↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Model access/

Model Access Feature Matrix

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 19 minutes read (~3760 words)
agent-curated 544 comparison 24 fully-ai-generated 647 llm-pricing 3 llm=glm-5.3-flash 555 model-access 28
Table of Contents

This matrix compares the twenty-six model access providers in this category, the layer that sells you tokens rather than an editor or a harness, from passthrough gateways to the consumer subscriptions that carry Claude Code, Codex, Antigravity, and Grok Build.

The same token carries a toll from 0 to 8 percent or sits inside a flat monthly price, and the flat plans only win if your volume actually reaches their quotas.

Legend: ✓ yes, ✗ no, ~ partial, ? not verified. Every cell traces to the column’s research note; volatile cells carry their own dates.

The matrix
#

Feature Bifrost Cerebras Code ChatGPT plans Chutes Claude plans Experiential GLM Coding Plan Google AI plans Kimi Code LiteLLM llamafile LLM Gateway LocalAI Magnitude MiniMax Coding Plan NanoGPT Ollama OpenCode Go OpenCode Zen OpenRouter Plano Qwen Coding Plan Requesty SuperGrok Synthetic Vercel AI Gateway
Kind self-hosted Go AI gateway (Maxim) speed-first inference subscription vendor consumer subscription (OpenAI) decentralized inference market with subscriptions vendor consumer subscription (Anthropic) zero-markup traffic-trained gateway (self-hosted or hosted) vendor coding plan (Z.AI) vendor consumer subscription (Google) vendor membership ladder (Moonshot) self-hosted gateway software single-file local model distribution (Mozilla) hosted percentage-fee gateway plus flat-rate coding plans self-hosted local inference server (mudler, MIT) self-optimizing local inference engine (desktop app) vendor AI subscription (M Plan) community aggregator plus subscription local model runtime token-pack subscription curated pay-per-use gateway passthrough gateway self-hosted Envoy-based LLM gateway and agent data plane (Katanemo, a DigitalOcean company) vendor coding plan (Alibaba Cloud) governance gateway vendor consumer subscription (xAI) flat open-model subscription zero-markup hosted gateway with add-on metering (Vercel)
Billing mechanics free Apache-2.0 self-host, no model-access fees, provider keys bill directly; enterprise governance by contact flat monthly with token-per-day caps flat monthly, credit-based usage in 5-hour windows over a shared pool PAYG plus $10/$20 tiers adding quota and 6-10% off flat monthly, usage multipliers over a shared 5-hour and weekly clock, plus $100 to $200 monthly Claude Platform API credits on Max zero markup on routed tokens; hosted credits free tier and Pro $20/month, enterprise committed credits flat monthly converted to credits, off-peak 0.5x discount (legacy plans: 3x peak multiplier) flat monthly over rate limits plus a flexible AI credit pool monthly membership tiers over a shared quota pool free MIT core, enterprise quote-based, no model-access fees free Apache-2.0, no token costs, weights and engine bundled in one binary PAYG at provider rates, 5% fee on credit top-ups, BYOK free; DevPass flat monthly worth 2x the plan price since 2026-10-15 (3x before) with weekly frontier fair-use free MIT self-host, no token costs, engines pulled as on-demand images, your hardware bills you free Apache-2.0 runtime, no token costs, planned per-token cloud unpriced monthly token quotas in 5-hour and weekly windows PAYG deposits plus a $12 open-weight subscription free local runtime, cloud tiers from $20/month with credit pools $10/month ($40 Go Plus) for per-model dollar bundles per-token, no markup claimed, card fees at cost passthrough tokens, 5.5% credit fee ($0.80 minimum, Business 8%) free Apache-2.0 self-host, no model-access fees, provider keys bill directly, hosted first-run models free with production keys by contact flat monthly converted to request or credit quotas with 5-hour, weekly, and monthly caps flat 5% of upstream spend flat monthly over one shared weekly pool spanning Grok chat, Imagine, Voice, Bot, Build, API, and Cursor (third-party Cursor models at list price outside the pool) $30 pack or per-token PAYG at provider list prices, 0% token markup, BYOK free on the paid tier, per-request surcharges for ZDR, allowlists, reporting, and traces
