Research

Official free tiers and cheap pay-as-you-go APIs

Check date: 2026-11-14. This is the ground truth for official free tiers and official pay-as-you-go APIs that can power an open-source agent. Every number below was printed on a vendor page opened that day. Where a vendor page does not print a number, this chapter says not published. Where a rate was not on a page that was opened, it says UNVERIFIED. Blog and forum figures are not used.

Two different “light” workloads appear in the notes. They are not interchangeable.

The standing $0 shortlist

Two official $0 options can run a tool-calling agent on a normal laptop. A light-day fit is skipped for both, because the quota that fit would be compared with is not published.

  1. Google AI Studio / Gemini Developer API, free tier. No billing account is required. Free qualifies as “Active project or free trial.” Custom function calling is part of the Gemini API. Code execution and URL context are priced “Free of charge” on the free tier. Computer use is “Not available” on the free tier, and the custom-tools endpoint for Gemini 3.1 Pro Preview is not on the free tier. Models priced “Free of charge” include gemini-4-hydrogen, gemini-3.6-flash, gemini-3.5-flash-lite, gemini-3.1-flash-lite, gemini-3-flash-preview (legacy), and Gemma 4 (no paid tier). gemini-4-carbon and gemini-4-argon are not on the free tier. The model to select is gemini-4-hydrogen. It publishes a 1M-token context window. Paid standard for that id is $1.50 input and $7.50 output per million. A previous-generation Flash id is also “Free of charge” on the free tier. Its paid standard price is $0.75 / $3.75 per million through 31 December 2026, then $1.50 / $7.50. The models page calls that id previous-generation Flash. It is not the model to select. The Gemini 3 developer guide says Gemini 3 models support 1M input / 64k output. Context for gemini-3.6-flash, gemini-3.5-flash-lite, gemini-3.1-flash-lite, and Gemma 4 is not published on the pricing page. This is the free tool-calling API this chapter keeps, and the model to select is gemini-4-hydrogen. The cost is data use: free-tier rows say content is used to improve Google’s products. Paid rows say it is not. Grounding with Google Search is “Not available” on the free tier for the current Gemini 3/4 Flash rows (2.5 Flash still prints 500 grounded requests/day, and that model is deprecated). Per-model requests per minute, tokens per minute, and requests per day are not published (they are shown per project in AI Studio). Limits are per project, not per key, and requests-per-day reset at midnight Pacific time. The rate-limit page says specified limits are not guaranteed and actual capacity may vary. The OpenAI-compatible base URL was not confirmed this check (https://ai.google.dev/gemini-api/docs/openai timed out). The native REST host that did load is https://generativelanguage.googleapis.com/v1beta/models/{model}:generateContent. gemini-2.5-flash and gemini-2.5-pro are no longer selectable in AI Studio.

From 1 Dec 2026, gemini-2.5-flash routes to gemini-3.5-flash-lite, and gemini-2.5-pro routes to gemini-3.8-flash. Neither destination is the model to select.

  1. Z.AI GLM-4.7-Flash. The pricing page lists input, cached input, cached storage, and output as Free, with no “limited-time” label. The model page gives a 200K context window, 128K max output, and function calling. The OpenAI-compatible base URL is https://api.z.ai/api/paas/v4/. A credit card is not stated on the pricing page. Rate limits are not published (console only; error 1302 is “Rate limit reached,” with no public number). This is second because the allowance might be tighter than the $0 price suggests. GLM-4.5-Flash and GLM-4.6V-Flash are also priced Free. Their context and tool support were not on a page opened in this check. GLM-5.3-Flash is a paid model at $0.15 / $0.50 per million input/output tokens. The Coding Plan is a separate paid subscription.

The public Z.AI privacy policy says it does not cover API Services and points developers to a Data Processing Addendum, which was not opened. Do not assume the consumer “improve our models” wording applies to this API.

