Claude got metered on June 15. Here's what AI is actually free — and what isn't.
After June 15, 2026 you don't need a bigger AI budget — you need the right free tier for interactive work, a couple of genuinely-claimable signup credits for bursts, and a local model for everything you run on a loop. Here's the verified list, with eligibility stated plainly, and the math.
- The June 15 change only meters programmatic / agent Claude usage. Interactive chat and terminal use are unaffected.
- Real free tiers cover a lot at $0 — Google's Gemini, Groq, Cohere — with rate limits.
- A few signup credits are genuinely claimable by anyone: Anthropic's ~$5 console trial, OpenRouter's $1 + free models. The headline "$1,000 free" numbers usually aren't.
- For anything you run on a loop, a local model (Ollama / MLX) is the only option that's truly $0 — no signup, no eligibility, no expiry.
- The expensive mistake is leaving an automated loop pointed at metered API rates by default.
What changed with Claude's pricing on June 15, 2026?
On June 15, 2026, Anthropic moved programmatic Claude usage — the Agent SDK, claude -p, Claude Code in GitHub Actions, and third-party agents — off the shared subscription pool and onto a separate monthly credit billed at standard API rates. Interactive chat and terminal use are unaffected.
- Monthly agent credit by plan: $20 on Pro, $100 on Max 5x, $200 on Max 20x.
- Credits are per-user, refresh monthly, and don't roll over; once exhausted, overages bill at standard API rates.
- Watch for Anthropic's claim email and claim your credit once through your account before the cycle starts.
For the full mechanics and the cost math behind running loops after the change, see our deep dive: the real cost of an autoresearch loop after June 15. (Change confirmed via InfoWorld and Anthropic's help center; verified June 2026.)
Do I actually have to pay more after June 15?
No — not for most solo workloads. What you pay comes down to one question: is a human watching the loop? That splits every job into three buckets, each with a different cheapest-honest answer.
| Your workload | Cheapest honest option | Cost |
|---|---|---|
| Interactive (you review each step) | Your existing Claude / Cursor subscription, or a provider free tier | ~$20/mo or $0 |
| Occasional bursts (a one-off bigger job) | A genuinely claimable signup credit | $0 until it runs out |
| Anything on a loop (runs unattended) | A local model (Ollama / MLX) | $0, always |
What AI is genuinely free after June 15?
Two categories: provider free tiers (great for interactive and burst work, always rate-limited) and local models (the only thing that's free for an unattended loop). Every figure below was checked against the provider's own page on June 4, 2026.
Free tiers you can use today
| Provider | What you get | The catch |
|---|---|---|
| Ollama / Apple MLX (local) | Run open models on your own machine, unlimited | Your hardware + electricity; it's local, not a hosted API |
| Google AI Studio (Gemini API) | Free tier — e.g. Gemini 2.5 Flash around 1,500 requests/day | Rate-limited; free-tier inputs may be used to improve Google's products |
| Groq | Free tier, no credit card; generous requests/day | Rate limits; you're limited to the models Groq hosts |
| Cohere | Free trial tier, ~1,000 calls/month | Trial-tier limits; instant trial keys |
| Mistral (La Plateforme) | "Experiment" tier, large monthly token budget | Prototype-only by terms — not for production |
Signup credits anyone can claim
These are small but real — claim-it-yourself, no company required. Use them to test, not to power a sustained loop.
| Program | Credit | Eligibility | The catch |
|---|---|---|---|
| Anthropic Console | ~$5 trial, all Claude models | New API account, phone verification, no card | Small — for testing, not loops |
| OpenRouter | $1 + access to free models | New account, no card, never expires | The $1 is tiny; the free models are the real value |
| Google Cloud | $300 welcome credit | New GCP customers only (never paid before) | Expires in 90 days |
| Microsoft Azure | $200 credit | New Azure accounts | Expires in 30 days |
Local models (Ollama, Apple MLX) are the only option with no eligibility and no expiry. For non-training work — prompt, copy, SEO, and headline optimization — both the agent and the evaluator run on your own machine at $0 per token. That's the answer for anything you run on a loop.
A 12-point check, one honest score — sent straight to your inbox. No card.
Take the free assessmentAre the "$1,000 free AI credits" deals real?
Sometimes — but they're almost always gated, and rarely the full number. The credits you see advertised as "$1,000 free on OpenRouter" or "$1,000 in Anthropic credits" are partner or startup-program offers, not something you can claim by signing up. The ungated, claim-it-yourself versions are far smaller: OpenRouter's public signup credit is $1; Anthropic's is about $5.
