Raw notes: topics

Cheap access to strong models

Check date: 15 November 2026. This log backs the guide page Strong models for less. Official vendor pages only. “not published” = the opened official page omits it. “UNVERIFIED” = the official page returned 403/404 and the figure comes only from a search-index snippet of that page. Standing free tiers (AI Studio, Groq, Z.AI, OpenRouter, Cohere, NVIDIA, Cerebras), relays, and subscription sign-ins are in research/02 and research/03 and are not repeated except where a fact changed the page.

Light month = 4,800 steps × (3,000 input + 800 output tokens), no caching, the same shape as research/02. Strong = a 4 or 5 on the guide's 1–5 judgment scale.

1. New-account credits and trials

Provider	What you get	Card	Agent via API?	Catch	Source
Google Cloud Free Trial	“$300 Welcome credit to spend over 90 days”	Yes, “a credit card or other payment method”	Effectively no for models. “The $300 credit can't pay for Gemini API in AI Studio costs.” It “can't pay for Gemini platform (formerly known as Gemini Enterprise Agent Platform, Vertex AI) services” “if you sign up for a free trial or upgrade to a paid account on or after November 16, 2026.” Also excludes “a generative AI partner model that is offered as a managed API”.	One-time. Only if you have “never been a paying user” and “haven't previously signed up”. Re-confirmed by a second fetch on 15 Nov 2026.	https://docs.cloud.google.com/free/docs/free-cloud-features
Google express mode	Free “for up to 90 days, within the specified quotas” for new Google Cloud users with an `@gmail.com` account	“You don't need to provide billing information”	Yes (API key). Models include `gemini-3.1-pro-preview`, `gemini-3-flash-preview`, `gemini-2.5-pro/flash/flash-lite`. Gemini 3 rows “Dynamic” RPM; 2.5 rows 10 RPM.	Preview/pre-GA. Removed for existing Cloud users or once billing is enabled. Free vs paid models, OpenAI compatibility, training use: not published.	https://docs.cloud.google.com/vertex-ai/generative-ai/docs/start/express-mode/overview
AWS Free Tier	“USD $100 in credits after you create an account” plus “up to an additional USD $100”	“A payment method isn't required to sign up for most new customers”	Bedrock coverage not published.	Free plan ends at 6 months or when credits run out. “Only to new AWS customers.”	https://docs.aws.amazon.com/awsaccountbilling/latest/aboutv2/free-tier.html, https://aws.amazon.com/free/free-tier-faqs/
Azure free account	“$200 credit to use on Azure services within 30 days” + 12 months of free services	Yes, credit or debit card (non-prepaid), $1 authorization reversed	Azure OpenAI / Foundry coverage not published. A Microsoft Q&A answer (not official docs) says Foundry quota is 0 on Free Trial: UNVERIFIED.	“Limit of one account with free services and $200 credit per new customer.”	https://azure.microsoft.com/en-us/pricing/purchase-options/azure-account
Anthropic Console	“New users receive free credits to test the API.” Amount not published.	not published	Yes	One-time.	https://platform.claude.com/docs/en/about-claude/pricing
OpenAI API	No signup grant published. $5 minimum prepaid purchase; purchased credits expire after 1 year (snippet; page 403): UNVERIFIED.	Yes	—	—	https://help.openai.com/en/articles/8264644
xAI	No standing free credit published; promo codes only; auto top-up minimum $5. Enterprise data-sharing credit wording: UNVERIFIED (403).	Yes	—	—	https://docs.x.ai/console/billing
DeepSeek	Granted balance is spent before topped-up balance; a new-user grant amount is not published.	—	—	—	https://api-docs.deepseek.com/quick_start/pricing
Together	“Together AI doesn't currently offer free trials.” $15 minimum first purchase.	Yes	—	—	https://docs.together.ai/docs/billing-credits
Fireworks	Signup credit not published.	not published	—	—	https://fireworks.ai/pricing
DeepInfra	“You have to add a card or pre-pay or you won't be able to use our services.” No signup credit published.	Yes	—	—	https://deepinfra.com/pricing

