Cloudflare Workers AI free models, and how I plug them into OpenCode#

I run a lot of coding through Cloudflare's edge GPUs instead of paying per-token to the big labs. Workers AI hosts 50+ models on Cloudflare's network in 200+ cities, and you call them with one API. The pricing unit is the Neuron, not the token, and every model publishes its own Neuron rate. The free part is real but bounded: 10,000 Neurons a day, resetting at 00:00 UTC, no rollover. A slow month does not buy headroom for a busy day.

One thing changed since I first wrote this guide. In July 2026 Cloudflare gated the heavy frontier models behind the Paid plan. If you try kimi-k2.6, kimi-k2.7-code, or glm-5.2 on a Free account now, you get a 403 (internal error 5035) telling you to upgrade. The Paid plan starts at $5 a month and keeps the same 10,000 free Neurons a day, with overage at $0.011 per 1,000 Neurons.

How the pricing actually works#

There is no unlimited free model. Everything burns Neurons, and the meter resets daily.

Tier	Free allocation	After the limit
Workers Free	10,000 Neurons/day	Request fails (error 3036, HTTP 429) until reset
Workers Paid ($5/mo)	10,000 Neurons/day	$0.011 / 1,000 Neurons

Two details worth knowing. First, when you go over on Free, the request does not queue or degrade, it just errors. Handle error 3036 specifically instead of treating every 429 the same. Second, per-model Neuron rates change as the catalog changes, roughly monthly. I quote August 2026 rates below, but read the current numbers on Cloudflare's Workers AI pricing page before you budget anything.

To put the free allocation in human terms: at 800 input tokens plus 200 output per request, 10,000 Neurons covers roughly 1,780 requests a day on Llama 3.2 1B, but only about 160 a day on Llama 3.1 70B. Small models make the free tier stretch. Large ones turn it into a demo budget.

Best coding models you can still use for free#

These are all reachable on the Free plan as of September 2026:

Model	Why I reach for it	August 2026 rate signal
@cf/ibm-granite/granite-4.0-h-micro	Agentic coding with function calling, very cheap per call	Budget tier, my default for routine work
@cf/meta/llama-3.2-1b-instruct	Basic but nearly free	2,457 neurons/M input tokens ($0.027/M)
@cf/meta/llama-3.1-8b-instruct-fp8-fast	Solid general coding	$0.045/M input tokens, my price baseline
@cf/qwen/qwen3-30b-a3b-fp8	MoE reasoning, only 3B active per pass	Great quality per Neuron
@cf/zai-org/glm-4.7-flash	131k context, 100+ languages, fast tool calling	Stays on Free
@cf/qwen/qwen2.5-coder-32b-instruct	Excellent dedicated coder, burns faster	Check current rate before a big session

My habit: granite or qwen3-30b for routine work, glm-4.7-flash when I need long context, and the 32b coder only for genuinely hard tasks.

These now need Paid, so don't burn an afternoon wondering why they 403:

Model	Note
@cf/moonshotai/kimi-k2.7-code	1T-param MoE coder, 262k context, Paid only
@cf/moonshotai/kimi-k2.6	Frontier agentic, $0.95/M input tokens, Paid only
@cf/zai-org/glm-5.2	Agentic coding flagship, Paid only
@cf/zai-org/glm-5.3-flash	320B MoE, 18B active, multimodal, Paid only

Also note: kimi-k2.5 was deprecated on May 30, 2026 and now aliases to kimi-k2.6 at the higher price. A batch of older models (llama-3.1-8b, llama-3.1-70b, mistral-7b v0.1, gemma-7b, phi-2 and others) was deprecated the same day. If a tutorial names a model that 404s, this is probably why.

New since the summer#

• GLM-5.3 and GLM-5.3 Flash (Aug 2026): Z.ai's flagship agentic coding models, 1M context on 5.3, multimodal on Flash. Both Paid.
• Qwen 3.8 27B: 27B instruction-tuned vision model for agents.
• Kimi K2.6 (Apr 2026) and Kimi K2.7 Code: frontier 1T-param MoE family, 262k context. Paid.
• DeepSeek v4 flash/pro (Aug 2026): 1M context options in AI Search.
• Gemma 4 26B A4B: 26B total, 4B active, 256K context, vision and thinking mode. Still on Free.
• Nemotron 3 Super (120B total, 12B active, hybrid Mamba-transformer): still on Free.

Full list lives at the Workers AI models page, which moves monthly.

