Access guide
Where to access AI models in 2026
One page for the question every model launch raises: where can you actually run it, what does each route cost, and which free tiers are real. Official APIs, routers, and local weights for GLM-5.2, Kimi K3, Claude, GPT-5.6, and Gemini - every price traced to a provider page or one of our verified deep dives.
Last updated: August 2026. Prices are snapshots from the linked source posts; pricing pages are the only source of truth and they move. The OpenCode Go link is a Developers Digest referral link; you may receive a signup benefit and we may receive credits.
Short answer
Open-weights models (GLM-5.2, Kimi K3) are cheapest through routers and inference clouds, where hosts compete on price - start with OpenRouter. Closed models are their vendor's API plus whatever their subscriptions bundle: Codex now ships with every ChatGPT plan including Free, Gemini CLI has the most generous free quota, and Claude's plan value sits in Opus 5 and Sonnet 5 - Fable 5 is API-only. Genuinely local means a smaller model from the local models hub.
The three access lanes
Official API
The vendor's own endpoint: the reference price, first access to new features, and the full context window. For closed models (Claude, GPT-5.6, Gemini) it is effectively the only lane, plus cloud platforms like Bedrock and Vertex AI for Claude.
Routers and clouds
One key, consolidated billing, failover, and price competition between hosts. This lane only exists for open-weights models - which is exactly why GLM-5.2 undercuts its own list price on OpenRouter while Kimi K3 holds one uniform rate. The optionality argument is the deeper reason to care.
Local and self-host
Zero per-token cost if the weights are open and your hardware is real. The 2026 frontier open models are datacenter-class, so local for most developers means a smaller dense model via Ollama - see the local models hub for what actually fits a laptop.
Access at a glance
| Model | Official API ($/1M in / out) | Cheapest documented route | Free path | Open weights |
|---|---|---|---|---|
| GLM-5.2 | $1.40 / $4.40 (Z.ai) | OpenRouter $1.20 / $4.10 | OpenCode Go credits, ZCODE quota | Yes (MIT) |
| Kimi K3 | $3 / $15, $0.30 cached (Moonshot) | Same rate across most hosts | Vercel AI Gateway $5 / 30 days | Yes (Kimi K3 License) |
| Claude Fable 5 | $10 / $50 | API only | None (plan access ended) | No |
| Claude Opus 5 | $5 / $25 (Fast Mode $10 / $50) | API | Included in Claude plan limits | No |
| Claude Sonnet 5 | $2 / $10 through Aug 31, then $3 / $15 | API (promo window) | Included in Claude plan limits | No |
| GPT-5.6 (Sol / Terra / Luna) | $5 / $30, $2.50 / $15, $1 / $6 | Luna $1 / $6 | Codex on ChatGPT Free | No |
| Gemini | See Google AI pricing | CLI free quota | ~1,000 CLI requests/day | No |
Anthropic prices verified July 27, 2026 on the official pricing page; GLM-5.2 rates verified June 20 and Kimi K3 rates July 27 in the source posts linked below. We do not list a Gemini API rate here because we have not verified one; check Google's pricing page.
Model by model
GLM-5.2
Z.aiVerified June 20, 2026; updated July 23
MIT-licensed open weights, so it is hosted almost everywhere and several hosts undercut the first-party price. The cheapest hosted routes serve fp4 quantized weights; test your task before optimizing purely on price.
Official
- Z.ai per-token API: $1.40 input / $4.40 output per 1M tokens, cached input $0.26 (fp8, the reference price).
- GLM Coding Plan: Lite $18/mo, Pro $72/mo, Max $160/mo. Exposes an Anthropic-compatible endpoint, so pointing Claude Code at it is an env-var swap; the [1m] model suffix enables the 1M-token context.
Routers and clouds
- OpenRouter auto-router: $1.20 / $4.10, routing across 13+ hosts. No :free variant exists for GLM-5.2.
- DeepInfra: $1.20 / $4.20 (fp4). Fireworks and Novita: $1.40 / $4.40 (fp8).
Local and self-host
- Weights on Hugging Face under MIT with first-class vLLM and SGLang support. At roughly 756B total parameters this is datacenter-class, not laptop-class.
- Ollama's glm-5.2:cloud tag runs on Ollama Cloud GPUs, not your machine.
Free and near-free
- OpenCode Go: $5 in credits plus $5 off the first month with referral code M6HEHM4JM5.
- Devin's Pro plan bundles "free use of leading open source models" including GLM-5.2 - inside a paid plan, not standalone.
- Z.ai's ZCODE CLI has carried a large free quota (community reports near 300M tokens); confirm before relying on it.
- Hugging Face Inference Providers opened a limited free launch window; check the model page for current status.