2. Credits bundled with a paid plan

Claude Max ($100/month) / Ultra ($200/month). Re-opened 15 Nov 2026. Monthly API credit “$100 USD (Max), $200 USD (Ultra)”; covers “Claude API, Claude Managed Agents, the Claude Agent SDK, and playground”; does not cover Claude Code, extra usage in the Claude apps, or Bedrock / Vertex / Foundry. Claimable after 7 days; expires at the end of each billing cycle, no rollover; “Payment method for Claude Platform: Not required.” When used up with no other credits, “API requests stop until your next monthly API credits arrive. Usage is never charged to your Claude plan.” Purchased credits and auto-reload continue spending if present. Credit usage counts toward the org's monthly spend cap. Free, Pro, and Enterprise are not eligible. https://platform.claude.com/docs/en/about-claude/api-credits-for-subscribers
- Fit: Sonnet 5.5 light month $67.20 and the light hybrid $17.46 fit inside $100; Opus 5.5 for every step ($134.40) does not.
Google AI Pro ($19.99/month). Re-opened 15 Nov 2026. “$10 USD per month in GenAI and Google Cloud developer credits” (Ultra 5x $40, Ultra 20x $100), spendable on the Gemini API, Google AI Studio, Firebase, Gemini Enterprise Agent Platform. Each grant expires one year after it is granted. Applied to a Cloud Billing account; card requirement not published. Paid-tier Gemini traffic is not used for training (research/02). $10 ≈ 20% of a paid gemini-4-hydrogen light month ($50.40). https://developers.google.com/profile/help/benefits

3. Students, researchers, startups, open source

GitHub Copilot Student: “Free”, “verified students”, “Auto model selection only”, “Excludes third-party agents”; AI credit allowance not published; eligibility re-checked monthly. Sign-ups paused April–June 2026; reopening gradually from 17 June; current status not re-confirmed. https://docs.github.com/en/copilot/get-started/plans, https://github.blog/changelog/2026-06-17-copilot-individual-plan-sign-ups-are-reopening/
Azure for Students: “$100 credit, which can be used within 12 months,” “No credit card required”; Azure OpenAI coverage not published. https://azure.microsoft.com/en-us/free/students
Google AI Pro for students: free 1 year for “eligible U.S. college students ages 18+,” redeem by 31 Dec 2026, “Valid form of payment required,” then $19.99/month. Whether the $10 developer credit is included: not published. https://gemini.google/students/
Codex for Students: “$100 in ChatGPT credits to use in Codex,” US/Canada university students, one per student, 12 months; not API credits. https://developers.openai.com/community/students
Anthropic AI for Science: “up to $20,000 in API credits” over 6 months, academic/nonprofit researchers. https://support.claude.com/en/articles/11199177
OpenAI Researcher Access Program: max $1,000 API credits (snippet only, page 403): UNVERIFIED. https://help.openai.com/en/articles/10139500
Claude for Startups: “We're not accepting new applications right now.” https://claude.com/programs/startups
AWS Activate Founders: “Start with $1,000 USD … Up to $5,000.” https://aws.amazon.com/startups/credits
Microsoft for Startups: up to $5,000 after business verification, up to $150,000 over time; covers “Azure OpenAI services.” https://www.microsoft.com/en-us/startups
Google for Startups Cloud: page truncated; amounts not confirmed.
Open source: free Copilot Pro for “a maintainer of a popular open-source repository” (criteria not published). OpenAI Codex for Open Source: “Six months of ChatGPT Pro with Codex” plus API credits from a $1M fund; per-project amount not published. https://docs.github.com/en/copilot/how-tos/manage-your-account/get-free-access-to-copilot-pro, https://developers.openai.com/community/codex-for-oss

OpenAI complimentary daily tokens for sharing traffic: help article 403. Search snippets say opt-in by org owners, daily pools per model group, a positive prepaid balance required, not every org eligible, and shared inputs/outputs used for training. Numbers and the covered model list: UNVERIFIED. The guide page says “could not confirm the current terms” and gives no numbers. https://help.openai.com/en/articles/10306912

