DeepSeek V4 is live: 1M-token context, 40% off list

Starting today the gateway serves the two official DeepSeek V4 releases: deepseek-v4-flash and deepseek-v4-pro. Both carry a 1,048,576-token context window — four times the 262K ceiling of our existing lineup — and both are in the launch promotion: 40% off list price, applied automatically.

What one million tokens actually buys you

A 1M window fits an entire mid-size repository, a year of support tickets, or a full legal discovery bundle in a single request — no chunking pipeline, no retrieval glue, no "which half do I drop" logic. You send the whole thing; the model sees the whole thing. We never truncate or rewrite your messages, so what fits, fits honestly.

The window counts input plus your max_tokens budget, and the limit is exactly 1,048,576 — oversized requests fail fast with context_length_exceeded before any generation starts, billed at $0. The context docs cover counting, the free /v1/count_tokens estimator, and per-language error bars we measured for the DeepSeek tokenizer.

Measured, because we do not quote marketing numbers

• `deepseek-v4-flash` is genuinely fast: ~1.0 s to first token (p50) and ~102 tokens/second streaming in our tests — the fastest throughput in the catalog, at a workhorse price.
• Prefill at scale is real work: ~45 ms per 1,000 input tokens on flash — a 100K-token prompt answered its first token in 4.6 s, 300K in 13.2 s.
• `deepseek-v4-pro` trades latency for depth: ~2.5 s to first token on short prompts, ~44 tok/s, and 76.5 s to first token on a real 600K-token prompt. Stream anything large.
• Function calling verified: in our contract tests, both models call tools with correct schemas — the agentic loop works, not just the spec sheet.

Your data stays on Cloudflare infrastructure

These are the official open-weight releases served on Cloudflare Workers AI — the upstream IDs @cf/deepseek-ai/deepseek-v4-flash-0731 and @cf/deepseek-ai/deepseek-v4-pro-0813 are public on every response. Your prompts are processed on that infrastructure and are not sent to DeepSeek’s servers; on our side we log request metadata only — tokens, latency, cost — never content.

Pricing: 40% off list, live on the models page

During the launch period deepseek-v4-flash bills at $0.264 / $0.792 per 1M tokens (input/output, list $0.44 / $1.32) and deepseek-v4-pro at $0.792 / $2.376 (list $1.32 / $3.96). The struck-through list price and the live rate are both on the models page, synced from the same billing database that rates your requests — the price you see is the price you pay.

Model IDs match DeepSeek’s own naming, so if your code already runs against an OpenAI-style endpoint, migration is base_url plus a key: sign up at app.clfaigateway.dev, top up with VietQR or an international card, and send the first request. The deep dives live at /deepseek-v4-api, /deepseek-v4-flash and /deepseek-v4-pro.