Kimi K3
Moonshot AIVerified July 27, 2026
Open weights under the Kimi K3 License (not MIT), served by a dozen providers at a near-uniform $3 / $0.30 cached / $15 per 1M tokens. The differentiators are billing model, data boundary, and harness, not price.
Official
- Moonshot API: model kimi-k3, OpenAI-compatible, unlocks after a $1 top-up. $3 input / $0.30 cached input / $15 output per 1M tokens, full 1M-token context.
- Kimi Code plus Kimi chat, Agent, and Work share Moonshot's membership credit system.
Routers and clouds
- Together, Fireworks (US-hosted, zero data retention, fine-tuning), Modal, SiliconFlow (OpenAI and Anthropic formats), and Vercel AI Gateway all publish the same $3 / $0.30 / $15 rate.
- RunPod runs a flat $15 per 1M tokens public endpoint. Cloudflare Workers AI shows pricing in the dashboard only. OpenRouter lists one upstream provider, so no failover yet. Baseten has no public token rate.
Local and self-host
- Weights are live on Hugging Face, but at 1.4TB in MXFP4 with roughly 8x GB300 per serving replica, self-hosting is a datacenter project, not a local option.
Free and near-free
- Vercel AI Gateway grants $5 in free gateway credits every 30 days to teams without a paid plan.
- OpenCode Go: $5 for the first month, then $10/mo, with K3 available.
- No verified K3 route exists in Devin, which does not take third-party keys.
Claude Fable 5
AnthropicTracked June-July 2026
Anthropic's heavyweight coding model is API-only: the promotional window that included it in Claude subscriptions ended in July 2026 after several extensions, and it now runs on usage credits at $10 / $50 per 1M tokens.
Official
- Claude API at $10 input / $50 output per 1M tokens, with the same model available through Amazon Bedrock, Google Vertex AI, and Microsoft Foundry.
- On subscriptions it drains the 5-hour rolling window roughly 2x faster than Opus or more; community analysis pegs Pro / Max 5x / Max 20x at roughly $20 / $100 / $200 of notional API spend per month (an estimate, not an official figure).
Routers and clouds
- No third-party hosts: Claude models are closed weights, so access is the Claude API and the cloud platforms above.
Local and self-host
- Closed weights - no local option.
Free and near-free
- None. Plan-included access was promotional and has ended; budget for usage credits or the API.
Claude Opus 5 and Sonnet 5
AnthropicVerified July 27, 2026
The subscription-friendly Claude tier: both models are included in Claude plan usage limits, and Sonnet 5 is in an introductory pricing window through August 31.
Official
- Opus 5: $5 / $25 per 1M tokens, with Fast Mode at 2x ($10 / $50). Opus 4.8 remains available as a legacy model at the same $5 / $25.
- Sonnet 5: promotional $2 / $10 through August 31, 2026, then $3 / $15.
Routers and clouds
- Closed weights: the API plus Claude subscriptions (Pro $20/mo, Max $100 or $200/mo) for Claude Code and the apps.
Local and self-host
- Closed weights - no local option.
Free and near-free
- No free API tier; plan-included usage inside Claude subscriptions is the practical low-cost path.
GPT-5.6 Sol family
OpenAIVerified July 27, 2026
OpenAI sells three tiers of the same generation - and bundles Codex access into every ChatGPT plan, including Free, which makes it the widest zero-setup route to a frontier model.
Official
- API: Sol $5 / $30 per 1M tokens for frontier reasoning, Terra $2.50 / $15 as the balanced tier, Luna $1 / $6 as the fastest and cheapest.
- Codex is included across all ChatGPT plans: Free, Go ($8/mo), Plus ($20/mo), Pro ($100 or $200/mo), Business, Edu, and Enterprise. Limits vary by plan and task size.
Routers and clouds
- Closed weights: access is the OpenAI API and ChatGPT plan bundling; there is no third-party hosting market.
Local and self-host
- Closed weights - no local option.
Free and near-free
- Codex on the ChatGPT Free plan is a real trial path for lighter use; there is still no standalone Codex free tier to budget around.
Gemini
GoogleVerified July 27, 2026
Google's free quota is the most generous no-card entry point of any frontier vendor, with the caveat that the free tier mostly serves Flash and throttles Pro quickly.
Official
- Gemini CLI is free at roughly 1,000 requests/day - mostly Flash in practice, with Pro throttling after a handful of large prompts (community reports put it at 4-15).
- Consumer plans: Google One AI Premium at $20/mo and Ultra at $250/mo. Paid API usage is a separate surface priced on Google's AI pricing page.
Routers and clouds
- Closed weights: the Gemini API on Google's own platform is the only serving path.
Local and self-host
- Closed weights - no local option.
Free and near-free
- The CLI free quota is the honest headline: generous for everyday use, but do not plan Pro-heavy workloads around it.