5. Prompt caching

Provider / model	Cache read	Cache write	Minimum	Lifetime	Automatic?	Source
Anthropic `claude-opus-5-5`	$0.20 (0.05x of $4)	5m $5 (1.25x); 1h $8 (2x)	512 tokens; shorter prompts “are processed without caching, and no error is returned”	5 min, “refreshed for no additional cost each time the cached content is used”	Opt-in (`cache_control`)	https://platform.claude.com/docs/en/build-with-claude/prompt-caching (re-opened 15 Nov)
Anthropic `claude-sonnet-5-5`	$0.10	5m $2.50; 1h $4	512	same	Opt-in	same
OpenAI `gpt-6-astra`	$1.00 (0.1x)	1.25x for default 30 min; 2.5x for 12h (“GPT-5.6 and later”)	1,024	“30m” default	“enabled by default”	https://developers.openai.com/api/docs/guides/prompt-caching (re-opened 15 Nov)
OpenAI `gpt-6.1-terra`	$0.05 (0.05x)	same multipliers	1,024	same	On by default	same
Google `gemini-4-hydrogen` paid	$0.075	not published	4,096 tokens for Gemini 4 Carbon, Argon, Hydrogen; explicit caching “not supported … for Gemini 4 models”	not published	“Implicit caching is enabled by default”	https://aistudio.google.com/docs/caching (redirect from ai.google.dev, re-opened 15 Nov)
DeepSeek `deepseek-flash`	$0.003 off-peak / $0.006 peak	none published	not published	“usually within a few hours to a few days”; “best-effort”	On by default	https://api-docs.deepseek.com/quick_start/pricing
DeepSeek `deepseek-v4-pro`	$0.022 / $0.044	none published	not published	same	On by default	same

A 3,000-token step cannot be cached on Gemini 4 (below 4,096). Caching never touches output; output is 57% of an Opus step ($0.016 of $0.028).

Illustration, not a measurement

Assumption: 2,000 of each step's 3,000 input tokens are a repeated prefix read from cache; 1,000 new; 800 output.

Opus 5.5: 2000×0.20 + 1000×4 + 800×20 = 20,400 µ$ = $0.0204/step; ×4,800 = $97.92 (vs $134.40). One 5-minute write per 20-step job (240 jobs): 240 × 2000 × (5 − 0.20)/1e6 = $2.30 → about $100.
Sonnet 5.5: 2000×0.10 + 1000×2 + 800×10 = $0.0102/step; ×4,800 = $48.96 (vs $67.20); writes 240 × 2000 × 2.40/1e6 = $1.15 → about $50.
Hybrid caution: with Opus on 1 step in 10, its 5-minute cache can expire between Opus steps. A step that writes every time: 2000×5 + 1000×4 + 800×20 = $0.030, above the uncached $0.028. Best case (all reads hit): 480×0.0204 + 4,320×0.000636 = $12.54 vs $17.46.
Automatic caches, same assumption: gpt-6.1-terra $24.48 before write charges (vs $33.60); deepseek-flash off-peak $3.05 (vs $4.46); deepseek-v4-pro off-peak $10.98 (vs $17.11), all-peak $21.96.

Whether any of the ten agents keeps its prefix byte-identical: not confirmed.