Why OpenRouter, SambaNova, GitHub Models, Hugging Face, Mistral, and Together do not count

OpenRouter :free models publish a hard 50 requests/day when all-time credits purchased are under 10, shared across free models, plus 20 requests/minute. Failed attempts count toward the daily cap. Tokens per day are not published. A light day is 120 requests, and 120 is above 50, so the cap stops the day at 50/15 ≈ 3 tasks. The notes call the token quality high. The request cap is the loss. The 1,000-requests/day ceiling starts only after a credit purchase (the written threshold is 10 credits, and the higher ceiling starts at 9 credits to absorb top-up fees). A credit card is not stated. :free variants do not require a credit purchase, and a negative credit balance blocks free models until credits are added. Free variant IDs end in :free.

SambaNova’s free tier publishes 20 requests/day per model, plus 20 requests/minute and 200,000 tokens/day, for every model in the free tables. A light day is 120 requests against 20, so the request cap binds first. At 4,000 tokens each, 20 requests are 80,000 tokens, under the token cap. 20/15 ≈ 1 task per model per day, with 5 calls left. The vendor’s own pages disagree about the card: the rate-limit page says the free tier applies when no payment method is linked, and the Plans page says Free means “Add a payment method and purchase credits to run your first requests.” Preview models “may be removed at short notice.”

GitHub Models is retired. As of 30 Jul 2026, the playground, model catalog, inference API, and bring-your-own-key are no longer available.

Hugging Face Inference Providers gives free users monthly credits of None. Extra usage requires a credits purchase. PRO and Team/Enterprise get $2.00 per month (per seat for Team/Enterprise), which is a paid subscription. The inference overview still says the product “includes a generous free tier.” That sentence disagrees with the pricing table. The pricing table is the one used here. Hub rate limits on the website and file resolvers are a different quota.

Mistral is dropped as a standing free API. Usage and limits: “Free mode includes a monthly usage budget for trying Vibe and Studio. It doesn't include API access. Creating API keys and using the API require billing activation and pay-as-you-go.” The public pricing page’s free plan is the Vibe chat product, not a developer token API. API pricing lists Mistral Moderation 2 as Free. That row does not override the billing-activation requirement for API keys, and it is a moderation model rather than a tool-calling agent model. Numeric free-tier RPM/TPM is not published (the limits page says the figures are shown per model in the admin console).

Together lists Prism-ML/Ternary-Bonsai-27B as Free input and Free output, context 262,144, function calling Yes. The credits page says Together does not currently offer free trials, and using the platform requires a purchase of at least $15 of credits. Until that purchase, the playground and API keys stay unavailable. A $0 token price behind a $15 gate is not a standing free API. If you do pay to enter, prompts and outputs are stored and may be used for product improvements unless Zero Data Retention is enabled. Model training is off by default.

Also checked, and not a standing $0 API

Cheap pay-as-you-go prices

Prices are USD per 1 million tokens, copied from vendor pricing pages opened on 2026-11-14. “In” is uncached input unless a note says otherwise. Output prices include thinking tokens only where the vendor page says so. Gemini’s paid rows do: output includes thinking tokens. These are API list prices for a pay-as-you-go agent runtime.

Gemini rows below are the paid tier, standard. On the free tier, the models named in the shortlist stay “Free of charge.” gemini-4-carbon and gemini-4-argon are paid-tier models.

