Free Tiers That Changed in 2026: Six Verified Pricing Shifts

Free tiers have a half-life, and this year it shortened. Six changes we verified against providers' own pages between July and September 2026 show the pattern clearly: sunsets get dates, starting credits get smaller, product categories get redefined, and "free" increasingly means "free until the promotion ends."

If you built on any of these, the change is not abstract — it is a broken script, a shrinking budget line, or a deadline on your roadmap. Here is what changed, what the official source actually says, and what to do about it.

The six changes at a glance

Change	Official source	Status
GitHub Models fully retired	GitHub Models documentation↗	Completed July 30, 2026
Perplexity Sonar API sunset	Perplexity Sonar quickstart↗	Dated September 27, 2026
Cerebras starting credit now $5	Cerebras pricing↗, Cerebras rate limits↗	Current as of 2026-09-17
OpenAI promo pricing with a date	OpenAI API pricing↗	Runs through November 21, 2026
Lambda repositioned as GPU cloud	Lambda GPU pricing↗	Structural, no end date
Amazon Q Developer IDE plugin end of support	Amazon Q Developer↗	Dated April 30, 2027

1. GitHub Models — fully retired on July 30, 2026

The biggest change is that the product is gone. GitHub's documentation states plainly that as of July 30, 2026, GitHub Models has been fully retired, and the retirement changelog↗ confirms the playground, model catalogue, inference API, and bring-your-own-key endpoints are no longer available.

What it means. The free, meterless access to a catalogue of open models under a GitHub account no longer exists as a product. The model-access surface inside GitHub is now Copilot, and the free tier is Copilot's free tier, which is measured in completions and chat requests per month rather than in models.

What to do:

1. Grep your repositories for the GitHub Models endpoint host and any workflow secrets that referenced it. Lingering references fail silently in scheduled jobs, which is worse than failing loudly.
2. Replace model-access calls with either the Copilot API surface or a standalone provider for the open models you were getting free there.
3. Update your dependency list and any onboarding docs — a stale tutorial is how the next developer loses an afternoon.

Lesson: a free product can be retired, not merely reduced. Watch the vendor's documentation, not its marketing page.

2. Perplexity Sonar API — sunsets September 27, 2026

Perplexity's original Sonar model family stops being supported on September 27, 2026, according to the provider's own Sonar documentation↗, and the changelog↗ records the deprecation history of related models. The provider ships a migration guide to its newer agent-style API, which routes third-party models at direct provider rates with no markup.

What it means. Any integration still pinned to sonar model IDs needs migration work against a hard deadline that is now weeks away, not months. The economics change as well: per-request search pricing in the range of a few dollars per thousand requests gives way to per-invocation tool pricing, where the web search call itself is billed individually (documented at $0.0025 per call). Workloads that make one request with multiple internal searches may see a different cost profile after the move, so model the new shape rather than assuming parity.

What to do:

1. Inventory every place a sonar model ID appears, including evaluation notebooks and prompt-regression tests that never reach production.
2. Run a parallel evaluation on the new API before you cut over — answer quality on the replacement route is a product question, not just a migration task.
3. Calendar the sunset date as a project milestone the day you read it, and give it an owner.

Lesson: when a vendor sunsets a model family, the sunset date is your deadline, and the migration guide is your work breakdown.

3. Cerebras — the starting credit is now $5, and it expires

Cerebras's current pricing page↗ advertises a self-serve developer tier with $5 of free credit to start. The rate-limits documentation↗ adds two details that matter even more than the amount: the credits arrive after adding a verified payment method, and they expire 30 days after they are granted. Older third-party pages, and figures that persisted in various comparison lists, still cite $30.

What it means. $5 is a speed demonstration, not a workload budget. The genuine value of the free allocation on this provider is that you get to measure inference speed on real models with your own prompts — the published throughput on its open-model offerings is the differentiator, not the credit balance. And the 30-day clock means the credit is a deadline-shaped asset: a grant you sit on is a grant you lose.

What to do: run your latency benchmark in the first week, capture the numbers in your own evaluation doc with the date, and treat the account as an evaluation environment rather than an ongoing free tier.

Lesson: when a credit figure conflicts between the vendor's page and everything else, the vendor's page wins — and the conflict itself is a signal that the terms moved.

4. OpenAI — promotional pricing with an expiration date

OpenAI's API pricing documentation↗ marks the discounted rate for GPT-5.6 Sol as "available at least through November 21, 2026," with the previous, higher rate shown struck through above it. This is unusually transparent: most vendors change prices silently, and most promos are undated.

What it means. Anything you budgeted at the promotional rate has a November deadline. A cost model that assumes the discount is permanent will overstate savings for every planning horizon that crosses that date — which includes most annual budgets signed in the second half of the year.

What to do:

1. Build durable forecasts at the struck-through list rate and treat the promo as headroom, not as the baseline.
2. Set a review reminder a few weeks before the date, when you still have room to renegotiate scope, pre-purchase credits, or shift workloads.
3. Where your architecture allows it, keep a second provider warm for the workloads that are price-sensitive, so a promo ending is a routing decision rather than an emergency.

Lesson: a dated promo is the most honest kind of pricing change. Use the date to set your own reminder.