6. Batch, flex, off-peak

Lever	Discount	Delay	Interactive agent?
Anthropic Message Batches	50% (Opus $2/$10, Sonnet $1/$5); stacks with caching	“most batches completing within 1 hour”; expire after 24 h	No
OpenAI Batch	50%	“completes within 24 hours”	No
OpenAI Flex (`service_tier: flex`)	“Tokens are priced at Batch API rates” (Terra $0.50/$2.50, Astra $5/$25)	“slower response times and occasional resource unavailability”; “429 Resource Unavailable”, not charged (re-opened 15 Nov, https://developers.openai.com/api/docs/guides/flex-processing)	Only if the agent can set `service_tier`: not confirmed for any agent. Light month Terra $16.80, Astra $168.
Gemini Batch	“50% of the standard cost”	target 24 h	No. A Flex row at the same price is on the pricing page; behaviour: UNVERIFIED.
xAI Batch	“20% off standard rates” for the grok-4.x family listed	not published	No
DeepSeek off-peak	50%, automatic outside 01–04 and 06–10 UTC Mon–Fri	none	Yes
Groq Batch / Flex	Batch 50% (no stacking with caching); Flex same price, 10x rate limits	24h–7d / none	Flex yes, no saving
Kimi batch	60% of standard (`research/02`)	not checked	No
Fireworks batch	50%	not checked	No

7. Flat-rate plans

Z.AI GLM Coding Plan. “Starting at just 18 USD per month” (https://docs.z.ai/devpack/overview). Pro/Max current prices not published on pages that rendered. Quotas in credits: Lite 2,000 per 5 h / 10,000 weekly; Pro 12,000 / 60,000; Max 28,000 / 140,000; off-peak 50% credits outside Mon–Fri 14:00–18:00 UTC+8 (report B). Supported tools page (https://docs.z.ai/devpack/tool/others, opened 15 Nov): “The GLM Coding Plan is limited to use within the following officially supported tools and product environments;” “users may not use their subscription benefits for tools or scenarios outside of this scope.” “1. Coding Agent Tool”: ZCode (1.5× Usage), Claude Code, Claude for IDE, Codex, OpenCode, Pi, Cursor, Cline, TRAE, Qoder, Droid, Kilo Code, Roo Code, Crush, Goose, Eigent. “2. General-purpose Agent Tool”: AutoClaw (1.5× Usage), OpenClaw, Hermes Agent, SillyTavern. (Report B saw a “best-effort support” label on the second group; the 15 Nov fetch did not render that label.) Agent Zero, Atomic Agent, NanoClaw, QwenPaw, Everfern, OpenManus: not listed. Endpoints on the same page: OpenAI Chat Completions https://api.z.ai/api/coding/paas/v4; Anthropic Messages https://api.z.ai/api/anthropic; model “Use custom” with example glm-5.2; “Incorrect endpoint configuration will result in inability to use GLM Coding Plan subscription quota.” Key: “Enter your Z.AI API Key”; separate plan key not confirmed. Usage policy (https://docs.z.ai/devpack/usage-policy): “Account sharing or multi-user access is prohibited”; rate limiting or account freezing; “Accounts with more than three violations may be banned”; “refunds are not supported.” FAQ (https://docs.z.ai/devpack/faq): “strictly limited to use within officially supported tools and products.”
- Fit: Z.AI's own GLM-5.3 estimate at 95% cache hit is 48–97M tokens/week on Lite (report B); a light month is about 4.3M tokens/week.
Alibaba Model Studio Token Plan (Personal). https://www.alibabacloud.com/help/en/model-studio/token-plan-personal-overview (opened 15 Nov): Lite $6, Essential $10, Standard $18, Pro $43; new-user Lite $3 / Essential $5 until 23:59:59 30 Nov 2026 (UTC+8), for accounts that never bought the plan. Models include qwen4-max-preview, qwen3.8-max, deepseek-v4-pro, deepseek-v4.1-flash, glm-5.3, glm-5.2; auto does not support tool_choice. “This plan is intended for interactive use only within coding tools and agent tools”; “It must not be used for automation scripts, custom application backends, or any non-interactive batch call scenarios”; violations “may result in subscription suspension or API Key banning”; “sharing is prohibited”; sharing a key “may be considered a violation and result in suspension.” “The only available region is Singapore.” Credits per call not published. Addresses: the tools page (https://help.aliyun.com/en/model-studio/more-tools) lists cn-beijing https://token-plan.cn-beijing.maas.aliyuncs.com/compatible-mode/v1 (OpenAI) and .../apps/anthropic (Anthropic), key from “Create or view API Key”; unsupported uses include Dify, n8n, Coze, Postman, and “automated scripts and application backends”. The Singapore address appeared only in a search snippet (https://token-plan.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1): UNVERIFIED. The quick-start page rendered without its steps.
- Older Coding Plan Pro ($50) carries older models and the same non-interactive ban (report B).
Kimi Code. https://www.kimi.com/code/docs/en/kimi-code/membership (opened 15 Nov): included on “Plus and above”; K3.1 Pro on Plus and above; 1M context on Pro and above; rolling 5-hour window, weekly limit removed for new members, monthly total; “Requests from the CLI, VS Code, Desktop, and third-party tools all count toward that quota.” Sidebar integrations: Claude Code, OpenCode, Codex, Hermes Agent. No tool restriction found. USD prices not published; China help page Plus ¥99, Pro ¥199, Max ¥699, Ultra ¥1,399 (report B). Base URL and API model ID: not on the page. Tool calling for K3.1 Pro: not confirmed this check.
MiniMax M Plan. FAQ (https://platform.minimax.io/docs/m-plan/faq): “M Plan is designed for individual, interactive developer use, with higher tiers offering higher usage limits.” “Pay-as-you-go is recommended for production use.” Names Claude Code, Codex, Cursor, OpenCode, Hermes Agent, OpenClaw; any OpenAI- or Anthropic-compatible tool. No stated penalty. Other tools page (https://platform.minimax.io/docs/m-plan/other-tools): https://api.minimax.io/v1 (OpenAI Compatible), https://api.minimax.io/anthropic (Anthropic Compatible), key starts sk-cp- from platform.minimax.io/console/plan, model MiniMax-M3.1-Flash. Prices and quotas not published; M3 inclusion not confirmed.
Ollama Pro/Max. Pro $20 ($200/yr) with $60 usage credits, 3 concurrent requests; Max $100 with $300, 10 concurrent. glm-5.3 $1.40/$4.40, deepseek-v4-pro $1.32/$3.96 (off-peak $0.66/$1.98 outside 12–18 UTC weekdays), minimax-m3 $0.60/$2.40, kimi-k3.1-pro $2/$10. No third-party-tool restriction printed. Light months: GLM-5.3 $37.06, V4 Pro off-peak $17.11, MiniMax-M3 $17.86, Kimi K3.1 Pro $67.20 (over $60). https://ollama.com/pricing
NanoGPT. $12/month, 60M input tokens/week (“Cached input counts in full”), output free on included open models, Monday 00:00 UTC reset; versions not named; excludes Claude, GPT, Gemini, Grok. https://nano-gpt.com/pricing
Cerebras Code. Pro $50 (24M tokens/day), Max $200 (120M/day), GLM 4.7, both “sold out”. https://www.cerebras.ai/pricing

