The clock
On July 24, DeepSeek turns off the deepseek-chat and deepseek-reasoner endpoints that most of the ecosystem has been building against since early 2025. Every integration pointed at those model names stops working that day, not gets deprecated with a grace period. If your production traffic still calls deepseek-reasoner, you have a migration to plan now, not a blog post to read later.
Default: move to V4 Flash
For most of what people used deepseek-reasoner for, DeepSeek V4 Flash is the right first-party landing spot. It runs 13B active parameters against a 284B total, a fraction of R1’s 37B active, and ships a 1M token context window that most reasoning workloads never got from the old API. Lower active params means lower inference cost per token on the provider side, and that tends to show up directly on the model’s live price.
The catch is that active-parameter count is a proxy for compute, not for reasoning depth. Flash’s routing is built for throughput on retrieval, summarization, and tool-calling, the same shape as GLM-4.5 Air or Qwen3 30B A3B: small active slice, large total capacity, optimized for volume. If your old deepseek-reasoner calls were mostly agent loops and structured extraction, Flash covers you and costs less.
The trap in the Pro price
If you’re comparing DeepSeek’s two current API tiers, check the number instead of trusting a chart you bookmarked. V4 Pro’s launch promo expired May 31, and a lot of comparison content still quotes that intro rate as current. Pro runs 49B active against 1600B total, more than triple R1’s active compute, and its live price reflects that now. The gap between what people remember and what the page shows today is exactly the kind of thing static pricing tables get wrong. Pull the current number before you size a migration around it.
Where Flash is the wrong call
If your product depends on R1-style multi-step reasoning traces, chain-of-thought that’s actually load-bearing for output quality, downgrading to Flash’s 13B active slice is a quality regression dressed up as a cost win. Flash was built to replace R1’s cost structure for tasks that never needed that reasoning behavior in the first place. Those are different problems, and treating them as one is how migrations quietly break output quality three weeks after everyone stopped watching the dashboards.
For that workload, don’t shrink the model. Move the same model to a different host. R1’s open weights are still served by multiple providers, and the R1 model page tracks all of them side by side. Spreads between cheapest and priciest provider on comparable open models routinely run past 20x as of this writing, gpt-oss-120b and qwen3-235b-a22b-instruct-2507 both show that pattern on their live pages right now. R1 hosting isn’t exempt from that spread. Check the provider table before assuming the first result you find is competitive.
The self-host math
Self-hosting R1 makes sense once your sustained volume is high enough that even the cheapest third-party host loses to owning the GPU time outright, and our breakeven guide already covers that volume threshold. What that guide doesn’t spell out is the hardware floor. R1 is 685B total parameters. At roughly 1.1GB per billion for FP8 weights, that’s about 750GB just to load the model, and reasoning workloads add another 25% or so in KV cache headroom for the long context chain-of-thought needs. Call it 950GB of VRAM before serving a single request. A single H200 at 141GB doesn’t get you there; you’re realistically looking at eight of them, or an equivalent H100 SXM cluster with more cards and more interconnect tax.
Dropping to INT4 to shrink that footprint is the obvious move, and it’s the wrong one for R1 specifically. Quantizing below FP8 hits hardest on models doing long sequential reasoning chains, because every dropped bit of precision compounds across the chain instead of averaging out the way it does in a single forward pass. Quantize R1 to fit fewer cards and you’re compressing the exact computation that made you choose R1 over Flash to begin with.
The rule
If your deepseek-reasoner traffic is agentic tool use, extraction, or anything where the model produces one answer and moves on, migrate to V4 Flash and take the active-parameter discount. If your traffic depends on genuine multi-step reasoning output, don’t downgrade the model to save on the migration: move to a different R1 host, checking the live spread on the R1 page before picking one, and only start pricing out an eight-card H200 cluster once your monthly spend on any single R1 provider clears what that cluster would cost sitting idle. This guide can’t tell you where your specific prompts fall on that line. That’s an eval against your own traffic, not a spec sheet you can read off a pricing page.