deepseek-chat and deepseek-reasoner Are Gone — Migrate to the Wrong Model and Your Bill Triples

By the Deep Seek AI editorial desk · July 25, 2026 · 8 min read

Every other migration guide tells you to rename a string. Most of them rename it to the wrong thing. Both retired aliases map to deepseek-v4-flash — migrate that way and your price is identical. Assume V4-Pro is the reasoner's successor and you pay about 3.1x more for the same tokens.

The documented mapping (both aliases go to the same model)

The mechanical change is trivial. Base URL, request schema, streaming format, tool-calling contract — all unchanged. But note where both aliases land: per DeepSeek's API docs, deepseek-chat maps to V4-Flash in non-thinking mode and deepseek-reasoner maps to V4-Flash in thinking mode. Neither maps to V4-Pro.

That distinction is the whole story. Migrate as documented and your bill is unchanged. Reach for deepseek-v4-pro because it looks like the successor to deepseek-reasoner and you have just tripled your per-token cost without changing a single prompt.

The billing impact, both ways

All rates below are per 1,000,000 tokens, USD, verified against api-docs.deepseek.com/quick_start/pricing on July 25, 2026.

Both aliases → deepseek-v4-flash (price-neutral)

Verdict: no price change at all. Both retired aliases were served by the same V4-Flash rate card, so a documented migration is exactly cost-neutral. Thinking mode (former deepseek-reasoner) is a request-level setting on V4-Flash, not a more expensive model.

The expensive mistake: assuming V4-Pro is the reasoner successor

Verdict: roughly 3.1x more expensive for identical traffic. V4-Pro is a bigger model (1.6T total / 49B active vs 284B / 13B) and may be worth it — but it is a capability upgrade you choose, not the migration target the docs prescribe.

Worked example: a mid-sized production chatbot

Assume 500M input tokens/month with a 30% cache-hit rate (350M cache miss, 150M cache hit) and 100M output tokens/month:

Run your own numbers with our interactive DeepSeek pricing calculator.

What about off-peak discounts and the surge pricing?

This is where most competing migration guides get it wrong, because they read the V3-era docs. Two things you need to know as of July 25, 2026:

• The old off-peak discount is gone. It was a V3/R1 feature and was retired together with those aliases. V4-Flash and V4-Pro do not currently offer an off-peak discount.
• Peak surge pricing is announced but not yet active. On June 30 DeepSeek announced that rates double during two daily windows — Beijing 09:00–12:00 and 14:00–18:00 (UTC 01:00–04:00 and 06:00–10:00) — tied to the official V4 release, which was announced for mid-July 2026. The official price list still shows a single flat rate and no confirmed switch-over date. We check this weekly and will update this page the day it flips.

30-minute migration checklist

1. Grep your codebase for 'deepseek-chat' and 'deepseek-reasoner' — check hardcoded strings, config files, environment variables, and any prompt-management tools (LangSmith, Helicone, Langfuse).
2. Replace both with 'deepseek-v4-flash'. That is the documented mapping for each alias: deepseek-chat → V4-Flash non-thinking, deepseek-reasoner → V4-Flash thinking mode.
3. Do NOT reach for 'deepseek-v4-pro' just because it sounds like the reasoner successor. Same traffic on Pro costs about 3.1x more ($0.435/$0.87 vs $0.14/$0.28 per 1M tokens). Pick Pro deliberately, for capability, not by default.
4. Set thinking mode explicitly where you relied on deepseek-reasoner: V4-Flash serves both modes, so the reasoning behaviour is a request-level choice now, not a model choice.
5. Re-check downstream schema assumptions: the reasoning_content field is still returned in thinking mode, so parsers written for deepseek-reasoner keep working.
6. Leave your cost model alone if you migrated to Flash — the rates are identical. If you moved anything to Pro, re-baseline your FinOps alerts upward by ~3x before they page you.
7. Move async/batch jobs to hours outside Beijing 09:00–12:00 and 14:00–18:00 now, before the announced peak surcharge activates.
8. Deploy behind a feature flag, roll out at 10% → 50% → 100% while watching latency and output-quality regressions.

What most guides get wrong

• Telling you deepseek-reasoner becomes deepseek-v4-pro. The docs map it to V4-Flash in thinking mode. Following that advice triples your bill for identical traffic.
• Framing the migration as a saving. Done correctly it is price-neutral; there is no discount to collect.
• Still quoting the "V3 off-peak 50% discount." That window closed with V3/R1.
• Treating the announced peak surcharge as already-active. As of July 25, 2026 it is not.

FAQ

This is an independent guide published by the Deep Seek AI editorial desk. We are not affiliated with DeepSeek.com. All rates verified against the official pricing page on July 25, 2026 and will be re-verified weekly.