Why the big numbers are a ceiling, not a check
Most "$X,000 free" figures come from startup and accelerator programs — AWS Activate (up to ~$100K), Google Cloud for Startups (up to ~$350K), Microsoft for Startups (up to ~$150K), the Anthropic startup program ($1,000+). To qualify you typically need a registered company, often a custom company email domain (generic Gmail is rejected), and sometimes accelerator or VC backing and "new, never-funded" status. The dollar figure quoted is the maximum, not what a typical account receives. A solo operator on a Gmail address usually doesn't qualify for any of it.
You'll also see these same credits bundled on "startup perks" deal portals. Those add a second locked door — you have to be a member of the portal on top of meeting the provider's own eligibility. Two gates, not one. None of it belongs in your plan as money you can count on.
How to tell a claimable credit from a marketing number
- 1Is it on the provider's own site — not a deals or perks portal?
- 2Can you meet the eligibility without joining a program or owning a company domain?
- 3Is the number a guarantee, or an "up to"?
If you can't answer yes to all three, treat it as a ceiling, not cash.
How much does it really cost to run an AI loop after June 15?
If you keep an overnight loop on Claude's metered API, a 100-experiment session runs roughly $15 at Opus-class rates (~$5 / $25 per million tokens) — about $450/month run nightly. Prompt caching cuts that to about $6 a session (~$180/month). Run the same non-training loop on a local model and it's $0.
| Setup | Per session | Per month (nightly) |
|---|---|---|
| Cloud API (Opus-class, no caching) | ~$15 | ~$450 |
| Cloud API + prompt caching | ~$6 | ~$180 |
| Local model (Ollama / MLX) | $0 | $0 |
Prompt caching is the lever on the cloud side: cache reads bill at about 10% of the input price, and a loop reuses the same large context (your strategy file, baseline data, rules) every iteration — close to the ideal case for caching.
One honest caveat: "$0 local" applies to non-training loops — prompt, copy, SEO, and headline work. If your loop trains a model, you still need a GPU, local model or not. Figures are approximate and depend on your token volume per cycle. You can run the cost math yourself in the economics breakdown.
How do I run AI A/B tests cheaply without a developer?
You point an AI coding agent — Claude Code or Cursor — at a plain-English instruction file that says: propose a change, measure it against a scorecard, keep it if it beats the baseline, revert it if it doesn't, then repeat. That keep-or-revert loop is the whole method, and on a local model it costs nothing to run.
Knowing local is the answer is step one. The part that takes work is writing the loop so it runs safely — a hard max_cost_usd stop so it can never bill you by accident, the right scorecard, the prompt-cache structuring. That's exactly what the Autoresearch Playbook is: 12 ready-to-run templates that already encode it, plus the guide that walks through running them on free credits and local models.
Which option should you pick?
| If you are… | Start with | Why |
|---|---|---|
| Just trying it out | Anthropic $5 + OpenRouter free models, or a Gemini free key | $0, no commitment |
| Running occasional bigger jobs | GCP $300 / Azure $200 new-customer credit | Real headroom, time-boxed |
| Running loops unattended / nightly | A local model (Ollama / MLX) | The only true $0 — no metered surprises |
| A funded startup with a company domain | Apply to AWS Activate / Anthropic startup program | Where the big credits actually live |
Frequently asked questions
Is the $1,000 free AI credits deal real?
Usually it's gated, and rarely the full number. Advertised credits like "$1,000 free on OpenRouter" or "$1,000 in Anthropic credits" are partner or startup-program offers — not something you can claim by signing up. The ungated, claim-it-yourself versions are much smaller: OpenRouter's public signup credit is about $1 plus free models, and Anthropic's new-account console trial is about $5.
What is the cheapest way to run an AI loop after June 15?
A local model via Ollama or Apple MLX. For non-training work — prompt, copy, SEO, and headline optimization — both the agent and the evaluator can run on your own machine at $0 per token, with no signup, no eligibility, and no expiry. It's the only option that's genuinely free for an unattended loop.
Do I have to pay for Claude Code after June 15, 2026?
Not for interactive use. Running Claude Code in your terminal and reviewing each step still draws from your normal subscription. Only programmatic, scripted usage — the Agent SDK, claude -p, and Claude Code in CI — moved onto a separate monthly credit (about $20 on Pro, $100 on Max 5x, $200 on Max 20x), after which it bills at standard API rates.
Can I A/B test with AI without a developer?
Yes. You point an AI coding agent such as Claude Code or Cursor at a plain-English instruction file: propose a change, measure it against a scorecard, keep it if it beats the baseline, revert it if it doesn't, then repeat. That keep-or-revert loop is the whole method, and on a local model it costs nothing to run.
Does Anthropic give free API credits?
New Anthropic console accounts receive roughly $5 in free trial credit that works across Claude models; you verify a phone number and no card is required. Larger amounts ($1,000 and up) come through the Anthropic startup program, AWS Activate, or accelerator partnerships, all of which have eligibility requirements.