From 59bdc28468f4d262816516a1a8053594241b7fa3 Mon Sep 17 00:00:00 2001 From: Eric Wang Date: Thu, 13 Aug 2026 02:03:23 -0700 Subject: [PATCH] feat(site): news - V4-Pro in the field, what practitioners report Real first-day HN accounts on deepseek-v4-pro-0813 (launch thread, 900+ pts), sourced with per-comment links: Pro is more direct than Flash (which flails at ~5x output); benchmarks vs a real deploy task diverge and the harness is why; the economics win only with disciplined caching (never rewrite history); it beats GLM/Qwen on price; and the quant/provider trust worry that ties to the rollback note. Anecdote, marked as anecdote, consistent with the bench page. Co-Authored-By: Claude Fable 5 --- site/build.py | 83 ++++++++++++++++++++++++++++++++++++++++++++ site/news/index.html | 83 ++++++++++++++++++++++++++++++++++++++++++++ 2 files changed, 166 insertions(+) diff --git a/site/build.py b/site/build.py index 03bb973..3075c4d 100644 --- a/site/build.py +++ b/site/build.py @@ -1384,6 +1384,89 @@ def jstr(s): live API where that is possible; the in-terminal feed is ds docs changelog.

+

2026-08-13 · V4-Pro in the field: what practitioners report

+

The launch table is DeepSeek's own. This is what +people running deepseek-v4-pro-0813 in real harnesses said in the +first day, mostly in the +Hacker News launch +thread (900+ points). It is anecdote, not measurement, but it is consistent +anecdote, and it lines up with the sober reading on the bench page.

+ +

Pro thinks less and says less than Flash

+

The clearest repeated observation is the split between the two models. One +reviewer running both through OpenRouter in a pi harness put it plainly: Flash +makes a lot more initial mistakes, and then has to re-check stuff, and +produces much more output compared to Pro … the output volume is often 5x +more … saying something wrong … and then ‘Wait, let me +re-check:’ before it lands on the right answer +(pixelesque). Pro +reaches the conclusion more directly. If you are paying for output tokens and +wall-clock, that difference is the argument for Pro on the hard steps and Flash +on the bounded ones, which is the route-by-role +advice from the other direction.

+ +

The benchmarks and a real task disagree, and the harness is why

+

The most useful counterweight was a like-for-like test: scan a repo and +produce a single docker-compose to deploy behind Caddy, with +wildcard certs provisioned outside, some ports already taken, and Postgres +built in. Run on deepseek-v4-pro-0813 and on +gpt-5.6-terra-high, the verdict was this one had few issues. +terra: none, with the note that the gap opens up as the project stops being +simple (freakynit, +repro). The +immediate reply from another user was the load-bearing caveat: +DS V4 is harness sensitive +(scrlk). A score is a +model plus a harness, and a cheap model with a good harness (self-verification +tools, an AGENTS.md it rereads) closes a lot of the gap the raw +number shows.

+ +

The economics are the story, but only if you cache right

+

This is where the field agrees with the kill +line. One practitioner built a cost simulation from their own pi sessions +and found that once caching is counted, deepseek-v4-pro comes out +cheaper than gpt-5.6-luna, on the strength of the cache-read +price (taosx). The +catch, stated bluntly by another: a 50% cache-hit rate is a misconfigured +harness, an agentic loop should see 99%+, and DeepSeek's cache pricing is low +enough that you should never change history … even more so with deepseek +because their cache hit pricing is so low (no stripping old thinking tokens, +no rewriting tool results to save context) +(p1necone). The 138x +cache-hit lever on the cost page is only real if the +prefix actually stays put.

+ +

Against the other open models

+

On GLM-5.2, one report from OpenCode work on an old Unix-clone codebase: +GLM ended up being far slower, and far more expensive, for approximately the +same results … There was never a problem that GLM could solve that DS +couldn't solve, faster, and significantly cheaper, with the honest coda that +DS isn't as good as GPT or Claude … but it's fast, and pretty darned +effective (spijdar, +who also runs a 3-bit quant locally at ~15 tok/s). Against Qwen3.8-max, released +the same day, the read was comparable capability at much lower price, with no +reason to reach for Qwen unless you need vision +(parsimo2010).

