Buying Tokens for Autonomous Agents: Who Sells API Credit, and What It Costs

Fourth in a set with Big Tech AI Subscriptions, If the Big Labs Block You and Chinese AI Services. Those three cover subscriptions — a human logging into a chat window. This one covers tokens: metered API credit, which is what an autonomous agent actually consumes. Prices as listed in early October 2026, USD per million tokens (input/output) unless noted. This tier moves monthly; verify before committing.

Why a subscription is the wrong product

A $20 subscription buys a seat for a person. An agent has no seat. It has a loop: read, call a tool, read the result, decide, call again — and every pass re-sends the accumulated context. Three things follow, and they matter more than the headline price.

Token amplification is the real bill. One user instruction becomes many model calls — thinking, searching, tool-calling, summarising — and you pay for every intermediate step. AWS documents this plainly for Bedrock Agents: there is no per-invocation fee, and that is precisely the trap, because the cost lands on the internal calls instead. Price the loop, not the prompt.

Cached input and batch discounts outrank the sticker price. An agent re-sends the same system prompt and the same conversation prefix on every turn, which is exactly the shape caching is built for. OpenAI reads cached input at 10% of standard rates and halves both directions on the Batch API; xAI charges $0.50 cached against $2.00 fresh on Grok 4.7; Moonshot charges $0.30 cache-hit against $3.00 cache-miss on Kimi K3 — a 10× difference on the input side alone. A loop that reuses its prefix pays a fraction of the table rate; one that rebuilds the prompt each turn pays all of it.

The spend tier binds before the price does. Every major vendor gates new accounts behind a monthly cap that an agent can hit in an afternoon. See the last section — this is the part most cost comparisons leave out, and the part that actually breaks things.

Big tech

The same eight companies as the subscriptions article, asked a different question: can you buy tokens, and at what rate?

The spread inside one vendor is wider than the spread between vendors. OpenAI’s own catalogue runs 25× on input between Luna and its top tiers; Amazon’s runs 70× between Micro and Premier. Choosing Google over OpenAI moves your bill by a fraction of what choosing Flash over Pro does. Pick the tier first, the vendor second.

Two smaller observations worth the ink. Microsoft is the odd one out — it sells a storefront rather than a model, which is a fine thing to buy and a bad thing to mistake for independence. And Meta’s Contributor tier is the only place in this table where the price is explicitly data rather than money; that is a clean trade if the workload is non-sensitive and a disqualifying one if it isn’t.

Alternatives from other countries

Same exclusions as the big-tech table above, and Chinese developers get their own section. What’s left is thinner than the subscription market — but not in the way you’d expect, because Cohere sells tokens perfectly well. It was excluded from the alternatives article for having no self-serve consumer subscription, and that disqualification is simply irrelevant here: there is a published per-token price and a card form.

Where I could not find a token to buy. Japan has a real model ecosystem and no self-serve API shop I could confirm a price for — Preferred Networks sells PLaMo access on enquiry, Sakana is geo-restricted. Israel’s AI21 halted Jamba development in May 2026, so the obvious entry is gone. For the UAE’s Falcon line I found downloadable weights but no first-party per-token price I could verify; treat it as self-hosting. In each case the honest statement is “not verified”, not “doesn’t exist” — but if you can’t find the price page, you can’t plan a budget around it either.

The pattern underneath: outside the US and China, the market sells you jurisdiction more than capability. Mistral is the one entry competing on both. Everything else in that table is either a compliance posture, a language specialism, or a European address wrapped around somebody else’s open weights — all legitimate reasons to buy, none of them a capability argument.

Chinese models

This is where the cheap tokens are, and it isn’t close.

Three things to weigh against those numbers, in descending order of how often they actually matter.

1. Latency compounds in a loop. China-hosted endpoints add roughly 100–200 ms per call from Europe or the US. On a chat turn that’s invisible. On a forty-step agent trace it’s four to eight seconds of pure wire time, before the model thinks about anything.
2. Jurisdiction is the same question as always. Prompts and tool outputs land with a company subject to PRC law. An agent is worse than a chat window here, not better: it sends whatever it happens to read, including files you never consciously pasted.
3. Sources disagree more than for Western vendors. Regional pricing gates mean third-party trackers and the in-console figure often differ — the GLM row above is a live example, and DeepSeek V4 Pro is quoted anywhere between $0.44 and $1.32 depending on who you ask.