8. Strong open-weight models, cheapest official host

Open weights confirmed on Hugging Face: zai-org/GLM-5.3, deepseek-ai/DeepSeek-V4-Pro, moonshotai/Kimi-K3.1-Pro, MiniMaxAI/MiniMax-M3 (license “minimax-community”). No Qwen3.8-Max repo: open weights could not confirm.

Model	Tool calling	Cheapest light month
MiniMax-M3, MiniMax API $0.30/$1.20 (“Permanent 50% off”; Fireworks same)	Confirmed 15 Nov: MiniMax's OpenAI-compatible API reference says “`MiniMax-M3.1-Flash` and `MiniMax-M3` support these additional Chat Completions parameters”, including `tools` (https://platform.minimax.io/docs/api-reference/text-openai-api); Fireworks marks “Function Calling” “Supported” (https://fireworks.ai/models/fireworks/minimax-m3). The HF card does not mention tool calling.	`4,800 × (3000×0.30 + 800×1.20)/1e6 = $8.93`
DeepSeek V4 Pro, off-peak $0.66/$1.98	Listed on DeepSeek's price card (`research/02`)	$17.11 off-peak; $34.21 all-peak; DeepInfra $1.30/$2.60 flat $28.70
GLM-5.3, DeepInfra	DeepInfra page tags the model “Function” (opened 15 Nov, https://deepinfra.com/zai-org/GLM-5.3); Ollama `glm-5.3:cloud` lists tools	Page shows $0.563/$2.50 (cached $0.125), “38% off”, list $0.90/$4.00, and a promotional Flex tier $0.45/$2.00. `4,800 × (3000×0.563 + 800×2.50)/1e6 = $17.71`; list $28.32; Z.AI $1.40/$4.40 = $37.06
Qwen3.8-Max, Global $1.65/$4.951	Not confirmed (Qwen pricing page silent)	$42.77
Kimi K3.1 Pro, DeepInfra $1.90/$9.50	Not confirmed this check	$63.84; list $67.20

