DeepSeek V4.1 Flash joins the free tier, and 95% of tokens are cache hits

The free tier now runs two models instead of one. Alongside GLM 5.3 Flash, every free key can call DeepSeek V4.1 Flash, on its own dedicated endpoint, with the same $10 of credit. Existing keys already have it, with nothing to re-claim.

Why a second model rather than a bigger grant

The single most common reason a free key looks broken is that the client is pointed at a model the key cannot use. A coding agent arrives with a model id already in its config, sends one request, gets a 403, and the person concludes the service does not work. Adding a second model that agents actually reach for removes a large share of that.

It also buys headroom. Each model has its own concurrency pool, so two models means bursts from different users stop competing with each other. On the busiest day so far the free tier served 2,378 requests and 380 million tokens through a single model. Two pools make that comfortable rather than tight.

What DeepSeek V4.1 Flash brings

• A 1M token context window, the same as GLM 5.3 Flash.
• Vision. It reads images, which the GLM route does not. Useful for screenshots, diagrams and anything a browser-driving agent picks up.
• Tool calling and structured output, both verified end to end before it went live.
• Prompt caching that actually reports itself, which matters more than it sounds. We checked that the endpoint returns cached token counts before listing it, because a cache you cannot measure is a cache you cannot bill honestly.

95% of all input tokens are cache hits

Across the whole life of the free tier, now 28,108 requests and 3.14 billion tokens from 72 accounts, 95% of input tokens have been served from cache. On recent traffic it runs higher still, above 98%.

This is the number that decides what agent work costs, and it is the reason a $10 grant goes as far as it does. An agent re-sends its whole conversation on every tool call, so the same prefix is billed again and again. If most of that is served from cache, the cache-hit price is effectively the only price that matters. We wrote up the arithmetic separately, and the free tier has now tested it at scale on traffic nobody curated.

What that means in practice: one account spent its entire $10 and got 284 million tokens for it, at a 99.3% cache-hit rate. That is not a projection, it is one person's finished balance.

What you get

• $10 of credit, no card, on a new account.
• Two models, glm-5.3-flash-modal and deepseek-v4.1-flash-modal, on one key.
• The same per-token rates paying customers get. It is a preview of the real service, not a throttled trial.
• $4 of usage per rolling 24 hours and 60 requests per minute, against no daily cap and 300 per minute on a credit pack.
• Cost visible per request. Every response carries x-cost-usd and x-input-cache-hit-rate, so you can check what a session cost and why, rather than taking a price list on trust. The API documentation lists both.

One thing to set before you start

Your key works with those two model ids and nothing else. Point a client at any other model and every request returns 403 model_not_allowed_for_key. The error now names the models you can use, but it is easier to set it correctly first. Buying a credit pack unlocks the whole catalogue on the same key, no new credentials.

Claim the free credit here, or read the numbers from the first round if you want to see how the last one went before signing up.