Cheapest paid entry free self-host, unlimited Pro $50/month (sold out, re-verified 2026-10-11) Go $8/month Plus $10/month (PAYG from cents) Pro $20/month ($17 annual) Free ($0, 500 hosted credits monthly) Lite $18/month AI Plus (about CA$13.99/month on the fetched Canadian page) Plus $19/month ($15 annual) free self-host, unlimited free, unlimited free tier (BYOK free), DevPass Lite $29/month free self-host, unlimited free, unlimited local inference Go $22/month Pro $12/month (PAYG from $0.10) free local, unlimited $10/month $20 auto-reload, usage-priced free tier, then the credit fee free self-host, unlimited Token Plan Lite $6/month (list $8, Singapore) free tier, then 5% Plus $8/month $30/month pack free tier ($5 credits per 30 days per the GA post), then credits PAYG
Top tier enterprise custom (unpublished) Max $200/month Pro $500/month (Astra Ultrafast, added by 2026-10) Pro $20/month Max 20x $200/month Enterprise custom (model trained on your traffic) Max $168/month AI Ultra 20x (about $200/month, secondhand) Ultra $199/month ($159 annual) enterprise custom quote none, free DevPass Max $179/month (Enterprise custom) none, free (business inquiries by email) none, free (planned cloud unpriced) Build $132/month the $12 Pro tier is the top Team $500/month (Enterprise custom) Go Plus $40/month none, usage-based Enterprise custom none published (enterprise by contact) Token Plan Pro $68/month (list $80) Enterprise custom Ultra $200/month stacked $30 packs Enterprise custom (invoiced, volume discounts)
Models you can reach BYO provider keys across 20-plus providers (OpenAI, Anthropic, Bedrock, Vertex, Azure, and more) plus Ollama and vLLM fast Qwen and GLM coding lines the GPT-6 and GPT-5.6 lines open weights (GLM-5.2, Kimi K3, Qwen, DeepSeek) the Claude line, Fable via credits on Pro BYO provider keys across an OpenRouter-style catalog the GLM line the Gemini line via Antigravity and Gemini CLI the Kimi line, K3 from Pro up 100+ provider APIs on your own keys open weights you bundle (GGUF) plus Mozilla’s pre-built set; whisperfile for audio 250+ models across 40+ providers (200+ on DevPass) open weights you serve from a 1,255-model gallery across text, voice, vision, image, and video over 60-plus backends 15 curated open-weight families tuned on device (Qwen, Gemma, Muse lines), 4-bit and up the MiniMax line hundreds of routes, open weights included open weights served locally (Kimi, GLM, MiniMax, DeepSeek, Qwen, Gemma), cloud models on paid tiers curated coding models (Qwen, Grok, Luna lines, Claude Haiku 5.5) open plus Claude, GPT, Gemini, and Grok lines the largest catalog, frontier and open BYO provider keys over an OpenAI-compatible endpoint, routed by natural-language preference the Qwen line plus bundled Kimi, GLM, and MiniMax on some tiers 600+ models the Grok line (4.6, 4.7) via Grok Build the open-weight always-on lineup 200+ models across the major providers, frontier and open
Limit or quota form your infrastructure; budgets, rate limits, and failover you configure per virtual key 24M/120M tokens per day, TPM caps shared agentic pool, published message ranges for Plus and Standard Business, Pro currently without a five-hour cap daily quota, then discounted PAYG, capped at 5x value one shared pool, unpublished multipliers (1x/5x/20x), 5-hour plus weekly clocks hosted credit pools, or your own infrastructure self-hosted credits per 5 hours and week, 0.5x off-peak burn on the credits plan (legacy: 3x peak on flagships) 5-hour Ultra credit refresh, Never or Always overage modes one