Model Input / 1M Output / 1M What the opened page adds
deepseek-flash (DeepSeek-V4.1-Flash-1017) $0.15 off-peak / $0.30 peak $0.60 off-peak / $1.20 peak Cache-miss input. Cache-hit input is $0.003 / $0.006. Both DeepSeek models list tool calls.
deepseek-v4-pro (DeepSeek-V4-Pro-0813) $0.66 off-peak / $1.32 peak $1.98 off-peak / $3.96 peak Cache-hit input is $0.022 / $0.044.
deepseek-chat, deepseek-reasoner UNVERIFIED UNVERIFIED Not on this price card. The 2026-04-24 note says both names would be inaccessible after 2026-07-24 15:59 UTC.
gemini-3.1-flash-lite $0.25 text, image, video; $0.50 audio $1.50 Cost-efficient, high-volume agentic tasks.
gemini-3.5-flash-lite $0.30 text, image, video, audio $2.50 Cost-efficient, high-volume agentic tasks.
gemini-3.6-flash $0.75 through 2026-12-31; $1.50 from 2027-01-01 $3.75 through 2026-12-31; $7.50 from 2027-01-01 On the price card. Not the model to select.
gemini-4-hydrogen $1.50 $7.50 Everyday coding and agentic workflows. Paid-tier rate. Free tier lists this model as Free of charge.
gemini-4-argon $2.00 through 2026-11-30; $4.00 from 2026-12-01 $10.00 through 2026-11-30; $20.00 from 2026-12-01 On 2026-11-14 the live rate is the introductory $2 / $10.
gemini-4-carbon $2.00 through 2026-11-30; $4.00 from 2026-12-01 $10.00 through 2026-11-30; $20.00 from 2026-12-01 Most capable Gemini 4 model, multi-step agents. Same introductory window as Argon.
Groq qwen/qwen3.8-27b $0.80 $4.00 Preview, not a free plan. The page says preview models are for evaluation and may be discontinued. The models page does not label any model free.
Claude Haiku 5.5, prompt ≤100,000 tokens $0.10 $0.50 5m cache write $0.125; 1h cache write $0.20; cache hit $0.01.
Claude Haiku 5.5, prompt >100,000 tokens $0.50 $2.50 5m cache write $0.625; 1h cache write $1.00; cache hit $0.05.
Claude Sonnet 5.5 $2 $10 5m cache write $2.50; 1h cache write $4; cache hit $0.10.
claude-opus-5-5, Claude Opus 5.5 $4 $20 5m cache write $5; 1h cache write $8; cache hit $0.20. Standard global rate. Fast mode is $8 / $40.
Claude Fable 5.2 $8 $40 Listed above Opus on the same table.
gpt-5-nano $0.05 $0.40 Standard, short context (≤272K). function_calling listed. Status: Deprecated.
gpt-6-luna $0.10 $0.50 Short context. Above 272K the whole request is $0.20 input and $0.75 output. Cached input on the short-context rate is $0.01. Function calling. Responses API for built-in tools. Chat Completions function calling only with reasoning_effort set to none. Cheapest current model whose page lists function calling and is not marked deprecated.
GPT-6 Astra (no confirmed API slug) $10 $50 The models page says to start with it. This chapter does not put a guessed slug in a command. Tool support was not re-opened on a model page.
gpt-6.1-sol $2 $10 Standard short context. Tool support was not re-opened on a model page.
grok-build-0.1, below 200k $1.00 $2.00 Cached input $0.20. At or above 200k the whole request is $2.00 / $0.40 cached / $4.00. Context 256k. Not in the batch-discount list.
grok-4.3, below 200k $1.25 $2.50 Cached input $0.20. At or above 200k: $2.50 / $0.40 / $5.00. Context 1M.
grok-4.7, below 200k $2.00 $6.00 Cached input $0.50. At or above 200k: $4.00 / $1.00 / $12.00. Context 500k.
grok-4.5, grok-4.6, below 200k $2.00 $6.00 Same input and output as grok-4.7 below 200k. grok-4.5 cached input is $0.30 below 200k.
qwen3.8-max, Singapore International, up to 1M $2 $6 The rows here are the International Singapore card. Global qwen3.8-max on the same page is $1.65 / $4.951.
qwen3.7-plus snapshot qwen3.7-plus-2026-05-26, up to 256K $0.40 $1.60 256K to 1M is $1.20 / $4.80. The live alias prints the same list numbers and adds “Limited-time 20% off.” A separate discounted dollar figure is not printed, so the discounted price is not stated.
qwen3.8-flash, up to 1M $0.15 $0.47
qwen3.7-flash, up to 32K $0.030 $0.130 32K to 256K: $0.100 / $0.400. 256K to 1M: $0.200 / $0.800.
qwen-flash, up to 256K $0.05 $0.40
qwen3.5-flash, up to 1M $0.10 $0.40
Kimi K3.1 Pro $2.00 $10.00 Cache write $2.00; cache hit $0.20. A distinct 1-hour write price did not render: UNVERIFIED.
Kimi K3.1 Pro Highspeed $4.00 $20.00 Cache write $4.00; cache hit $0.20.
Kimi K3 $3.00 $15.00 Cache write $3.00; cache hit $0.30. K2 numeric rates did not render: UNVERIFIED.
MiniMax-M3.1-Flash $0.15 $0.60 Cache read $0.015.
MiniMax-M3 standard, ≤512k input $0.30 $1.20 Cache read $0.06. The page labels M3 “Permanent 50% off” and strikes through $0.60 / $2.40 / $0.12. The table uses the lower figures.
MiniMax-M3 standard, >512k input $0.60 $2.40 Cache read $0.12. Struck-through pre-discount figures: $1.20 / $4.80 / $0.24.
Z.AI GLM-5.3-Flash $0.15 $0.50 Paid. The free model is GLM-4.7-Flash.
Cerebras qwen-3.8-27b $0.99 $1.49 Public token price beside the $5 / 30-day credit.

