DeepSeek retired the model names deepseek-chat and deepseek-reasoner at 15:59 UTC on 24 July 2026. Both had pointed at DeepSeek-V4-Flash — the first at its non-thinking mode, the second at its thinking mode.
For anyone still calling the old names, the fix is a string change. The more useful question is what the V4 rate card does to a workload once you look past the headline number, so we priced the migration.
How the deadline was communicated
-
V4 lands, deprecation announced in the same breath
The changelog entry that introduces V4-Pro and V4-Flash also gives the legacy names three months to live.
-
No dedicated announcement follows
The deadline appears in the changelog and as an inline note on the pricing page. There is no standalone migration post.
-
deepseek-chat and deepseek-reasoner deprecated at 15:59 UTC
Both names had pointed at V4-Flash — non-thinking and thinking mode respectively.
Three months is reasonable notice. Where it thins out is the channel: the deadline lives in a changelog entry that also announces V4, and as an inline note on the pricing and docs pages. There is no standalone migration announcement. A team that reads release notes caught it in April; a team that does not may have found out when a call failed.
What the migration actually costs
| What you were on | What it actually was | Input $/M | Output $/M | Cache-hit input $/M |
|---|---|---|---|---|
| deepseek-chat | V4-Flash, non-thinking mode | $0.14 | $0.28 | $0.0028 |
| deepseek-reasoner | V4-Flash, thinking mode | $0.14 | $0.28 | $0.0028 |
| (the upgrade path) | V4-Pro | $0.435 | $0.87 | $0.003625 |
Both retired names resolved to V4-Flash, so a like-for-like migration is a rename, not a repricing. Moving to V4-Pro is a deliberate upgrade: 3.11× the input price and 3.1× the output.
Like-for-like, nothing changes. Both retired names resolved to V4-Flash, so pointing your client at deepseek-v4-flash preserves your pricing exactly. The migration is only expensive if you treat it as an opportunity to move up a tier.
The upgrade is a real decision. V4-Pro costs 3.11 times V4-Flash on input and 3.11 times on output. That is a considered trade for harder reasoning work, not a default.
Where DeepSeek sits against the field
| Model | In / Out $/M | Blend 3:1 | Blend 10:1 | Cached in |
|---|---|---|---|---|
| DeepSeek V4-Flash | $0.14 / $0.28 | $0.18 | $0.15 | $0.0028 |
| DeepSeek V4-Pro | $0.435 / $0.87 | $0.54 | $0.47 | $0.003625 |
| Claude Haiku 4.5 | $1.00 / $5.00 | $2.00 | $1.36 | $0.10 |
| GPT-5.6 Luna | $1.00 / $6.00 | $2.25 | $1.45 | $0.10 |
| Gemini 3.6 Flash | $1.50 / $7.50 | $3.00 | $2.05 | $0.15 |
Blended $/M = (r × input + output) ÷ (r + 1), using CACHE-MISS input — the honest basis for comparison.
At an agent-shaped 10:1 blend, V4-Flash lands at US$0.15 per million tokens — roughly a ninth of Claude Haiku 4.5 and a thirteenth of Gemini 3.6 Flash. Even V4-Pro, at US$0.47, undercuts every non-DeepSeek model in the workhorse tier. On price alone this end of the market is not close.
The number to be careful with
DeepSeek publishes a cache-hit input price of US$0.0028 per million on V4-Flash, which is 50 times lower than its cache-miss price of US$0.14. On V4-Pro the gap is wider still: US$0.003625 against US$0.435, a factor of 120.
That is a real discount on genuinely repeated prompt prefixes, and it is why every comparison in this article uses cache-miss pricing. Quote the cache-hit figure in a business case and you are assuming a cache-hit rate nobody has measured on your traffic. The honest planning number is the miss price, with cache savings treated as upside.
The caveats that matter
- The failure mode after the cutoff is undocumented. DeepSeek's docs say only that the names are deprecated; they describe no grace period and no error contract.
- No peak or off-peak pricing appears on the current V4 rate card. Time-of-day discounting existed in the V3 and R1 era; the pricing page as published carries flat rates only.
- Both V4 models share a 1M-token context and a 384K max output, so the migration does not change what fits in a request.
Key takeaways
- The deadline.
deepseek-chatanddeepseek-reasonerstopped working at 15:59 UTC on 24 July 2026. - The fix. Both names resolved to V4-Flash, so a like-for-like migration is a rename with no price change.
- The upgrade. V4-Pro is 3.11 times V4-Flash on both input and output — a decision, not a default.
- The comms. Three months' notice, delivered through a changelog entry and an inline docs note, with no dedicated announcement.
- The number to watch. The sub-cent headline is a cache-hit rate, 50 times below cache-miss on Flash and 120 times below on Pro. Plan on the miss price.
View author profile → · Editorial policy