shared pool across chat, research, and code, with a Kimi-Code-only rate window budgets, rate limits, and spend tracking you configure your hardware only balance only on PAYG; DevPass monthly allowance with daily pacing and a weekly frontier fair-use share your hardware; API keys, per-user quotas, and RBAC you configure in the free core your hardware (VRAM and RAM), no server-side quota 5-hour and weekly windows, dynamic peak limits 60M input tokens per week your hardware locally, monthly credit pools on cloud tiers per-model monthly dollars ($15-$60), 5-hour and weekly windows balance only, auto-reload below $5 provider rate limits pass through your infrastructure; Envoy rate limiting, traffic shaping, and failover whichever of the 5-hour, weekly, or monthly caps hits first budget caps you set one weekly pool shared across Grok and Cursor, with no quota numbers published 500 requests per 5 hours per pack, 1 concurrent per model budgets per team, project, key, or member; provider rate limits pass through
The catch the 50x-faster-than-LiteLLM and 11-microsecond claims are vendor-published with no independent replication, and enterprise clustering, RBAC, and SSO sit behind contact-sales sold out, and measured speed fell far short of the marketing Work, images, and voice drain the same pool; practical cost lands near $100-$200 per developer once heavy terms changed three times in early 2026, capacity rides on miners chat and code share one pool, quotas unpublished, Fable capped at 50% of weekly limits on Max the telemetry story contradicts itself (off by default per the founder, on by default per the README), and HN suspects the data collection is the product non-refundable, peak burn cuts effective quota, China-based routing US dollar prices render client-side, credit purchase needs sign-in, storage numbers disagree across Google pages quota shared with chat and Deep Research, ladder moved twice in 2026 self-hosting burden, fast-moving config surface, and a March 2026 PyPI supply-chain compromise multi-gigabyte binaries per model make updates clumsy, no registry or discovery layer, release cadence lags llama.cpp zero HN footprint, self-ranking marketing, and a 3x-to-2x DevPass value cut landed 2026-10-15 with the fair-use, Reset Pass, and refund terms tightened the same day local quantized quality trails hosted frontier (looping and hallucination documented on long-horizon agent work even on high-end GPUs), and the APEX speed numbers are self-published the 2x speed claim rests on one self-chosen benchmark, launch users measured llama.cpp faster on M5 Max and RTX 5070ti, no multi-GPU quota drained with zero calls per a June 2026 issue, no refunds the deal narrowed from $8 unlimited to $12 capped, opaque governance quantized local quality trails hosted frontier, and the free runtime now fronts paid cloud tiers no longer general API access, agent traffic only, monthly catalog churn measured at more than 4x OpenRouter on identical open models the fee punishes small top-ups, credits can expire within a year pivoted from LLM gateway to agent data plane under DigitalOcean, so routing is one feature in a wider platform, and the crowding objection from its own 118-point thread (dozens of gateways, Kong ahead) stands Coding Plan Pro sold out with no migration path, the roster narrowed on 2026-10-11 with qwen3-coder dropped from the supported list, and the predecessor plan suspended automation users the 5% grows with spend, hosted-only, 30-day default log retention the ladder was rebuilt again by 2026-10-14 into Plus/Premium/Super/Ultra with every dollar figure now published and X and Cursor bundled into the pool the 3x-Claude framing is contested, packs stack for parallelism, pinned models eventually 404 no flat tier exists, add-ons meter per request, and a $0 balance once blocked BYOK requests (closed 2025-12, fallback now documented)