DeepSeek peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday–Friday, excluding Chinese public holidays. All other hours, including weekends and those holidays, are off-peak. Off-peak is half of peak. Legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are accepted and billed at the Flash price. The 2026-04-24 note says V4-Flash performs on par with V4-Pro on simple agent tasks.

Anthropic US-only inference is a 1.1x multiplier on the standard global rates. The hybrid below uses standard global Opus 5.5, not fast mode. On the Claude chat pricing page, the same token card calls Haiku the fastest, most cost-efficient model, Sonnet the high-performance model for coding and agents, and Opus the daily driver for agentic coding and enterprise work.

xAI bills a request at or above the long-context threshold entirely at the higher rate. Batch is 20% off for grok-4.7 and the other 4.x text models listed on the page. Grok 4.7 Fast is not on the public API. The public API price used here is the standard global rate.

The Qwen pricing page does not say which of these IDs support tool calls. Kimi’s docs page says K3-series context is 1,048,576 tokens, billing is pay-as-you-go with no subscription, and batch for kimi-k3.1-pro and kimi-k3 is 60% of the standard price. K2 is closed to new users from 2026-10-31 and planned to retire by the end of December 2026. China-site Moonshot rates in another currency were not opened: UNVERIFIED. MiniMax Priority is 1.5x standard.

OpenRouter bills inference at the provider’s list price and does not mark up individual requests. The fee is on credit purchases: Standard 5.5%, and the FAQ also states a $0.80 minimum on that fee; crypto 5%; Business 8% with no monthly minimum. BYOK is $25,000 of list-price inference per month with no OpenRouter fee on Standard and Business, then 5% of list price. Per-model token rates are whatever that provider posts and are not repeated as a second price.

What one routine call costs

For a 3,000-input / 800-output call, cache miss, standard tier, the opened rates that are both cheap and attached to an agent or tool-calling description are:

Model In / out per 1M Cost of one such call
qwen3.7-flash, request ≤32K $0.030 / $0.130 $0.000194
Claude Haiku 5.5, prompt ≤100k $0.10 / $0.50 $0.000700
deepseek-flash off-peak, cache miss $0.15 / $0.60 $0.000930
MiniMax-M3.1-Flash $0.15 / $0.60 $0.000930
qwen3.8-flash $0.15 / $0.47 $0.000826
gpt-6-luna, short context $0.10 / $0.50 $0.000700
gemini-3.1-flash-lite text $0.25 / $1.50 $0.001950

qwen3.7-flash is the lowest number. The Alibaba page does not describe it as a multi-step agent model. Groq’s models page does not label a free model and does not describe tool-calling quality. DeepSeek’s price card marks tool calls on deepseek-flash, and the 2026-04-24 note says that flash line performs on par with V4-Pro on simple agent tasks. That is the cheap model in the hybrid. The frontier model is Claude Opus 5.5 at $4 / $20. deepseek-flash peak is exactly double off-peak. The worked months use off-peak cache-miss. A month that fell entirely in peak hours is stated beside each total.

Hybrid: DeepSeek Flash off-peak and Claude Opus 5.5

Routine tool steps go to deepseek-flash (DeepSeek-V4.1-Flash-1017). Hard planning goes to Claude Opus 5.5, API id claude-opus-5-5. No cache hits, no batch, no OpenRouter fee inside the totals, global standard Opus (not fast mode). Each call is 3,000 input tokens and 800 output tokens. The 800 output tokens are the whole billed output. DeepSeek does not list a separate thinking SKU. Opus prompts here are far under 100,000 tokens, so no long-context multiplier applies.

