DeepSeek V4.1 Flash: cheaper routing, Pro remains and peak pricing matters
September 22, 2026 update: Flash moves to 4.1; Pro remains available
DeepSeek released V4.1 Flash on September 10, 2026. It entered this weekly catch-up through AI ta Krajta on September 15 and other roundups. A video's date is not the model's release date.
The current deepseek-flash alias points to multimodal V4.1 Flash. Older Flash and Vision Exp were replaced, with their aliases temporarily redirected. V4 Pro remains available as deepseek-v4-pro.
There is a source conflict: the original announcement described redirecting Pro from September 14. The current changelog explicitly preserves Pro and its billing. Operational decisions should follow current documentation rather than that initial plan. Do not merge the models in your router without testing.
Current prices replace the August table
The official pricing page, checked September 22, lists USD per million tokens:
| Model and period | Cache hit | Cache miss | Output |
|---|---|---|---|
| V4.1 Flash off-peak | 0.003 | 0.15 | 0.60 |
| V4.1 Flash peak | 0.006 | 0.30 | 1.20 |
| V4 Pro off-peak | 0.022 | 0.66 | 1.98 |
| V4 Pro peak | 0.044 | 1.32 | 3.96 |
Peak hours are Monday–Friday, 01:00–04:00 and 06:00–10:00 UTC, excluding Chinese public holidays. Weekends are off-peak. Convert from UTC when scheduling so daylight-saving changes do not break the queue.
For invoices or CRM enrichment, retain the queue design below, update the alias and cost calculation, and rerun extraction and tool-calling evaluations. Non-urgent batches can wait; a customer incident cannot. Flash is a candidate for the first pass, while escalation to Pro should depend on errors measured on your own data.
Historical context: V4 Pro GA in August 2026
The original analysis below explains the introduction of time-based prices. Its August Flash rates and future-tense announcements are historical; use the table and rules above for current budgets. The queue, validation and cost-per-completed-task principles still apply.
DeepSeek V4 is no longer just an April preview. On August 13, 2026, DeepSeek moved V4‑Pro to general availability and announced a new API pricing model: from August 16, the price will also depend on when a workload runs.
That matters more to companies than another benchmark. Batch AI jobs should now be scheduled like cloud compute: interactive requests run immediately, while invoice processing, CRM enrichment and overnight document checks move to off‑peak hours.
What changed
The API name deepseek-v4-pro remains unchanged, but it now points to DeepSeek‑V4‑Pro‑0813. DeepSeek highlights three practical changes:
- stronger agent performance in production scenarios,
- reasoning effort controls for low, high and max,
- native OpenAI Responses API support, explicitly adapted for Codex.
The one-million-token context window from the April release remains. For long-running agents, selective reasoning is more useful than filling every request with the entire archive. Cheap steps can run at low effort; difficult exceptions can escalate.
DeepSeek reports 87.9 on Terminal Bench 2.1 and 31.8 on the public AutomationBench. These are vendor-reported results, not a replacement for your own evaluation. Treat them as a reason to test, not as a reason to migrate automatically.
Pricing from August 16
The new rates take effect on August 16, 2026 at 16:00 UTC. Peak hours are 01:00–04:00 and 06:00–10:00 UTC; all other hours are off-peak. Off-peak rates are half the peak rates.
| Model and window | Cache-hit input / 1M tokens | Cache-miss input / 1M tokens | Output / 1M tokens |
|---|---|---|---|
| V4‑Flash off-peak | $0.007 | $0.22 | $0.66 |
| V4‑Flash peak | $0.014 | $0.44 | $1.32 |
| V4‑Pro off-peak | $0.022 | $0.66 | $1.98 |
| V4‑Pro peak | $0.044 | $1.32 | $3.96 |
Until the change, V4‑Pro costs $0.003625 for a cache hit, $0.435 for a cache miss and $0.87 for output. The “more than 1,000% increase” is mathematically true only for the cheapest cache-hit input during peak hours. Regular V4‑Pro input rises by about 52% off-peak and roughly triples at peak. Output becomes 2.28× more expensive off-peak and 4.55× at peak.
That is more accurate than saying DeepSeek became ten times more expensive.
A router is no longer enough
A model router selects a model by task complexity. The new pricing adds a second axis: how soon the result is needed.
- Run incoming support email or lead triage immediately, preferably on Flash, and escalate only ambiguous cases to Pro.
- Queue invoices, monthly CRM enrichment, contract checks and content pipelines for off-peak processing.
- Split long agent runs into cheap steps and escalations. Extraction, normalization and validation rarely need max; a difficult decision might.
- Measure cache-hit rate for stable prompts and repeated context. Without it, the attractive cache price is only a number on a pricing page.
In n8n or Make, store each job with a priority and deadline instead of calling the model directly from every trigger. A worker processes urgent items immediately and defers the batch. Add an idempotency key, retry limit, audit log and human review for expensive or sensitive outputs.
Example: 10,000 invoices overnight
A sensible pipeline is:
- V4‑Flash extracts fields into structured JSON.
- Deterministic rules validate totals, VAT, currency and duplicates.
- Only ambiguous cases escalate to V4‑Pro.
- The batch runs off-peak; an urgent manual request can jump the queue.
- Store the model, cost, source documents and escalation reason with the result.
This saves more than swapping one model for another. The largest gain comes from keeping easy cases away from the expensive model and avoiding peak capacity for work that can wait.
Before production
Recalculate the budget from the real cache-hit ratio and output length. Then run an evaluation on representative business tasks and measure not only success rate but repeated calls, latency and the cost of correcting errors.
DeepSeek supports OpenAI Chat Completions, the Anthropic API and now native Responses API, but protocol compatibility does not guarantee identical behavior. For agents, verify tool calling, reasoning-history handling, structured output, timeouts and retries.
Bottom line
DeepSeek V4‑Pro GA matters because it combines a stronger agent model with time-dependent pricing. DeepSeek remains inexpensive, but the April price and the blanket 10× cheaper claim are no longer safe assumptions.
The model router decides what handles the job. The queue decides when it runs.
Sources: DeepSeek V4‑Pro GA release, DeepSeek changelog, current and scheduled API pricing, April V4 Preview announcement, Marek Bartoš's video.