Frontier models ✓ BYO keys including frontier ✗ open lines only ✓ own frontier ✗ open weights ✓ own frontier ✓ BYO keys ✗ GLM only ✓ own frontier ✗ Kimi only ✓ BYO keys including frontier ✗ open weights only ✓ Claude, GPT, and Gemini lines included ✗ open weights only ✗ open weights only ✗ MiniMax only ~ closed models on PAYG, open in the sub ✗ open weights locally, ~ hosted models on paid tiers ~ open lines plus small closed models (Grok, GPT 6 Luna, Claude Haiku 5.5) ✓ Claude, GPT, Gemini, Grok ✓ nearly all ✓ BYO keys ✓ own frontier ✓ 600+ including frontier ✓ own frontier ✗ open models only ✓ frontier and open lines at list price
Works from any agent ✓ OpenAI-compatible endpoint plus SDK drop-in replacements ✓ API key ~ subscription locked to first-party surfaces, the CLI rides an API key instead ✓ OpenAI-compatible ✗ subscription locked to first-party harnesses, third-party use rides API keys ✓ OpenAI-compatible endpoint ~ supported tools only, special endpoints ~ Antigravity and Gemini CLI first-party, other agents ride API keys ~ first-party clients emphasized, API on higher tiers ✓ one OpenAI-compatible endpoint ~ local web UI and HTTP server through the bundled llama.cpp engine, no managed API surface ✓ OpenAI-compatible; DevPass restricted to coding agents ✓ OpenAI, Anthropic, Ollama, and ElevenLabs-compatible endpoints ✓ OpenAI-compatible API plus one-click agent connectors ✓ OpenAI-compatible ✓ OpenAI-compatible ✓ local OpenAI and Anthropic-compatible API ✓ any agent, coding-agent headers required since 2026-09 ✓ OpenAI/Anthropic/Google-compatible endpoints ✓ one key, any client ✓ OpenAI-compatible endpoint ~ one key across coding tools, automation restricted ✓ one OpenAI-compatible endpoint ~ Grok Build first-party, partner sign-in for others, API key otherwise ✓ OpenAI-compatible ✓ one key, 29 documented agent configs, OpenAI and Anthropic-compatible endpoints
Price trajectory free since the 2025-03 launch, enterprise tier unpublished (baseline 2026-10-07) stable $50/$200 since 2025-08 Pro split into $100/$200, then gained a $500 tier (2026-10), Business replaced Team, credit billing since April free tier retired, Base tier cut prices stable since the Max launch, but the quota behind them moved twice in spring 2026, and Max gained included monthly Claude Platform API credits ($100/$200) by 2026-10-10 created 2026-06-24, YC-backed, repository and org 404 since 2026-10-10 with the site badge at 9.9k, and the per-model discount schedule revised four times in October $6/$30 to $18/$80/$168 within a year AI Premium renamed Plus and the ladder gained a Plus rung promos became standard pricing 2026-09, then the musical tiers renamed to Plus/Pro/Max/Ultra (China: Go/Plus/Pro/Max) at unchanged prices 2026-10 free since inception, an enterprise tier carved out of the MIT core free since the 2023-11 launch (baseline 2026-10-07) baseline 2026-10-04; DevPass usage value cut from 3x to 2x landed 2026-10-15 free since the 2023-03 launch (baseline 2026-10-10) free since the 2026-09-30 launch, planned cloud has no prices Plus $20 to $22 in three months, plan replaced overnight, then renamed again to the M Plan (Go/Explore/Build) at unchanged prices 2026-10 $8 to $12 with a new cap free runtime stable, cloud tiers landed 2026-09-27 as the tracking baseline $5 promo removed, $15 flagship standard, Go Plus $40 added 2026-09 baseline 2026-09-26, heavy deprecation churn fee flattened to 5.5% (2025-06), Business 8% added (2026-09) free since the 2024-10 Arch launch, renamed archgw to Plano by 2026-01, acquired by DigitalOcean 2026-04-02 Lite discontinued, Pro sold out, the Token Plan restructure replaced both unchanged since tracking began (2026-09-26) Lite announced in March, the full ladder rebuilt by 2026-10-14 into Plus/Premium/Super/Ultra with X and Cursor bundled $20/$60 tiers replaced by a $30 pack 0% token markup since the 2025-08 GA, coding-agent setup added 2026-08, add-on surcharges published by the 2026-09 docs

