OpenAI API Review 2026

How OpenAI’s platform is put together

The OpenAI API is the pay-per-token door to the GPT model families, run by OpenAI out of San Francisco. In late 2026 the developer documentation lives at developers.openai.com while the dashboard, billing and keys stay at platform.openai.com. There’s a lot on the menu: the GPT-6 and GPT-6.1 flagships, the GPT-5.6 and GPT-5.4 lines, o-series reasoning models, image, realtime audio, transcription and embedding models, plus hosted tools such as web search at $10 per 1,000 calls.

Billing is organization-based. An organization holds projects, projects hold API keys, and both usage limits and spend limits can be set at either level. That structure matters for credits, because the ChatGPT subscriber grant attaches to exactly one organization and only keys inside it draw on the credit. Most of the friction we see in forum threads comes from people creating a key in the wrong organization.

What free credit exists and what doesn’t

There’s no sign-up credit to find. OpenAI’s pricing, rate limits and data pages don’t mention one, and the quickstart offers a single free test request followed by a Go to billing button. The $5 and $18 grants that still circulate in third-party guides are history. Budget a $5 purchase as the real entry fee; it’s also the threshold that moves your organization from Free to Build.

The live credit program is the API credits benefit for ChatGPT subscribers, added on October 28, 2026. Pro subscribers receive $100, $200 or $500 a month matching their tier price. Business workspaces receive $20 per paid Standard seat and $100 per Premium seat, pooled and capped at $500 per workspace per month. The subscription must be paid and at least seven days old, the same email must own both accounts, and the first grant is prorated.

Grants are spent before purchased credits and expire at the end of each billing cycle without rolling over. They cover the Responses, Agents, Decisions, GPT-Live, Realtime and Images APIs and the built-in tools, but not Codex usage through API keys. Cancelling doesn’t claw back credits already issued; it simply stops new ones.

What the Free tier really is

OpenAI’s Free tier is a usage ceiling, not an allowance. Its only qualification is that you’re in an allowed geography, and it carries a $100 a month usage limit with no published per-model rate limits. What’s actually free on the platform is small: the omni-moderation-latest model, 1 GB of file search storage, and 1 GB of ChatKit storage per account per month. Everything else costs money from the first token.

Since October 6, 2026 the paid tiers are Build, Launch and Grow, down from five. Build requires $5 in total credit purchases and allows $500 of usage a month, Launch requires $100 for $5,000, and Grow requires $500 for $200,000. On Build the Astra, Sol and Terra families get 5,000 requests and 1 million tokens per minute, with Luna at 2 million. Grow lifts that to 15,000 requests and 40 million tokens, and 30,000 requests with 180 million tokens on Luna. Upgrades are automatic once purchases cross each threshold.

What it costs per million tokens

The current spread runs from gpt-6-luna at $0.10 input and $0.50 output per million tokens, through gpt-6.1-terra at $1 and $5 and gpt-6.1-sol at $2 and $10, to gpt-6-astra at $10 and $50. The GPT-5.6 line sits between, with Sol at $4 and $20 on promotional pricing that holds at least through November 21, 2026. Standard rates apply up to 272K input tokens, and long-context pricing kicks in above that.

Three modifiers change those numbers. The Batch API charges roughly half, so gpt-5.4 falls from $2.50 and $15 to $1.25 and $7.50. Flex processing matches Batch rates on synchronous calls at lower priority. Fast mode, which replaced Priority processing on July 30, 2026, costs about twice standard. Cached input runs at 10% of the input rate on GPT-5.6 and later, 5% on GPT-6.1 Sol and Terra, with cache writes at 1.25x for the default 30-minute lifetime or 2.5x for the 12-hour option added on October 20, 2026.

Billing is prepaid. Purchases start at $5, $10 is the default amount, and purchased credits expire after one year. The dashboard has supported hard monthly spend limits per organization and project since July 22, 2026, and a Prompt Caching dashboard since August 20 shows hit rates and write-read splits so you can see what caching is actually saving you.

Credits for startups and researchers

OpenAI for Startups is VC-gated. Companies whose investor is in OpenAI’s network can receive free API credits, rate limit upgrades and solutions engineer time, and the application needs a referral code from that investor along with a product description, funding details and your org ID. No credit amount is published. If your VC isn’t a partner, there’s no self-serve route.

The Researcher Access Program is the one open door. The application form offers up to $1,000 in API credits valid for 12 months toward any publicly available model. Credits are tied to the account and institution named in the application, can’t be transferred, and expire without reinstatement. There’s no dedicated API education credit; students get pointed to Codex community programs rather than platform grants.

The mode matters more than the model

For cost control the model choice matters less than the mode. The same gpt-6.1-sol request costs $2 and $10 at standard, $1 and $5 on Flex or Batch, and $12 and $60 on the Ultrafast tier listed for Astra and Sol. That’s a 12x spread on identical tokens. Decide per workload: batch for anything that can wait, Flex for background agents, standard for users waiting on a screen, Fast only where latency is measured.

Caching is now explicit on GPT-5.6 and later. You can place cache breakpoints yourself, the minimum cacheable prefix is 1,024 visible tokens, and Prompt Cache Diagnostics has been generally available since September 8, 2026. Earlier models such as gpt-4.1 and gpt-4o still cache implicitly but at worse rates, 25% and 50% of input respectively, which is one more reason to migrate.

The 429 that isn’t a rate limit

The most common failure is a 429 that isn’t a rate limit. OpenAI enforces an approved monthly usage limit per tier in addition to any spend limit you set, and both stop requests. The dashboard path is Settings, Organization, Limits. A second trap is the credit_balance_exhausted error, which fires when prepaid credit and grants are both gone. Auto recharge prevents it, but set the monthly recharge cap or a runaway agent can keep buying.

Data handling is conservative by default. API content isn’t used for training unless you opt in, abuse monitoring logs are retained up to 30 days, and Zero Data Retention or Modified Abuse Monitoring require prior approval. Support for Free and Build organizations is the help center and community forum; there’s no published response time, so check the status page before you escalate.

Who should pay for it

Yes for teams that will spend real money, and a 4.0 rather than higher because of how the first dollar works. No trial credit, a $500 cap on Build and VC-only startup credits make the first week pricier than Google Cloud or AWS. Once you’re past that, the combination of a $500 subscriber credit, Batch and Flex pricing, explicit caching and per-project spend limits gives more cost control than most competitors. If you already pay for ChatGPT Pro, link the API credit today.