools lists zero-cost alternatives to get started.

This shift isn't new: we were already documenting it in AI Trends (May 2026). November gave it its critical mass — the majority of the market, across all regions.

🔀 Model Routing: The Skill That Replaces "What is the best LLM"

The best model of the moment is useless if you call it at the wrong time. In November 2026, routing transitioned from an insider hack to a product feature, both at GitHub and Google.

Take Copilot Fusion, unveiled at GitHub Universe on October 28th: its automatic multi-model routing adapts to the requested effort level (Standard, High, Max). The result: Fusion Max competes with Claude Code by leveraging Opus 5.5 Max for about half the cost per task, with a 30% discount on model costs, according to CMultimedia. To choose your stack, our comparison of AI tools for code details these trade-offs.

Google is pushing the logic further—to the breaking change. Starting December 1st, Gemini 2.5 Pro and Flash requests will be routed to gemini-3.8-flash and gemini-3.5-flash-lite. The same logic applies to multimodal: Omni 1.1 Flash routes to 1.2 starting December 3rd, with a 30% discount on video ($12.00/1M, or about $0.07/second in 720p), according to TokenBill.

Analysts agree. IBM summarizes the shift: AI leadership is now defined by systems and cooperative model routing, not by models alone—and agentic parsing (Unstructured + Docling pipeline) is replacing monolithic document processing, according to IBM Think. Kanerika ranks routing as a top priority, alongside tiered pricing models and the deadlines of the AI Act (Kanerika).

Microsoft is taking it further, making it one of its 7 trends for 2026: AI is becoming a true collaboration partner, which requires efficient infrastructure—and thus intelligent routing (Microsoft Source).

My opinion is clear: measure the cost per task completed, not the price per token. That is the only metric that matters when your router can choose between a Terra at $1/5 and an Opus 5.5 at $4/20.

🤖 AI Agents: from demo to monthly quota

November 2026 marks the turning point: AI agents are no longer sold as pilot projects, but as per-user licenses. Three major vendors went generally available within 48 hours.

Salesforce kicks things off on November 5: Agentforce Coworker reports usage up 21x in three months, priced at $50/user/month (November 2026, check salesforce.com). The GA brings Microsoft Teams integration, an iOS app, Mac and Windows desktop apps, an MCP server, and connectivity to Google Workspace and Microsoft 365 (Salesforce).

Adobe follows on November 4 with CX Enterprise Coworker, available directly inside ChatGPT, Claude, Databricks Genie, Gemini, and Microsoft Copilot — with AWS, IBM, and Snowflake to come. It all runs on MCP and the Agent2Agent (A2A) protocol, according to Adobe's press release. Translation: your agents no longer live in a vendor silo, but right where your teams already work.

Snowflake, presented at its Expedition 2026 on November 3, is rolling out always-on bots that work in the background: CoCo Automations, Async Agent API, and Snowflake Mobile. The adoption numbers set the tone — CoWork has topped 5,800 accounts, CoCo 9,100 (Snowflake).

Even business planning is getting in on the act. Abra, launched by Abacum on November 2, orchestrates forecasting workflows across the enterprise — and more than 85% of Abacum customers already use agentic capabilities every month (Abacum).

At the same time, the cost of agentic action is collapsing: GPT-6.1 Terra, built for computer use, bills at $1/$5 per million tokens. Multiplying an agent's actions finally costs less than thinking them through.

MCP and A2A: the quiet standardization that changes everything

MCP server at Salesforce, MCP + A2A at Adobe, Async Agent API at Snowflake: November cements open agentic protocols. That's what it takes for a coworker to stay portable from one platform to the next — and the reason I'd be wary of any agent locked into a proprietary ecosystem.

Governance remains the weak link. According to the Okta Global CISO Insights 2026 report covered by The AI News, 81% of CISOs worry about the excessive access granted to AI, and only 47% can identify every agent present in their environment. The European Commission, through its AI Act Service Desk, reminds us that agents do not form a distinct legal category: since August 2, 2026, an agent classified as high-risk must meet the requirements of both the AI system and the underlying GPAI model.

🖼️ Image generation: the leader costs twice as much as its runner-up

ChatGPT Images 2.5 remains number one in the Elo rankings (1222), but Google's Nano Banana 2.1 is only 46 points behind — for half the price. That's THE key takeaway from Digitiz's rankings for November 2026.

The market has settled into three tiers:

Per-image prices now range from $0.009 to $0.211 depending on the model and provider. At the scale of a content production operation, that gap translates into thousands of euros per month.

On the API side, OpenAI breaks down GPT-Image-2.6 into two variants, Sunburst and Flare: $5/1M input text tokens, $8/1M image input, and $30/1M image output, according to DataStudios. For its part, xAI is replacing grok-imagine-image-quality with grok-imagine-image-2.0, $0.01 cheaper per image (TokenBill).

Two competing strategies stand out: FLUX 3 is betting on self-hosting with its open-weights lite version, while Adobe Firefly is playing the legal protection card — an argument that carries more and more weight for exposed corporations.

AurionX's price index confirms this positioning: image editing is the cheapest category of AI tools, with a median of $9.18/month — versus $35 for customer service, the most expensive category.

My advice: for volume work (blog visuals, A/B variants), Nano Banana 2.1 or MAI-Image 2.6 are more than enough. Reserve ChatGPT Images 2.5 for cover visuals, where quality makes all the difference.

⚖️ EU AI Act: December 2, 2026 is not a postponement

No, the high-risk postponement exempts you from nothing. On December 2, 2026, two deadlines hit systems already placed on the market: machine-readable marking of synthetic content and a new prohibition.

