↓ Skip to main content
tomrochette.com
  1. Agents/

Model access

Author
Tom Rochette
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: in progress / Readability: 3 / 4 minutes read (~715 words)
agents 49 model-access 28
Table of Contents

The layer that sells access to models themselves: gateways and routers taking a percentage or, in Vercel’s case, nothing on the token itself, vendor plans selling a quota (including the consumer subscriptions that carry Claude Code and Codex), flat subscriptions selling a ceiling, and the self-hosted software (LiteLLM, Bifrost, Ollama, LocalAI, Magnitude, Plano, llamafile) that routes, serves, or distributes the same access for free. Editors and harnesses live in their own categories; this is where the token bill gets paid.

  • Bifrost - Maxim’s self-hosted Apache-2.0 Go gateway unifying 20-plus providers with failover, budgets, and an MCP gateway, the performance challenger to LiteLLM, 8.8k stars.
  • Cerebras Code - wafer-scale inference sold as speed, $50/$200 per month, currently sold out.
  • ChatGPT plans - OpenAI’s subscription ladder from Go $8 to Pro tiers up to $500 per month, every tier carrying Codex.
  • Chutes - decentralized inference with pay-as-you-go plus $10/$20 subscriptions capped at 5x pay-as-you-go value.
  • Claude plans - Anthropic’s Free/Pro/Max subscriptions, the only non-API way to run Claude Code, from $20 to $200 per month.
  • Experiential - a YC-backed Apache-2.0 gateway at zero markup that mines agent traces to train routers and, on enterprise, a model you own.
  • GLM Coding Plan - Z.AI’s flat monthly quota for the GLM line, from $18, restructured twice since launch.
  • Google AI plans - the Google Plus/Pro/Ultra ladder carrying Antigravity and Gemini CLI, with credits over rate limits.
  • Kimi Code - Moonshot’s membership ladder for coding, $19 to $199 monthly on the Plus/Pro/Max/Ultra tiers.
  • LiteLLM - a self-hosted MIT gateway unifying 100+ provider APIs behind one OpenAI-compatible endpoint, with routing, budgets, and spend tracking.
  • llamafile - Mozilla’s single-file LLM distribution folding llama.cpp and Cosmopolitan Libc into one cross-platform executable, the family’s distribute-instead-of-serve answer, 26.2k stars.
  • LLM Gateway - an open-source gateway charging 5% on credit top-ups with free BYOK, plus DevPass flat-rate coding plans from $29.
  • LocalAI - mudler’s MIT self-hosted inference server serving a 1,255-model gallery behind OpenAI, Anthropic, Ollama, and ElevenLabs-compatible APIs with multi-user quotas, 49.5k stars.
  • Magnitude - a self-optimizing local inference engine (YC S25) that tunes its kernels on your device, free under Apache-2.0.
  • MiniMax Coding Plan - token-quota subscriptions for the MiniMax line, $22 to $132 per month, born from a silent plan replacement and renamed again to the M Plan in October.
  • NanoGPT - the community aggregator: hundreds of routes pay-as-you-go plus a $12 open-weight subscription.
  • Ollama - the free local runtime and registry for open models, now paired with paid cloud tiers.
  • OpenCode Go - the OpenCode team’s $10/month open-model token pack, usable from any agent.
  • OpenCode Zen - the OpenCode team’s curated pay-per-use gateway over benchmarked endpoints.
  • OpenRouter - the largest model gateway, passthrough tokens plus a 5.5% credit fee, now joining Stripe.
  • Plano - the Envoy-based self-hosted LLM gateway and agent data plane behind the Arch-Router routers, acquired by DigitalOcean in April 2026.
  • Qwen Coding Plan - Alibaba Cloud’s flat quota for the Qwen line plus bundled rivals, in transition to a Token Plan.
  • Requesty - EU-residency gateway charging a flat 5% markup on upstream spend.
  • SuperGrok - xAI’s subscription ladder, rebuilt in mid-October 2026 from the $30-to-$300 SuperGrok tiers into Plus $8 to Ultra $200 with X Premium and Cursor bundled and one shared weekly pool.
  • Synthetic - a flat $30/month subscription for open-weight coding LLMs aimed at agent users.
  • Vercel AI Gateway - Vercel’s hosted gateway at provider list prices with 0% token markup, one-command setup for 29 coding agents, and no flat tier.

Its members are compared on shared rows in the Model Access Feature Matrix.

Changes
#

  • 2026-09-26 - Added Cerebras Code.
  • 2026-09-26 - Added Chutes.
  • 2026-09-26 - Added GLM Coding Plan.
  • 2026-09-26 - Added Kimi Code.
  • 2026-09-26 - Added MiniMax Coding Plan.
  • 2026-09-26 - Added NanoGPT.
  • 2026-09-26 - Added OpenCode Go.
  • 2026-09-26 - Added OpenCode Zen.
  • 2026-09-26 - Added OpenRouter.
  • 2026-09-26 - Added Requesty.
  • 2026-09-26 - Added Synthetic.
  • 2026-09-26 - Added ChatGPT plans.
  • 2026-09-26 - Added Claude plans.
  • 2026-09-26 - Added Google AI plans.
  • 2026-09-26 - Added Qwen Coding Plan.
  • 2026-09-26 - Added SuperGrok.
  • 2026-09-27 - Added LiteLLM.
  • 2026-09-27 - Added Ollama.
  • 2026-09-27 - Added Experiential.
  • 2026-10-04 - Added LLM Gateway.
  • 2026-10-06 - Added Magnitude.
  • 2026-10-06 - Added Vercel AI Gateway.
  • 2026-10-07 - Added Plano.
  • 2026-10-07 - Added Bifrost.
  • 2026-10-07 - Added llamafile.
  • 2026-10-10 - Added LocalAI.