Also worth knowing
Kimi K2.5 via Kimi Code
Moonshot's previous-generation model runs at $0.60 / $2.50 per 1M tokens through Kimi Code, one of the cheapest agent-capable rates going if K3-level capability is more than the task needs.
Meta Muse Spark 1.1
Meta's first paid developer model: $1.25 / $4.25 per 1M tokens with a 1M-token context, tool calling, and MCP support, and $20 in free credits for US developers. Details in the pricing tracker.
The free stack
If the budget is zero: Windsurf's free tier, Augment's free Dev plan, Copilot Free (2,000 completions and 50 premium requests a month), and Cline with your own key or a local Ollama model cover a surprising amount of real work.
Gateways and proxies
An Anthropic-compatible proxy can point Claude Code at NVIDIA NIM, OpenRouter, DeepSeek, LM Studio, llama.cpp, or Ollama backends. The tradeoffs - auth, logging, tool-call translation - are covered in the model gateway breakdown.
How to pick a route
Just trying a model: take the bundled or credited route - Codex on any ChatGPT plan, Gemini CLI's free quota, OpenCode Go's first-month credits, or Vercel AI Gateway's rolling $5 for K3. None requires a standalone API commitment.
Production tokens on a budget: for open-weights models, let the router market work for you and accept that the cheapest hosts serve quantized weights - validate on your own task first. For closed models the vendor price is the price; the lever is picking the right tier (Sonnet 5's promo window, Terra or Luna instead of Sol).
Daily agentic coding: flat-rate plans beat per-token billing once you code with a model every day: the GLM Coding Plan, Claude Pro or Max, or a ChatGPT plan with Codex. Watch burn multipliers - Fable 5 consumes a Claude window roughly twice as fast as Opus, which changes which plan you need.
Control or compliance: self-host open weights on vLLM or SGLang, or pick a host with an explicit data boundary (Fireworks documents US hosting and zero retention for K3). This only pays off at steady, high volume.
FAQ
What is the cheapest way to access GLM-5.2?
On hosted APIs, OpenRouter's auto-router ($1.20 input / $4.10 output per 1M tokens) and DeepInfra ($1.20 / $4.20) are the cheapest documented routes, both serving fp4 quantized weights. Z.ai's first-party API is the fp8 reference at $1.40 / $4.40. For trials, OpenCode Go referral credits and Z.ai's ZCODE CLI quota cost nothing up front.
Is there a free way to use Kimi K3?
Close to free: Vercel AI Gateway grants $5 in gateway credits every 30 days to teams without a paid plan, and OpenCode Go is $5 for the first month with K3 available. The weights are open, but at 1.4TB in MXFP4 needing roughly 8x GB300 per replica, self-hosting is not a free-tier substitute. There is no verified K3 route in Devin.
Is Claude Fable 5 included in Claude subscriptions?
Not anymore. The promotional window that included Fable 5 in Pro, Max, Team, and Enterprise plans ended in July 2026 after several extensions. Fable 5 now runs on usage credits or the API at $10 input / $50 output per 1M tokens, and it burns the 5-hour usage window roughly twice as fast as Opus.
Should I use an official API, a router, or local weights?
Official APIs give the reference price and first access to new features. Routers like OpenRouter consolidate keys and billing, add failover, and surface price competition between hosts - most useful for open-weights models served by many providers. Local weights only make sense for open models sized for your hardware; GLM-5.2 and Kimi K3 are both datacenter-class, so most developers should use a hosted route.
Can I run any of these models locally?
GLM-5.2 has MIT-licensed weights with vLLM and SGLang support, but at roughly 756B parameters it targets multi-GPU rigs; Ollama's glm-5.2:cloud tag runs on hosted GPUs, not your machine. Kimi K3's weights are open but not local-viable at 1.4TB. Claude, GPT-5.6, and Gemini are closed weights. For models that genuinely run on your hardware, see the local AI models hub.
Continue reading
- Where to Run GLM-5.2 Free and Cheap: Every Provider Compared - the full GLM-5.2 route map
- Where to Access Kimi K3: Every Provider and Price Compared - all twelve K3 providers in detail
- AI Coding Tools Pricing 2026 - the running tracker behind the closed-model numbers here
- How Claude's Usage Limits Actually Work With Fable 5 - windows, multipliers, and burn rates
- Free Claude Code and the Model Gateway Tradeoffs - routing one agent across many backends
Pick the tool once you have picked the route
The tool directory has verified profiles for every router, CLI, and agent on this page, and the local models hub covers what runs on your own hardware.
Get Smarter About AI Dev
New tutorials, open-source projects, and deep dives on coding agents - delivered weekly.
One email per weekReal code, not theoryFree forever