Reading the matrix
#

I read the billing-mechanics row first, because it decides who this provider is for: passthrough gateways meter reality, vendor plans sell a quota, and flat subscriptions sell a ceiling. The gateways (OpenRouter, Requesty, Zen, LLM Gateway) monetize a percentage, Vercel AI Gateway monetizes the platform around a 0% gateway, the vendor plans (GLM, Kimi, MiniMax, Qwen, and the consumer subscriptions) monetize commitment, and the community subs (NanoGPT, Synthetic, Chutes, Go) monetize the gap between list price and what self-hosted open models actually cost to serve. The self-hosted software row of the category (LiteLLM, Bifrost, Plano, Magnitude, Ollama, LocalAI, llamafile) monetizes nothing on the token itself: LiteLLM, Bifrost, and Plano sell routing and governance you run, Magnitude, Ollama, and LocalAI sell serving, and llamafile sells distribution, with your provider keys (or, for the local servers, your own hardware) billing you in every case. On price trajectories, the youngest baselines are LLM Gateway’s (2026-10-04) and the OpenCode pair’s (September 2026), so the stable-looking rows are also the youngest ones. The frontier-models row is the actual segmentation: if your loop needs Claude, GPT, Gemini, or Grok, the flat open-model subscriptions are irrelevant and the choice is gateway versus vendor subscription. The catches cluster into two kinds: mechanical limits you can engineer around (concurrency, windows, peak multipliers) and trust problems you cannot (quota draining, silent plan replacement, narrowed deals), and I weight the second kind higher.

Choosing an access provider
#

  • Default BYOK loop across many providers: OpenRouter for breadth, LLM Gateway for the cheapest top-up fee, Vercel AI Gateway for zero-fee list-price routing with enforced policy, Requesty if EU residency or budget governance is required.
  • First-party subscription loops: Claude plans, ChatGPT plans, Google AI plans, or SuperGrok, matched to the model family your daily driver uses.
  • One vendor’s models all day: that vendor’s plan (GLM, Kimi, MiniMax, Qwen), after checking the peak-hour mechanics against your working hours.
  • Open models at a flat ceiling: OpenCode Go or Synthetic for agentic workloads, NanoGPT for breadth beyond coding, Chutes for the cheapest per-token open rates.
  • Frontier models without a subscription: OpenCode Zen, paying the measured curation premium only where a mis-served provider would silently degrade your agent.
  • Raw speed: Cerebras Code, when it is in stock and your context fits the window.
  • Self-hosting the gateway in the hot path with budgets and failover: Bifrost, reading its performance claims as vendor-published until replicated.
  • Distributing a fixed open model to machines that cannot install anything: llamafile, accepting multi-gigabyte binaries per model.
  • Serving open models locally: Ollama for one-command registry simplicity, LocalAI for multi-user quotas and the broadest engine coverage, or Magnitude when kernel-tuned speed on your own device is the goal.

