Andy Garcia · freelance dev
FRENES
Back to the blog
AI & Pricing

Anthropic Is Giving Startups a Free Year of Claude Team and $1,000 in API Credits. The Credits Expire Six Months Earlier.

October 7, 2026 6 min read

On Tuesday, October 6, during SF Tech Week, Anthropic expanded its Claude for Startups program. Approved startups get a free year of Claude Team with up to five Premium seats, $1,000 in Claude API credits, and up to $45,000 in offers from partner companies. Eligibility is broad: founded in the last five years, or funded in the last two. Bootstrapped companies can apply. Startups backed by a VC in Anthropic's partner network can get up to $100,000 more in API credits through that VC.

TechCrunch's URL for the story still says "enterprise service". Its headline and the program page say Claude Team, the group plan. That difference matters, and it is not the only line worth reading twice.

I build a LegalTech product on Claude and ship side projects on cheap API tiers, so I read the program page, the official terms and the addendum that took effect the same day. Here is what the offer is worth, what it ties you to, and what I would set up while it is free.

The seats are the real gift

A Premium seat on Claude Team costs $100 per month billed annually, or $125 billed monthly. Five seats for twelve months is 5 × $100 × 12 = $6,000 at the annual price, $7,500 at the monthly one. The API credits are $1,000. The plan is worth six times the credits.

So the offer is mostly a year of Claude, Claude Code included, for a small team, plus a smaller budget for the thing you are actually building. One condition: only organizations new to Team qualify.

What $1,000 buys on the API

Current list prices, per million tokens: Claude Sonnet 5.5 is $2 input and $10 output. Claude Opus 5.5 is $4 and $20. Claude Haiku 4.5 is $1 and $5.

On Sonnet 5.5, $1,000 is 500 million input tokens, or 100 million output tokens. Nobody spends it like that, so take a typical product call: 4,000 tokens in, 500 out.

On Sonnet 5.5, that is 4,000 × $2 / 1M = $0.008, plus 500 × $10 / 1M = $0.005. Total $0.013 per call, so about 76,900 calls for $1,000. The same call costs $0.026 on Opus 5.5 (about 38,500 calls) and $0.0065 on Haiku 4.5 (about 153,800).

Now put the 3,000 tokens of stable instructions in the prompt cache. Cache reads on Sonnet 5.5 cost $0.20 per million, 10 percent of the input price. The call becomes 3,000 × $0.20 / 1M = $0.0006, plus 1,000 fresh tokens at $0.002, plus the same $0.005 of output: $0.0076, about 131,600 calls, ignoring the occasional cache write. Same credits, about 70 percent more calls.

Two things fall out of that arithmetic. Once the cache does its job, output is about two thirds of the cost of the call, so response length becomes the next lever. And the credits expire six months after they are granted. 76,900 calls over six months is about 420 a day. That is a prototype with a few real users, not a launch.

Two clocks that do not line up

The Team year runs twelve months. The credits run six. From month seven to month twelve, your seats are still free and your API traffic is billed at list price. If your product found users in that window, the first real API invoice lands while the team still feels like everything is free.

At month twelve, the program page does not say what happens to the seats. The general pricing FAQ says that to avoid the next charge on a plan, you cancel at least 24 hours before the renewal date. I would put both dates in a calendar the day I claim the offer.

OpenAI's startup page, for what it is worth, is explicit about its own credits, which it hands out through partner VCs: when they run out, "you'll automatically switch to standard billing." That is the default pattern in this industry. Plan as if it applies.

What the credits tie you to

The credits work only on the first-party Claude API through the Claude Console. Not on Amazon Bedrock, not on Google Cloud Vertex AI, not on other third-party platforms.

For many startups, that is fine. For mine, it is the main constraint. In the LegalTech I build, where a model runs is a clause in the client contract. The first-party API's inference_geo parameter accepts two values: "global", the default, and "us", at 1.1 times the price. Workspace geo, where data is stored at rest, currently offers only "us". If a client needs a guarantee about another region, the documented route is a regional endpoint on Bedrock or Google Cloud, at a 10 percent premium. The credits do not follow you there.

For a European startup with sensitive data, then, the $1,000 is a prototyping budget on a path you may not ship on. Still useful. Just do not architect around the free path.

Team is also not Enterprise. On Anthropic's plan comparison, audit logs, custom data retention controls and the Compliance API are marked "No" for Team. If your clients ask for those, the free year does not answer them.

A few smaller lines. Anthropic may use your name and logo to identify you as a program member and customer. The program can be modified or cancelled at any time. Whether $1,000 of credits is taxable where you are is your problem, the terms say. The partner offers come from third parties, and the addendum treats them as having no attributable value for liability purposes. The Applied AI office hours carry a residuals clause: Anthropic may use general know-how retained in the unaided memory of staff who saw your confidential information, though not the information itself. I would bring architecture questions to those sessions, not client documents.

What I would set up on Monday

None of this is a reason to skip the offer. Six thousand dollars of seats is real. But the free months are the cheapest time to build the habits that make the paid months predictable.

Set a spend limit in the Console, under Settings and Billing, and one per workspace if you split environments. Then handle the two errors. Your own limit returns HTTP 400. Your tier's monthly cap returns HTTP 429 with the error code enforced_spend_limit_reached and no retry-after header, and the SDK's automatic retries keep failing until the first of the next month, unless you get a higher limit sooner. Treat that one as an outage, not a hiccup.

Track usage from day one with the Usage and Cost Admin API. It reports uncached input, cached input, cache creation and output tokens per minute, hour or day, grouped by model, API key or workspace. It is not available on individual accounts, so set up an organization first. One key per feature makes the report readable.

Design for the cache now: stable instructions first, variable content last. A five-minute cache write costs 1.25 times the input price and pays for itself on the first read.

Pick the model per task and measure cost per completed task on your own traffic. On my side projects, yes or no decisions go to a small judge model, not a frontier one. For anything that can wait, the Batch API halves input and output prices.

And keep one thin wrapper around your model calls, with provider and model as configuration. The day a client asks where inference runs, moving from the first-party API to a cloud provider should be a config change, not a rewrite.

The offer is generous where it is visible, the seats, and tight where it is not, the credits. Read it as six months of measurement paid for by Anthropic. Then the bill after it is a number you already know.

Sources

A project like this one?

I design and deploy products like this. Let's talk.

Let's talk