Agents log 2026-09-26

(owner-prompted, model access seeded)

#

• Owner request (“Do we track opencode go? Let’s track all model access providers and their pricing”): created the Model access category for the layer that sells access to models (gateways, routers, vendor coding plans, flat subscriptions), with eleven member notes and Pricing and Price history first-class in every note [glm-5.3-flash]
• Membership: OpenCode Go ($10/month), OpenCode Zen, OpenRouter (passthrough plus a 5.5% credit fee, with the official Stripe joining announcement replacing the draft’s unconfirmed-talks framing), Requesty (flat 5% markup, EU residency), Synthetic ($30/month), NanoGPT ($12/month), Chutes ($10/$20 tiers), Cerebras Code ($50/$200, sold out today), GLM Coding Plan ($18/$80/$168), Kimi Code ($19-$199), MiniMax Coding Plan ($22/$55/$132) [glm-5.3-flash]
• Boundary drawn: hosted access providers only; editor subscriptions stay in Surfaces, vendor API list prices in the model provider matrix, and self-hosted gateway software out unless the owner asks [glm-5.3-flash]
• model-access-feature-matrix created with eleven columns and ten rows; the category index page created with eleven entries; section _index.md gained the matrix and category one-liners [glm-5.3-flash]
• Cross-links added (opencode, openrouter-rankings, model-selection-for-coding-tasks), each with a dated Changes bullet [glm-5.3-flash]
• Corrected in-run before commit: opencode-zen’s price table dropped four pre-baseline deprecation rows, requesty lost a stray verification line and a mangled link label, and openrouter’s Status was rewritten onto the Stripe announcement [glm-5.3-flash]

(owner-directed, verification stamps removed)

#

• Removed the “Facts below verified as of” line from all 213 research notes across the 21 categories; volatile facts keep their inline “as of” qualifiers and AGENTS.md’s note skeleton no longer includes the line [glm-5.3-flash]
• Caveats from three stamps: lmarena’s client-rendered-app caveat moved into its Changes; the other two lost nothing [glm-5.3-flash]
• AGENTS.md: the Status evidence line now says “as of a stated date” and the Changes rule no longer exempts verification-date bumps [glm-5.3-flash]

(owner-prompted, Claude and ChatGPT plans added)

#

• Owner question (“Why aren’t you covering Claude code and openai subscriptions?”): added claude-plans and chatgpt-plans, accepting that vendor consumer subscriptions shipping an agent are in scope for model access [glm-5.3-flash]
• claude-plans: Free/Pro $20 ($17 annual)/Max 5x $100/Max 20x $200, with the 2026 record (May 6 permanent 5-hour doubling, May 13 weekly +50 percent promotion through August 19, July 20 Fable tightened to 50 percent of weekly limits on Max and credits-only on Pro, July 25 Opus 5, the never-shipped Agent SDK split) [glm-5.3-flash]
• chatgpt-plans: Free/Go $8/Plus $20/Pro $100 (5x) and $200 (20x), with Business $25/user and the April 2 Team replacement [glm-5.3-flash]
• Boundary updated: editor-bundled access (Cursor, Copilot Pro) stays in Surfaces pending owner direction [glm-5.3-flash]
• model-access matrix extended from eleven to thirteen columns and re-sorted; category index, cross-links, and the section matrix one-liner updated [glm-5.3-flash]
• On owner instruction the commit also carried a concurrent refresh session’s in-flight work across roughly eighty notes [glm-5.3-flash]

(owner-prompted, model access completeness pass)

#

• Owner question (“Any other subscription providers?”): added qwen-coding-plan, supergrok, and google-ai-plans, all inside the vendor-subscription boundary, correcting my leads where fetched sources disagreed (the Qwen night rate is 40 percent of Credits) [glm-5.3-flash]
• qwen-coding-plan: Coding Plan Pro $50/month sold out, Lite discontinued 2026-03-20, Token Plan $6-$68 Singapore with a China early-bird ladder, Qwen plus bundled Kimi, GLM, and MiniMax; supergrok: Free with limited Grok Build, $30/$100/$300, one shared weekly pool across chat, Build, and API, Lite in testing; google-ai-plans: the Plus/Pro/Ultra ladder carrying Antigravity and Gemini CLI [glm-5.3-flash]
• Rejected with reasons: Poe, Mistral Le Chat, Perplexity (no coding-agent story), Together, Fireworks, Groq, DeepInfra (no subscription product), Bedrock, Vertex, Azure (enterprise clouds), and the discount marketplaces; Copilot and Cursor remain the owner’s pending boundary call [glm-5.3-flash]
• model-access matrix extended from thirteen to sixteen columns and re-sorted; category index, cross-links, and the section one-liner updated; an insertion-order defect was caught and fixed in-run [glm-5.3-flash]