URGENT: DEEPSEEK API DEADLINE JULY 24 AT 15:59 UTC
● deepseek-chat and deepseek-reasoner aliases break July 24 at 15:59 UTC — no grace period
● Migrate: deepseek-chat → deepseek-v4-flash; deepseek-reasoner → deepseek-v4-flash (or v4-pro for heavy reasoning)
● V4-Flash defaults to thinking ON — add thinking disabled if you need the old deepseek-chat behaviour
● Kimi K3 open weights: July 27 — self-hosting eliminates hosted API data residency concern
Full Comparison Table
| Model |
Input /1M |
Output /1M |
Context |
AA Index |
SWE Marathon |
Open weights |
| Kimi K3 |
$3 ($0.30 cached) |
$15 |
1M (Allegretto+) |
57 (#4 of 189) |
42.0% #1 |
July 27, 2026 |
| DeepSeek V4 Pro |
$0.44 (disc.) |
$0.87 (disc.) |
1M |
Not ranked |
Not published |
Available now |
DeepSeek V4 Pro discounted 75% — verify current rate at api.deepseek.com. List price: $1.74/$3.48/M. Both models subject to China National Intelligence Law for hosted API use. K3 Moderato/Vivace tiers: 256K context. Allegretto+ required for 1M.
Intelligence vs Price — The Core Trade-Off
Kimi K3 ranks #4 of 189 models on the Artificial Analysis Intelligence Index (score 57) — above GPT-5.6 Terra, Claude Opus 4.8, and Claude Sonnet 5. DeepSeek V4 Pro has not been independently ranked on this index. K3 leads SWE Marathon at 42.0% — the highest published agentic coding score. DeepSeek V4 Pro's benchmark scores on neutral evaluators have not been published as of July 2026. K3 costs $3/$15/M (or $0.30/M cached input). DeepSeek V4 Pro costs $0.44/$0.87/M at discounted rates — approximately 7x cheaper on input and 17x cheaper on output. The intelligence gap is meaningful. The price gap is decisive for high-volume workloads.
The Open Weights Situation — July 27 Changes the Equation
DeepSeek V4 Pro is already available as open weights — you can self-host it today on your own infrastructure. Kimi K3 open weights ship by July 27, 2026. Before July 27, K3 is API-only (Moonshot hosted), meaning the China National Intelligence Law concern applies to both models equally for hosted use. After July 27, K3 self-hosting on Western infrastructure eliminates that concern — while DeepSeek self-hosting has been possible for months. Both models become equally viable for regulated industry self-hosting once K3 weights ship. The 7-day gap (today vs July 27) is the only remaining window where DeepSeek has a self-hosting advantage.
Which to Use
Choose Kimi K3 when: intelligence ranking matters (AA Index #4 vs unranked), you do frontend or long-horizon agentic coding (Design Arena #1, SWE Marathon 42.0%), or you want to wait for open weights July 27 to self-host with full data control.
Choose DeepSeek V4 Pro when: maximum cost efficiency is the primary driver (7-17x cheaper than K3), you already have DeepSeek integrations and need to migrate before July 24, or you need open weights today rather than waiting until July 27.
If you use DeepSeek today — migrate by July 24. deepseek-chat and deepseek-reasoner aliases break at 15:59 UTC Thursday. The fix is one line: swap to deepseek-v4-flash (same tier) or deepseek-v4-pro (heavy reasoning). V4-Flash defaults thinking ON — add thinking disabled if you need the old non-thinking behaviour.
Last updated July 2026. Related: DeepSeek migration guide — July 24 deadline → · Kimi K3 full review → · DeepSeek V4 Pro vs Sonnet 5 vs Terra →