Per call:

Light: about $17

8 tasks/day × 20 calls × 30 days = 4,800 calls. 10% frontier = 480 Opus calls and 4,320 Flash calls.

Same month entirely at Flash peak ($0.00186 per cheap call): $8.0352 + $13.44 = $21.4752, which rounds to $21 per month.

If all 4,800 calls were Opus: 4800 × 0.028 = $134.40, which rounds to $134. The 10% frontier share is most of the $17 bill. The Flash share is about $4.

Heavy: about $185

40 tasks/day × 20 calls × 30 days = 24,000 calls. 25% frontier = 6,000 Opus calls and 18,000 Flash calls.

Same month entirely at Flash peak: $33.48 + $168 = $201.48, which rounds to $201 per month.

If all 24,000 calls were Opus: 24000 × 0.028 = $672 exactly.

Direct list price only. Paying the same token bill through OpenRouter Standard adds the credit-purchase fee on top of these inference dollars: 5.5%, with a $0.80 minimum stated in the FAQ; crypto 5%; Business 8%. The notes do not compute a new total that includes that fee.

A $20 chat plan is not the agent runtime

Claude Pro is $20 billed monthly, or $17/month when $200 is billed annually. Claude Max is $100 and Claude Ultra is $200. Max includes $100/month of API credits and Ultra includes $200/month of API credits. Pro does not include that monthly API-credit line. The ChatGPT Plus help page says API usage is not included; Plus is $20, and ChatGPT Pro 100 / 200 / 500 is $100 / $200 / $500. A $20 chat plan is in the same ballpark as the light hybrid’s $17 API bill and does not buy this call volume on the API. The heavy hybrid’s $185 is near one month of ChatGPT Pro 200 or Claude Ultra and is still a different product: those plans meter chat usage. They do not replace pay-as-you-go tokens for an open-source agent. API tokens are billed separately.

Cost rows

The 1–5 scores are a judgment from the vendor wording in the 2026-11-14 notes. They are not a benchmark, a leaderboard, or a measured eval. Google is a 4 because the free tier stays in this chapter, the model to select is gemini-4-hydrogen, and Carbon and Argon stay off that tier. Groq is not a $0 row in the table below: the models page does not label a free model, and it does not describe tool-calling quality. Z.AI is a 3 because GLM-4.7-Flash documents function calling and the public allowance is not published. The light hybrid is a 4 because most calls are Flash, which DeepSeek says is on par with V4-Pro on simple agent tasks, with 10% on Opus. The heavy hybrid is a 5 because a quarter of its calls are Opus, which Anthropic’s price card calls the daily driver for agentic coding. Most heavy calls are still Flash.

Option Real monthly cost Credit card or hardware How smart it feels (1-5) Catch
Google AI Studio / Gemini Developer API, free tier, select gemini-4-hydrogen $0 No card. Hosted API. 4, judgment Privacy (prompts may train the provider)
Z.AI GLM-4.7-Flash $0 Card: not stated. Hosted API. 3, judgment Privacy policy skips the API. Rate limits: not published
Light hybrid: deepseek-flash (DeepSeek-V4.1-Flash-1017) off-peak + claude-opus-5-5 about $17 Pay-as-you-go API bill. Hosted API. 4, judgment Safe official. Peak month about $21
Heavy hybrid: deepseek-flash (DeepSeek-V4.1-Flash-1017) off-peak + claude-opus-5-5 about $185 Pay-as-you-go API bill. Hosted API. 5, judgment Safe official. Peak month about $201

The light and heavy dollar figures are the off-peak totals already rounded in the notes ($17.4576 to about $17, $184.74 to about $185). A month that fell entirely in DeepSeek peak hours is the notes’ side case: about $21 light, about $201 heavy. The OpenRouter credit-purchase fee is an add-on on top of those inference dollars and is not included in the table.

Sources

Vendor pages already cited in the 2026-11-14 notes. Checked or opened that day.