A reminder of the timeline. Regulation (EU) 2026/1744, known as the AI Omnibus, in force since July 27, 2026, has postponed the application of Annex III high-risk rules to December 2, 2027, and that of regulated products to August 2, 2028. But Article 50 (transparency), the prohibitions, and enforcement powers remain unchanged, according to the analysis by Twarx.

Since August 2, 2026, the disclosure of deepfakes and AI-generated content without human review is already mandatory. On December 2, 2026, machine-readable marking of synthetic content applies to systems already placed on the market — and a new prohibition, targeting non-consensual sexual imagery and CSAM, comes into force the same day, according to CMultimedia.

The fine? Up to €15M or 3% of global turnover, with the DGCCRF in charge of enforcement in France. This is not a theoretical fine: the marking is technically verifiable, and therefore actionable.

For AI agents, the official AI Act Service Desk FAQ is clear: no separate legal category, the rules for AI systems and GPAI models apply, and Article 50 transparency obligations kick in as soon as an agent interacts with people.

The advice from The AI News is worth quoting as is: Annex III high-risk rules will only apply from December 2, 2027, but governance questions arise as soon as the agent can act. With 47% of CISOs unable to inventory their agents, waiting would be a management failure before it is a compliance failure.

We already covered these deadlines in AI Trends (September 2026) — what has changed since is the number of companies affected despite themselves, driven by the massive adoption of AI coworkers.

🧭 What this changes for your roadmap

Three priorities depending on your profile: locking in costs, testing open weights, preparing for regulatory labeling. Here's the concrete translation.

If you're a developer: add a router in front of your LLM calls and batch whatever can be batched. The automatic switchovers on December 1st (Gemini) and December 3rd (Omni) will happen without asking for your opinion — test your prompts on the replacement models now.

If you're a marketer: Nano Banana 2.1 for volume, ChatGPT Images 2.5 for key visuals. With a median of $9.18/month for image editing, cost is no longer the obstacle — production consistency is.

If you're an executive: take inventory of your agents (47% of CISOs can't identify them), prepare machine-readable labeling before December 2nd, and pilot a coworker with a pilot team — Agentforce Coworker serves as the pricing benchmark at $50/user/month.

If sovereignty concerns you: GPT OSS 120B, Muse Spark and ERNIE-5.5-Flash-Base now cover ground that was reserved for closed models six months ago. The cost of self-hosting has never been lower.

❌ Common Mistakes

Mistake 1: Building a budget on promo pricing

The GPT-5.6 Sol promo (4/20 $) ends at the end of November, as does the introductory price of Gemini 4 Argon (2/10 $). Solution: forecast at full price (5/30 $ and 4/20 $) and treat any decrease as a bonus, never as a baseline.

Mistake 2: Ignoring switch-over dates

MiniMax M2.5/M2.7 disappear on November 18, Claude Sonnet 4.5 on the 30th, and the Gemini 2.5 auto-routing on December 1st with breaking changes. Solution: test your prompts on the replacement models before the date, not after the production incident.

Mistake 3: Deploying agents without an inventory

47% of CISOs cannot identify all the agents in their environment, and 81% worry about excessive access. Solution: centralized inventory, least privilege principle, and action logs from the very first agent in production.

Mistake 4: Believing the AI Act has been postponed

The postponement only concerns Annex III high-risk systems (December 2, 2027). Article 50, the prohibitions, and the penalties (€15M or 3% of global revenue) apply right now. Solution: machine-readable marking and deepfake reporting integrated into your pipeline before December 2.

❓ Frequently Asked Questions

What are the most notable AI models of November 2026?

Gemini 4 Argon (1M context), GPT-6 Cyber (restricted access), GPT-6.1 Terra (computer-use at $1/$5) and Claude Opus 5.5 on the closed side. On the open-weights side: GPT OSS 120B, ERNIE-5.5-Flash-Base, Meta Muse Spark and Kimi K3.1 Pro. Thirteen models in total, eight of which with open weights.

Are AI API prices really increasing in November 2026?

Yes: 23 increases versus 9 decreases were recorded in early November, out of 32 changes in total. The nuance: the decreases target agentic use cases (GPT-6.1 Terra at $1/$5) and small models, while the high end and the expiration of promotional offers are pushing the bill upward.

Can you self-host a competitive AI model in 2026?

Yes, and the offering has never been broader: GPT OSS 120B, ERNIE-5.5-Flash-Base, Muse Spark (1M context), LFM2.5-24B-A2B for embedded agentic workloads and a lightweight FLUX 3 for images. Compact models run on a simple VPS; the 120B models require dedicated GPU hardware.

What does the AI Act change as of December 2, 2026?

Two obligations for systems already on the market: machine-readable labeling of synthetic content and the entry into force of a new prohibition (non-consensual sexual imagery, CSAM). Annex III high-risk, meanwhile, is postponed to December 2, 2027. Penalties: up to €15M or 3% of worldwide revenue.

How much does an AI agent cost for a team in November 2026?

Salesforce's Agentforce Coworker serves as a reference: $50/user/month (check on salesforce.com). For AI tools in general, the AurionX index puts the median of the cheapest paid plan at $15/month, with 33% of the 772 tools analyzed under $10/month.

✅ Conclusion

In November 2026, value is shifting from the model to the system: intelligent routing, governed agents, and well-mastered open weights are worth more than the top spot on a leaderboard. To make sure you don't miss any of December's releases, follow our AI news and our selection of the best AI tools, updated continuously.