The rest of the catalog#

Image generation#

Model ID	Provider
@cf/black-forest-labs/flux-1-schnell	Black Forest Labs
@cf/black-forest-labs/flux-2-dev	Black Forest Labs
@cf/black-forest-labs/flux-2-klein-4b	Black Forest Labs
@cf/black-forest-labs/flux-2-klein-9b	Black Forest Labs
@cf/leonardo/phoenix-1.0	Leonardo AI
@cf/leonardo/lucid-origin	Leonardo AI
@cf/bytedance/stable-diffusion-xl-lightning	ByteDance

Speech and audio#

Model ID	Type	Provider
@cf/openai/whisper	ASR	OpenAI
@cf/openai/whisper-large-v3-turbo	ASR	OpenAI
@cf/deepgram/nova-3	ASR	Deepgram
@cf/deepgram/flux	ASR (conversational)	Deepgram
@cf/deepgram/aura-1	TTS	Deepgram
@cf/deepgram/aura-2-en	TTS	Deepgram
@cf/deepgram/aura-2-es	TTS (Spanish)	Deepgram
@cf/myshell-ai/melotts	TTS (multilingual)	MyShell

Embeddings#

Model ID	Provider
@cf/google/embeddinggemma-300m	Google
@cf/qwen/qwen3-embedding-0.6b	Qwen
@cf/baai/bge-base-en-v1.5	BAAI
@cf/baai/bge-small-en-v1.5	BAAI
@cf/baai/bge-large-en-v1.5	BAAI
@cf/baai/bge-m3	BAAI

Other models#

Model ID	Task	Provider
@cf/meta/m2m100-1.2b	Translation	Meta
@cf/ai4bharat/indictrans2-en-indic-1B	Translation (Indic)	AI4Bharat
@cf/microsoft/resnet-50	Image Classification	Microsoft
@cf/huggingface/distilbert-sst-2-int8	Text Classification	HuggingFace
@cf/meta/llama-guard-3-8b	Content Safety	Meta
@cf/baai/bge-reranker-base	Reranking	BAAI

---

Plugging it into OpenCode#

Step 1: get Cloudflare credentials#

1. Go to https://dash.cloudflare.com/ and sign up or log in with Google or GitHub.
2. Open AI and then Workers AI in the sidebar.
3. Copy your Account ID from the dashboard.
4. Create an API token and copy it somewhere safe. Treat it like a password.

Step 2: connect the provider in OpenCode#

1. Open OpenCode (CLI or GUI, both work).
2. Press Ctrl + P for the command palette.
3. Pick Connect Provider.
4. Find Cloudflare Workers AI in the list.
5. Paste your Account ID and API Token when asked.

Step 3: pick a model and start working#

1. Press Ctrl + P and pick Switch Models to see what is available.
2. Start with something cheap like granite-4.0-h-micro for everyday work, and switch to a heavier coder when a task earns it.
3. Code normally. OpenCode routes the requests through your Cloudflare account.

Step 4: build from a PRD file#

The workflow I keep coming back to:

1. Write a PRD (Product Requirements Document) as prd.md in your project.
2. Open the project folder in the OpenCode CLI.
3. Prompt it with something like:

   According to my requirement build fullstack application using prd.md

4. OpenCode reads the prd.md, figures out the project shape, installs dependencies, and generates the frontend and backend code.
5. Review the diff before you run anything it produced.

Step 5: switch models when the task changes#

• Ctrl + P and then Switch Models works anytime.
• Cheap models (granite, small llama, qwen3) live comfortably inside the 10k daily budget.
• The Paid-only frontier models cost real money per call, so I save them for the tasks that actually need them.

---

Things that tripped me up#

• Some model names from older tutorials no longer exist. Kimi K2.5 now aliases to K2.6, and a set of 2024-era models (llama-3.1-8b, mistral-7b v0.1, phi-2 and friends) was deprecated in May 2026. If a name 404s, check the current catalog.
• Rate limits are per task type (text generation sits around 300 requests/min). That is separate from the Neuron budget.
• Watch usage at https://dash.cloudflare.com/?to=/:account/ai/workers-ai - the Neuron meter is the one that matters.

Where to read the current numbers#

• Cloudflare Workers AI models: https://developers.cloudflare.com/workers-ai/models/
• Cloudflare Workers AI pricing: https://developers.cloudflare.com/workers-ai/platform/pricing/
• Paid-plan change (July 2026): https://developers.cloudflare.com/changelog/post/2026-07-28-models-require-workers-paid/
• OpenCode