DeepSeek switched off deepseek-chat and deepseek-reasoner yesterday. The rename is one line; the reasoning default it flips on is what moves your bill.
On July 24 at 15:59 UTC, DeepSeek retired its two oldest API model names. Requests to deepseek-chat and deepseek-reasoner, the strings half the tutorials on the internet still use, now fail outright. The official replacement for both is deepseek-v4-flash, at $0.14 input and $0.28 output per million tokens on a cache miss, and an almost absurd $0.0028 on a hit. The prices are fine. The trap is that V4-Flash turns thinking on by default, and deepseek-chat was the non-thinking name, so the obvious one-line rename quietly converts a chat workload into a reasoning workload. Here is the exact mapping, the rate card that died against the one that replaced it, and the peak-hour pricing announcement still hanging over all of it.
Photo by Oliver Tsappis on Unsplash
The fix, in thirty seconds
- Swap the model string to deepseek-v4-flash. Both retired names point there, not to V4-Pro.
- If you were on deepseek-chat, explicitly disable thinking in the request body. V4-Flash defaults to thinking on, and that default is the whole cost story of this migration.
- If you were on deepseek-reasoner, you were already paying for reasoning tokens. The rename changes nothing about your bill.
Three months of warning, then a hard stop
Credit where due: this was not a surprise shutdown. DeepSeek announced the retirement on April 24, the same day the V4 preview shipped, and spent the whole transition routing the legacy names to the new model. If your integration broke yesterday, the three months of notice were sitting in the changelog the whole time. The sequence matters for understanding your invoices, so here it is in order.
| Date | What happened | Effect on your bill |
|---|---|---|
| April 24 | V4 preview ships. Legacy names silently rerouted to V4-Flash. | deepseek-chat callers moved to Flash rates without touching code. |
| May 22 | V4-Pro's 75% launch discount made permanent. | Pro settles at $0.435 in / $0.87 out, today's rates. |
| June 29 | Peak-hour pricing announced to API users by email. | Nothing yet. Still not on the rate card as of July 25. |
| July 20 | The official V4 release DeepSeek had promised for mid-July, per press reports. | No price change. GA kept the May rates. |
| July 24, 15:59 UTC | deepseek-chat and deepseek-reasoner fully retired. | Requests on the old names stop billing because they stop working. |
The April routing detail has a consequence people keep missing: if you were still calling deepseek-chat last week, you were not running V3.2. You had been on V4-Flash weights and V4-Flash prices for three months. Whatever quality or cost drift you noticed since late April was the V4 migration happening to you early, not a preview of what the rename will do.
The rename trap: thinking is on unless you say otherwise
DeepSeek's own docs state the mapping plainly. deepseek-chat corresponds to the non-thinking mode of deepseek-v4-flash, and deepseek-reasoner corresponds to the thinking mode. But deepseek-v4-flash, called bare with no thinking parameter, defaults to thinking on. Put those two facts together: a developer who fixes yesterday's outage by swapping the model string and redeploying has just moved a non-reasoning workload onto a reasoning model without deciding to.
The per-token rates do not move, which is what makes this easy to miss in a quick test. The token counts move. Reasoning traces bill as output at $0.28 per million, and a thinking response carries its chain of work along with the answer. Take a small chat workload at 10M input and 2M output a month: $1.96 at Flash rates with thinking off. If thinking triples the output tokens, and multiples in that range are normal for reasoning traces on short answers, the same traffic bills $3.08, a 57% jump from a one-line diff. On a chat product doing hundreds of millions of tokens, that is a budget conversation, plus the latency of a model that now deliberates before replying.
The reverse mistake costs more. Some deepseek-reasoner users will read "reasoner" as the premium tier and migrate to deepseek-v4-pro. That is not the mapping. Reasoner maps to Flash with thinking on, and Pro is a separate model at a bit over three times the Flash price on both cache-miss input and output. Choosing Pro might be the right call, its benchmark scores are better, but it is an upgrade decision, not a migration step.
The rate card that died against the one that replaced it
For anyone who genuinely was on the old rates until April, the move to V4 is a straight discount. The last published V3.2 pricing behind deepseek-chat was $0.28 cache-miss input and $0.42 output per million. Flash halves the input and takes a third off the output. And the cache-hit line went from a 10x discount to a 50x one.
| Per 1M tokens | V3.2 (final deepseek-chat rates) | deepseek-v4-flash | deepseek-v4-pro |
|---|---|---|---|
| Input, cache miss | $0.28 | $0.14 | $0.435 |
| Input, cache hit | $0.028 | $0.0028 | $0.003625 |
| Output | $0.42 | $0.28 | $0.87 |
| Context / max output | 128K / 8K | 1M / 384K | 1M / 384K |
Pro's cache hit is the strangest number on the card. Pro costs 3.1x Flash on a cache miss, but only $0.003625 against $0.0028 on a hit, a 1.3x gap. On a cache-heavy agent loop where 90% of input tokens are repeated context, the input side of Pro is nearly Flash-priced and the real premium collapses toward the output line. If you were staying on Flash purely for input cost, check your cache hit rate first; the calculator handles cached-rate math.
