Guides Industry Report

Free LLM APIs in November 2026: What Changed, What's New, What to Watch

F
Free-LLM Editorial Team
• Published: November 4, 2026 • 12 min read

Two weeks after our last re-check, the free LLM API market moved again, and not in one direction. A gateway that was free in October now charges for every model, Ollama dropped its free cloud plan, and AI21 quietly switched off the Jamba API. At the same time we found more genuinely free providers than in any previous month, including one run by a Harvard lab. Here's everything that changed, checked against each provider's own docs, pricing page or public model list on November 4, 2026.

The short version

  • New free providers: FreeInference, AnyAPI, Zro, Inferbase, Gonka Broker and OpenTyphoon, plus new trials at Regolo, GonkaRouter, NovAI and Vectoree. Hyperbolic is back with a $1 credit.
  • No longer free: Onomeo (all API models paid), Nebius (bank card mandatory, no trial credit documented), Nscale ($5 top-up to start). Ollama Cloud is now pay-as-you-go, with only Beam still free.
  • Shut down: AI21 Labs sunset its Jamba API on August 9.
  • Best new free models: Step 5 Preview, Mistral Large 4 and Kimi K2.6 on NVIDIA's free endpoints, Kimi K2.7 Code, GLM-5.3 and DeepSeek V4 Pro on Cloudflare, Gemma 4 on the Gemini API, Recursive Flint on OpenRouter.
  • Dates to watch: Scaleway GLM-5.2 ends November 16, IteraCompute's free promotion ends November 24, Gemini 2.5 Flash/Pro redirect to 3.x on December 1, the free Dots 3 Note preview on OpenRouter ends December 31.

Part 1: New free providers worth knowing

FreeInference: free, no card, and it works with Claude Code

The most interesting newcomer comes from academia. FreeInference is run by the MadSys Lab at Harvard SEAS "for open-source, research and education". You sign up with an email, no card, and get GLM-5.3 Flash, DeepSeek V4 Flash and MiniMax M3 (all with 1M-token context) plus Qwen3.6 35B. It has both an OpenAI endpoint and an Anthropic one, so Claude Code connects with two environment variables. The catch: the quota is only described as "generous", and the lab studies agent traffic, so don't send anything private.

Daily and monthly allowances on gateways

  • AnyAPI: 100K tokens a day on its free models, no card.
  • Inferbase: 5,000 requests a month and $0.50 of model usage a day on any managed model under $2 per million output tokens (GPT-OSS 120B, DeepSeek V4 Flash, Kimi K2.7 Code…). No card, no time limit.
  • Zro (MoonMath): a daily allowance on GLM-5.3 Flash and DeepSeek V4.1 Flash, zero request retention, and a CLI that launches Claude Code for you.
  • Gonka Broker: a monthly free token grant once you verify a real mobile number. It runs on the decentralized Gonka network.

Regional and specialist

OpenTyphoon from SCB 10X is a free "research showcase" API for Thai and English (Typhoon 2.5, 200 requests per minute). If you build for Thai users, it's the best free option we know. Cloudflare also added SEA-LION v4 27B, a Southeast Asian language model, to its free catalog.

New trials (one-time credit)

  • Regolo.ai (Italy, EU hosting, zero retention): 30 days with 1M tokens a day, no card.
  • GonkaRouter: 1M free tokens for new users.
  • NovAI: $2 welcome credit without a card, and Zhipu's GLM Flash models free at 5 requests per minute.
  • Vectoree: $5 welcome credit on an agent backend.
  • Hyperbolic relaunched serverless inference at a new address (api.hyperbolic.ai/inference/v1). Verifying a phone number unlocks the API and adds $1. Our October report said its API was retired; that was true then.

Part 2: Who stopped being free

  • Onomeo listed free-chat, free-code and free-vision models in October. Its catalog now shows only paid models and lists those three as "removed this week". Earned rewards only work in its web chat.
  • Nebius Token Factory: its billing docs say onboarding can't be finished without a bank card and no longer mention the sign-up credit. A visitor told us the same thing in a comment on October 15.
  • Ollama Cloud replaced its free plan with "Pay as you go, $0", which bills every cloud model per token. Reflection's Beam is the only model still labelled free. DeepSeek V4.1 Flash, Gemma 4 and GPT-OSS now cost money there (a reader reported it first). Local Ollama is still completely free.
  • AI21 Labs sunset the Jamba API, Maestro and its file library on August 9, 2026 and says a new platform is coming.
  • Nscale needs a $5 prepaid top-up before the first request, so it moved from "trial" to "no longer free".
  • Alibaba Model Studio removed the new-user free quota from older models on October 29 (Qwen-Max, Qwen-Plus, Qwen3.6-Plus and more). Newer models such as Qwen3.8 Flash and Qwen3.8 Max still come with about 1M free tokens for 90 days.