And the standing escape hatch from the Chinese services article still applies, more strongly for agents than for chat: the weights are downloadable. Running DeepSeek or GLM on your own hardware removes points 1 and 2 completely, and an agent’s token volume is exactly the profile where owning the hardware starts to beat renting the tokens. See the multi-GPU survey.

Gateways: the fourth place to buy

Most agent frameworks don’t call a vendor directly — they call a gateway, and for failover that’s the right instinct, since a single account outage stops an autonomous run cold.

• OpenRouter passes through vendor pricing and charges 5.5% on credit purchases, routing to Fireworks, Together, Groq and Cerebras underneath. You pay a margin for one API key and one bill across every vendor in this article.
• Groq and Cerebras publish per-token rates with no purchase fee and run 600–750 tokens/sec decode on 70B-class models — often 2–3× Together’s price for the privilege. On Llama 3.1 8B: Groq $0.05 / $0.08, Cerebras $0.10 / $0.10, Together and Fireworks $0.18 both directions.
• Fireworks positions explicitly for agentic workloads — structured output, function calling, JSON mode — which is a product claim rather than a price one, but the right claim for this use case.

For an agent, the speed difference is not cosmetic. A forty-step loop at 700 tokens/sec finishes while a 70-tokens/sec endpoint is still on step six.

What actually stops an agent

Not the price. The cap. Every major vendor gates new accounts, and an unattended loop reaches the ceiling far faster than a human ever would.

Three practical consequences:

• Credits expire after a year at all three, and the expiry can’t be extended. Prepaying a large balance to secure a rate is a bet on your own burn rate.
• Auto-recharge is on by default at OpenAI, with a $5 minimum top-up. An agent in a retry loop plus auto-recharge is an unbounded bill — set the optional monthly cap the same day you create the key.
• Google is the only one of the three offering postpay, which sidesteps both the prepayment and the expiry.

My take

Buy on tier, not on brand. The honest conclusion from the big-tech table is that the vendor choice barely moves the bill compared to the tier choice. A sensible agent runs a cheap workhorse for the ninety percent of steps that are routine — Gemini Flash-Lite at $0.10, Qwen3.5 Flash at $0.10, GPT-5.6 Luna at $0.20 — and escalates to a frontier model only for the steps that need judgement. Paying Astra or Fable 5 rates for “read this file and tell me the next step” is the single most expensive habit in agent engineering.

Then make the prefix cacheable. The cached-input discounts in these tables range from 4× to 10×, and an agent loop is the one workload shaped to collect them in full. That is a larger saving than any vendor switch on this page, and it is a code change rather than a procurement decision.

Hold two accounts. Not for price arbitrage — for the run that dies at 3am because one tier cap was reached or one provider refused one prompt. A gateway account at 5.5% is a reasonable insurance premium against a stalled autonomous job; a second direct account is cheaper and more work.

On the cheap Chinese tokens: three cents a million is real, and so is the jurisdiction. For public data, scraping, summarising, classification — the bulk of what agents actually do — the price is hard to argue with. For anything touching client data, the answer that keeps working is the one from the companion article: take the weights and run them yourself. An agent’s volume is exactly what makes that arithmetic close.

The thread running through all four of these articles: the subscription market is crowded and the token market is concentrated. Eleven of the sixteen entries above are American or Chinese, and the non-US alternatives are mostly selling a jurisdiction rather than a capability. If you want genuine independence for an autonomous agent — the thing that keeps running when an account, a country or a policy changes — the only complete answer is still holding the weights.

Further reading

• OpenAI API pricing, every model · Gemini API pricing and the January increase · Mistral API pricing
• Anthropic pricing — primary source
• Amazon Bedrock pricing and token amplification
• API credits, tiers and expiry compared · Cross-vendor token price table
• Asian and European LLM APIs for production · Inference API price comparison
• Chinese AI Services — the subscription side of the same labs
• Self-Hosting the Frontier — the alternative to buying tokens at all