Where Flash sits in the cheap tier now
Since spring the budget end of the market has gotten crowded, so it is worth restating what $0.14/$0.28 means in context. Every western lab's cheap model costs multiples of Flash. Only Alibaba's qwen-flash undercuts it, at rates so low they read like typos. Prices below are list, per million tokens, from each provider's pricing page this week.
| Model | Input | Cached input | Output | 12M in / 3M out |
|---|---|---|---|---|
| qwen-flash (up to 256K) | $0.05 | n/a | $0.40 | $1.80 |
| DeepSeek V4-Flash | $0.14 | $0.0028 | $0.28 | $2.52 |
| Gemini 3.5 Flash-Lite | $0.30 | $0.03 | $2.50 | $11.10 |
| Claude Haiku 4.5 | $1.00 | $0.10 | $5.00 | $27.00 |
| GPT-5.6 Luna | $1.00 | $0.10 | $6.00 | $30.00 |
| GLM-5.2 | $1.40 | $0.26 | $4.40 | $30.00 |
The 12M/3M column is a modest chat product's month, and the spread runs from under $2 to $30 for the same traffic. We have argued before that the cheap end of this market is Chinese, and the V4-Flash card does not change that. What it changes is the floor: a 1M context window and 384K max output at these prices did not exist anywhere in March.
The benchmark sheet assumes the expensive mode
One more reason the thinking default matters: it is the mode the marketing numbers come from. The scores on both HuggingFace model cards are labeled Max, DeepSeek's highest reasoning setting. Run Flash with thinking off, the old deepseek-chat behavior, and you should not expect the numbers below. Both models are MIT licensed with open weights; Pro is a 1.6T-parameter mixture with 49B active, Flash is 284B with 13B active.
| Benchmark | V4-Flash (Max) | V4-Pro (Max) |
|---|---|---|
| SWE-bench Verified | 79.0 | 80.6 |
| LiveCodeBench (Pass@1) | 91.6 | 93.5 |
| Terminal-bench 2.0 | 56.9 | 67.9 |
| GPQA Diamond | 88.1 | 90.1 |
| BrowseComp | 73.2 | 83.4 |
Flash within 1.6 points of Pro on SWE-bench Verified, at under a third of the price, is the headline for coding workloads. Where Pro earns its 3.1x is the agentic stuff: an 11-point gap on Terminal-bench and a 10-point gap on BrowseComp are real differences in task completion, not benchmark noise. Pick by workload shape, and remember every number in that table was produced by a model burning reasoning tokens you will be billed for.
What to watch between now and the first peak-hour bill
The shutdown was the easy part of DeepSeek's summer. The open question is peak pricing. In late June, DeepSeek told API customers by email that daily windows of 9:00 to 12:00 and 14:00 to 18:00 Beijing time will bill at roughly double, with changes taking effect on 24 hours' notice. We covered the announcement in our peak-hour surge pricing piece. As of July 25, neither the English nor the Chinese pricing page shows a peak table. The flat rates in this post are what is billing today, and the doubling is a scheduled storm that has not made landfall.
The arithmetic when it lands: those windows cover 7 of 24 hours, so a uniform around-the-clock workload would see its DeepSeek spend rise about 29%. A US-shaped workload mostly will not care, because 9:00 to 12:00 Beijing is 9pm to midnight US Eastern and the afternoon window is 2am to 6am. If your traffic peaks in American business hours, you are already off-peak. Either way, do one thing this week: log the model name and thinking mode on every request. You just watched a provider reroute and retire model names with the bill as the only witness. When the peak email arrives, you will want your own numbers, not a guess.
Sources
- DeepSeek: Models and pricing - V4-Flash and V4-Pro rates, deprecation footnote with the exact mapping
- DeepSeek: V4 preview release notes - April 24 announcement, July 24 retirement date, transition routing
- HuggingFace: DeepSeek-V4-Pro model card - Benchmark table, 1.6T/49B active parameters, MIT license
- HuggingFace: DeepSeek-V4-Flash model card - Flash benchmarks, 284B/13B active parameters
- TechNode: DeepSeek peak-time API pricing - June 29 email announcement, peak windows and 2x multiplier
- Google: Gemini API pricing - Gemini 3.5 Flash-Lite comparison rates
- OpenAI: API pricing - GPT-5.6 Luna comparison rates