Card at sign-up: a growing trend

We now show a "Card required" label in the directory, because "free" increasingly means "free, once your card is on file". Cerebras, Nebius, Nous Portal and Hetzner's experimental API fall in that bucket today. Hetzner's inference is still free of charge, but you sign in with a normal Hetzner account, and two readers (one on GitHub, one in our report form) were asked for full billing details and a card.

Part 3: Free models that came and went

New free models

  • NVIDIA NIM: Step 5 Preview (StepFun's agentic vision model, 1M context; thanks to the visitor who asked for it), Mistral Large 4, Kimi K2.6, GLM-5.3 Flash, MiniMax M3.1 Flash, Gemma 4 31B and Nemotron 3 Super are all on the free trial endpoints.
  • Cloudflare Workers AI: Kimi K2.7 Code, GLM-5.3 (1M context), DeepSeek V4 Pro and Nemotron 3 Super joined the 10,000 free neurons a day. The big models burn neurons fast.
  • Google AI Studio: Gemma 4 31B and 26B are free on the Gemini API. Gemini 4 "Argon" launched, but it's paid only.
  • Mistral: Mistral Large 4 (1M context) is in public preview. Mistral Large 3 and Medium 3.5 were deprecated on October 31 and retire in April 2027.
  • OpenRouter: Recursive's Flint (a 40B coding MoE) is free, and a 1M-context stealth model, Crouton Alpha, appeared on OpenRouter, Kilo and Nous.
  • ModelScope and OVH AI Endpoints added a dozen models each to their free catalogs.

Free models that disappeared

  • Poolside Laguna (S 2.1 and XS 2.1) is gone from every free gateway we checked (OpenRouter, Kilo, Nous); the free period ended around October 31.
  • OpenRouter also dropped the :free versions of MiniMax M3.1 Flash (reported by a reader), Thinking Machines Inkling and Inkling Small, and Nemotron 3 Super. The paid versions remain.
  • Nous Portal's free list shrank to Solar Mini 4, Step 3.7 Flash, Beam and Crouton Alpha.
  • Scaleway renamed DeepSeek V4.1 Flash to deepseek-v4.1-flash-1017, so the old id stopped working (another reader report). The 1M free tokens still apply.

Part 4: The big releases this autumn

Most frontier launches since September are closed and paid: Claude Opus, Sonnet and Haiku 5.5, GPT-6 Sol and Luna, GPT-6.1 Terra, Grok 4.7 and Gemini 4 Argon. You won't find them on a real free tier (and services that sell them "80% off" are reselling shared accounts; we reject those). The good news is on the open-weight side, where the models are free to run and quickly show up on free APIs:

  • Gemini 3.8 Flash (closed, but free on Google AI Studio)
  • DeepSeek V4.1 Flash and its October 17 update, MIT license
  • Mistral Large 4 (1T total, 52B active), open weights
  • Reflection Beam 1.0 (501B), Apache 2.0, free on several gateways
  • MiniMax M3.1 Flash, Step 5 Preview, Nemotron 3.5 Super VL
  • Meta Muse Spark Open (816B), Meta's first big open release in a while
  • Kimi K3.1 Pro: weights published on November 2 (2.8T parameters, not yet free anywhere)
  • Small and local: Flint 40B, LFM3, Nemotron 3.5 Bolt, Gemma 4 and Qwen3.8 27B, which has become the default local model.

Part 5: Claude Code on free models, done right

A reader pointed out on GitHub that our Claude Code instructions were wrong, and they were right. Claude Code speaks Anthropic's Messages API, so pointing it at an OpenAI-style URL fails. We rewrote the setup generator and wrote a guide to free coding models and Claude Code. In short: OpenRouter, Z.AI, FreeInference, Ollama and LM Studio have native Anthropic endpoints. For Groq, Cerebras, NVIDIA or Mistral, put LiteLLM or claude-code-router in front.

Dates to put in your calendar

  • November 16: Scaleway retires glm-5.2 (requests go to GLM-5.3).
  • November 24, 00:00 PT: IteraCompute's 30M tokens/month promotion ends.
  • December 1: Google redirects gemini-2.5-flash to 3.5 Flash-Lite and gemini-2.5-pro to 3.8 Flash.
  • December 31: the free Dots 3 Note preview on OpenRouter and Kilo expires.
  • April 30, 2027: Mistral Large 3 and Medium 3.5 retire.

How we checked

Every change above comes from the provider's own pricing page, documentation or public /v1/models catalog, fetched on November 4. Where a page needs a login we couldn't complete (Coze, ModelScope's quota page, AnyAPI's model list), the provider page says so instead of guessing. Coze is now marked "unconfirmed" for that reason. Every single change is in the changelog, with its source, and you can follow it by RSS.

Spotted something we got wrong? Use the Report an Issue button on any provider page. This month alone, five reader reports led to fixes.

Previous report: Free LLM APIs in October 2026. Related: best free LLM APIs and building a fallback chain.