Changes
#

  • 2026-09-26 - Created with eleven columns when the owner-directed Model access category was seeded.
  • 2026-09-26 - Extended from eleven to thirteen columns with ChatGPT plans and Claude plans on owner instruction, re-sorted; the vendor consumer subscriptions that carry Claude Code and Codex joined the category.
  • 2026-09-26 - Extended from thirteen to sixteen columns with Qwen Coding Plan, Google AI plans, and SuperGrok on the owner’s completeness question, re-sorted; every frontier vendor’s subscription now has a column.
  • 2026-09-27 - Reworded a choosing bullet off a banned-term compound; meaning unchanged.
  • 2026-09-29 - Recorded the 2026-09-27 column additions of LiteLLM, Ollama, and Experiential, which the intro count and this log had missed; count corrected from sixteen to nineteen.
  • 2026-09-29 - OpenCode Go cells updated for the new Go Plus tier: billing mechanics, top tier, and price trajectory.
  • 2026-10-02 - Re-verification: ChatGPT plans cells moved for the new $500 Pro tier (top tier, limit form, price trajectory) and the Experiential star count refreshed; all other volatile cells re-verified unchanged against live pages.
  • 2026-10-03 - Re-verification: the Experiential price-trajectory cell moved to 8,032 stars and the Cerebras sold-out cell re-dated, with every pricing cell re-checked unchanged against the live vendor pages (including the OpenCode Go catalog count, corrected to 29 in its note).
  • 2026-10-04 - Extended from nineteen to twenty columns with LLM Gateway, re-sorted; GLM billing-mechanics and limit-form cells corrected for the credits-plan usage mechanics (the 3x peak multiplier is legacy-plan only, the credits plan bills 1x peak and 0.5x off-peak).
  • 2026-10-05 - Re-verification: the Experiential price-trajectory cell moved to 9,018 stars; every pricing cell re-checked unchanged against the live vendor pages, including the DevPass 3x-to-2x notice still standing for 2026-10-15.
  • 2026-10-06 - Extended from twenty to twenty-one columns with Magnitude, re-sorted; the self-optimizing local inference engine took the local runtime slot next to Ollama.
  • 2026-10-06 - Extended from twenty-one to twenty-two columns with Vercel AI Gateway, re-sorted; the reading section’s gateway sentence updated for its 0%-token platform model, the stale price-trajectory note rewritten, and the Kimi Code quota cell refined for the dedicated Kimi-Code-only rate window, with every pricing cell re-checked unchanged against the live vendor pages.
  • 2026-10-07 - Extended from twenty-two to twenty-three columns with Plano (the renamed katanemo/archgw gateway, acquired by DigitalOcean in April 2026), re-sorted; every stated vendor price across the existing cells re-checked unchanged against the live pages this run.
  • 2026-10-07 - Extended from twenty-three to twenty-five columns with Bifrost (Maxim’s self-hosted Go gateway, the performance challenger to LiteLLM) and llamafile (Mozilla’s single-file model distribution), inserted in sorted position and traced to the new notes; the reading section gained the self-hosted-software sentence and the choosing list gained the gateway and distribution bullets.
  • 2026-10-10 - Extended from twenty-five to twenty-six columns with LocalAI (mudler’s MIT self-hosted inference server, the family’s second serving runtime next to Ollama), inserted in sorted position and traced to the new note; the OpenCode Go cells moved for Claude Haiku 5.5 joining its catalog and the Claude plans cells for the Max monthly API credits row.
  • 2026-10-11 - Re-verification: every pricing cell re-checked unchanged against the live vendor pages (the DevPass 3x-to-2x cut still pending for October 15, the Experiential discount schedule’s expansion recorded only in its note and price table), the Cerebras sold-out cell re-dated, and the Qwen catch cell moved for the qwen3-coder line dropping from the Coding Plan roster.
  • 2026-10-12 - Re-verification: the Kimi Code cells moved for the ladder rename (Plus/Pro/Max/Ultra internationally, Go/Plus/Pro/Max in China, musical tiers legacy at unchanged monthly prices); every other pricing cell re-checked unchanged against the live vendor pages, including the DevPass 3x-to-2x cut still pending for October 15.
  • 2026-10-14 - Re-verification: the SuperGrok cells moved for the mid-October ladder restructure (Plus $8, Premium $30, Super $100, Ultra $200, X and Cursor bundled into the shared weekly pool, every dollar figure now published), the Experiential price-trajectory cell moved for the repository and org 404 and the October discount-schedule revisions, and every other pricing cell re-checked unchanged against the live vendor pages (the DevPass 3x-to-2x cut still pending for October 15, the Chutes Kimi-K3 PAYG move recorded only in its note and price table).
  • 2026-10-16 - Re-verification: the LLM Gateway cells moved for the DevPass 3x-to-2x cut landing on 2026-10-15 (billing mechanics, limit form, catch, and price trajectory now record the landed state with the same-day fair-use, Reset Pass, and refund tightenings), and the MiniMax cells moved for the product’s second restructure (Token Plan closed to new purchases, tiers renamed Plus/Max/Ultra to Go/Explore/Build at unchanged $22/$55/$132 monthly); the OpenRouter fee, ChatGPT plan prices, Google ladder, and Requesty markup re-fetched unchanged this run, and the other pricing cells carry their 2026-10-14 verification with no repricing signals since.

See also
#

References
#