5. Lambda — from per-token inference to hourly GPUs

Lambda's own pricing pages↗ now describe the company as a GPU cloud, selling H100, H200 and B200 capacity by the hour with no egress fees, rather than as a per-token inference API. The old per-token inference on-ramp that third-party summaries still describe is not the product anymore.

What it means. "Lambda" in a cost model now means hourly GPU capacity, not per-token inference. The unit of cost changed from tokens to node-hours, which is a different business with a different break-even: your utilisation rate, not your token spend, becomes the number that decides whether it is cheap. That is good news for sustained, GPU-hungry workloads and bad news for spiky, low-volume API use — an idle rented GPU bills whether or not it generates anything.

What to do: if your plan relied on a per-token free on-ramp from this provider, replace it with a provider that still sells tokens. If you were going to move to GPUs anyway, model utilisation honestly — the break-even against per-token pricing depends on how many hours a day you can keep the node busy.

Lesson: a vendor can change what it is while keeping its name. Re-read the product page, not only the pricing table.

6. Amazon Q Developer — the client has an end-of-support date

Amazon Q Developer's free tier is intact: AWS's product page↗ and the end-of-support announcement↗ both describe a free allowance of tens of agentic chat interactions per month plus a substantial code-transformation allowance, available perpetually. But the same pages carry an end-of-support notice: IDE plugin support ends April 30, 2027, with migration guidance to the newer surface.

What it means. The free tier is not going away, but the client you are using is on a clock. End of support means no security fixes, not just no new features, so the risk curve bends upward well before the date itself. Teams that standardised on an IDE plugin workflow have a twelve-month migration window that is already partly spent.

What to do: identify who owns the migration, run the newer surface in parallel on one team for a sprint, and set an internal deadline comfortably before April 2027 rather than treating the vendor's date as the deadline.

Lesson: end-of-support dates on tooling matter as much as free-tier changes on the service.

The pattern: none of these were surprises

All six changes share one property — they were stated in writing on an official page before they bit anyone. GitHub said it in its docs and shipped a changelog entry. Perplexity published the date and a migration guide. Cerebras printed the amount and the 30-day expiry in its own rate-limit docs. OpenAI dated the promo and showed the struck-through price. AWS gave an end-of-support date with a one-year runway.

The people who got hurt were not victims of a secret change. They were victims of a missed re-read: a figure remembered from a blog post, a tutorial copied from a forum, a credit balance assumed rather than checked. That is the most expensive kind of staleness, and it is entirely preventable.

A monthly re-verification routine you can actually run

1. List your dependencies with their source URL. For every provider you depend on, record the exact pricing or limits page you verified, plus the date. A dependency without a source URL is an unverified assumption.
2. Re-read the primary source monthly, not the summary. Ten minutes a month on the official page beats an hour of incident response later.
3. Convert every date you find into a task. A sunset is a deadline; a promo end is a forecast input; an end-of-support notice is a migration project; a credit expiry is a calendar entry with an owner.
4. Diff what you recorded against what the page says now. The diff is your change log, and over a quarter it becomes a genuinely useful artefact for budget reviews.
5. Re-run one small workload on the affected provider. A canary request tells you whether the documented change is already live in behaviour, which is not always the same day as the announcement.
6. Keep a fallback route configured. Providers change; a router that can move traffic means a change is a routing update rather than a rewrite.

Why the directory is built around dated verification

Every entry in our directory carries the verified state of a provider's free tier with the date we read it, and the ones that changed keep their change history rather than silently updating. We treat "what was true in March" as one of the most expensive sentences in AI budgeting, which is why a number without a date is not treated as verified at all.

How we verified this

Each change in this article was confirmed by reading the provider's own documentation between July and September 2026; the source page is linked inline for every claim, and this article was last checked on 2026-09-17. Where a change is a repositioning rather than a dated announcement — the GPU-cloud shift, for example — we describe what the official pages now say rather than asserting an announcement date we could not find. We have deliberately excluded rumoured changes, leaked roadmaps, and any figure that could only be sourced to a third-party roundup.

FAQ

Do these changes mean free AI APIs are ending? No. New free allocations appeared during the same period. What is ending is the assumption that a specific free tier is permanent: the pattern in 2026 is dated terms, smaller starting credits, and shorter-lived promotions rather than the disappearance of free access itself.

Which of these changes should I act on first? Anything with a date inside your current planning horizon. A sunset weeks away outranks a repositioning with no deadline, because migration work has to be scheduled before the deadline, not after it.

How do I tell a real change from an out-of-date blog post? Read the provider's own pricing or limits page and check for a date on it. Third-party summaries lag by months — the Cerebras credit figure is the clearest example, where older pages still cite a number the vendor's own page no longer shows.

Is a dated promotion a reason to avoid a provider? No, but it is a reason to model it correctly. Build forecasts at the permanent rate, use the promotional window for work you would otherwise defer, and keep a fallback ready for when the window closes.

What is the cheapest way to hedge against these changes? Multi-provider routing plus prompt caching. Caching reduces the cost of the input you repeat, and routing means a provider-level change degrades one route rather than breaking your application — see the reset-shape comparison for how to map workloads to more than one allowance.