The DeepInfra Flex tier's behaviour was not opened, so the guide uses the standard discounted price.

9. Spend controls

Anthropic: hard Monthly spend limit below the tier cap (Start $500, Build $1,000, Scale $200,000); HTTP 400 at your own limit, 429 enforced_spend_limit_reached at the tier cap. https://platform.claude.com/docs/en/api/rate-limits
OpenAI: optional hard limit (429 organization_spend_limit_exceeded / project_spend_limit_exceeded); “Enforcement is not instantaneous”; alerts “do not enforce a cap”. https://developers.openai.com/api/docs/guides/spend-limits
Google: project spend cap “Experimental”, ~10-minute billing latency, batch and agent sessions “may incur overages”; prepay $5–$5,000, keys stop at $0 (402), credits expire after 12 months.
DeepSeek: prepaid balance only, granted balance first.
OpenRouter: per-key limit, daily/weekly/monthly or lifetime reset; 403 “Key limit exceeded”. https://openrouter.ai/docs/api_reference/authentication

10. Ranked light months used on the guide page

Rank	Setup	Light month	Arithmetic
0	Gemini free key, `gemini-4-hydrogen`	$0 within project limit	`research/02`
1	MiniMax-M3, MiniMax API	$8.93	above
2	DeepSeek V4 Pro off-peak	$17.11	`4,800 × 0.003564`
3	Light hybrid, Flash + Opus 5.5	$17.46	`research/02`
4	GLM-5.3, DeepInfra discounted	$17.71	above
5	GLM Coding Plan Lite	$18 flat	listed tools only
6	Ollama Pro	$20 flat ($60 credits)	GLM-5.3 uses $37.06
7	`gpt-6.1-terra`	$33.60	`4,800 × 0.007`
8	Sonnet 5.5, caching illustration	about $50 ($67.20 uncached)	section 5
9	Opus 5.5, caching illustration	about $100 ($134.40 uncached)	section 5

11. OpenRouter free models (from the parent's 15 Nov check)

The 1,000 requests/day free limit is keyed to credits purchased over the account's life (“all time”, https://openrouter.ai/docs/api/reference/limits), so a one-time $10 purchase lifts it for good and the $10 stays as credit; a negative balance blocks free models. The Free Models Router openrouter/free “falls back only to other free models” and accepts tools / tool_choice (https://openrouter.ai/docs/guides/routing/model-variants/free, https://openrouter.ai/openrouter/free). A fallback list that includes a paid model is billed.

12. Routes outside a provider's terms (page policy)

Per the site owner's 15 Nov 2026 instruction for this page, terms-only routes (coding plans in unlisted agents, Alibaba Token Plan unattended, MiniMax M Plan unattended, personal chat plans in third-party agents, plan sharing) are described with their documented setup and the provider's stated consequences. Excluded entirely as possible fraud or unauthorized access: fake student or identity details, extra accounts to re-claim new-user credits or multiply free limits, stolen/leaked/resold keys, gpt4free-style scrapers, chargebacks and card tricks. Provider wording on re-claiming: Azure “limit of one account with free services and $200 credit per new customer”; Google Free Trial requires you “haven't previously signed up”; AWS credits “only to new AWS customers”; Codex for Students “one per student”.