+ +

The trust question: is the endpoint even serving the real model?

+

One thread ties straight to the entry below. An open-weights user asked to be +able to prove to me … that this output was generated by FP8 DeepSeek V4 +Pro 0813. Not some cheaper quantization +(Phemist), and others +noted they pin to a single provider for deepseek specifically. The model will +not tell you which checkpoint or precision it is, and its self-description is +worthless for it. The signals that hold up are the ones you can check without +asking it: the endpoint's model list and the docs version cell, which is exactly +the point of the rollback question below.

+ +

The bottom line from the field

+

It matches the bench page. V4-Pro-0813 sits in the frontier-adjacent cluster, +not clear of it; a stronger closed model still lands the hardest task in fewer +turns; and the reason to run it anyway is that cheap cached input makes retries, +review passes and parallel workers affordable. Route by role, keep the prefix +stable, and measure successful-task cost rather than the per-token price or the +launch chart.

+

2026-08-13 · No, V4-Pro did not roll back

Ask deepseek-v4-pro what model it is and it may tell you it is GPT-4o. A model that answers with a competitor's name looks like a swap, and diff --git a/site/news/index.html b/site/news/index.html index 155c025..64f0ca7 100644 --- a/site/news/index.html +++ b/site/news/index.html @@ -81,6 +81,89 @@

News

live API where that is possible; the in-terminal feed is ds docs changelog.

+

2026-08-13 · V4-Pro in the field: what practitioners report

+

The launch table is DeepSeek's own. This is what +people running deepseek-v4-pro-0813 in real harnesses said in the +first day, mostly in the +Hacker News launch +thread (900+ points). It is anecdote, not measurement, but it is consistent +anecdote, and it lines up with the sober reading on the bench page.

+ +

Pro thinks less and says less than Flash

+

The clearest repeated observation is the split between the two models. One +reviewer running both through OpenRouter in a pi harness put it plainly: Flash +makes a lot more initial mistakes, and then has to re-check stuff, and +produces much more output compared to Pro … the output volume is often 5x +more … saying something wrong … and then ‘Wait, let me +re-check:’ before it lands on the right answer +(pixelesque). Pro +reaches the conclusion more directly. If you are paying for output tokens and +wall-clock, that difference is the argument for Pro on the hard steps and Flash +on the bounded ones, which is the route-by-role +advice from the other direction.

+ +

The benchmarks and a real task disagree, and the harness is why

+

The most useful counterweight was a like-for-like test: scan a repo and +produce a single docker-compose to deploy behind Caddy, with +wildcard certs provisioned outside, some ports already taken, and Postgres +built in. Run on deepseek-v4-pro-0813 and on +gpt-5.6-terra-high, the verdict was this one had few issues. +terra: none, with the note that the gap opens up as the project stops being +simple (freakynit, +repro). The +immediate reply from another user was the load-bearing caveat: +DS V4 is harness sensitive +(scrlk). A score is a +model plus a harness, and a cheap model with a good harness (self-verification +tools, an AGENTS.md it rereads) closes a lot of the gap the raw +number shows.

+ +

The economics are the story, but only if you cache right

+

This is where the field agrees with the kill +line. One practitioner built a cost simulation from their own pi sessions +and found that once caching is counted, deepseek-v4-pro comes out +cheaper than gpt-5.6-luna, on the strength of the cache-read +price (taosx). The +catch, stated bluntly by another: a 50% cache-hit rate is a misconfigured +harness, an agentic loop should see 99%+, and DeepSeek's cache pricing is low +enough that you should never change history … even more so with deepseek +because their cache hit pricing is so low (no stripping old thinking tokens, +no rewriting tool results to save context) +(p1necone). The 138x +cache-hit lever on the cost page is only real if the +prefix actually stays put.

+ +

Against the other open models

+

On GLM-5.2, one report from OpenCode work on an old Unix-clone codebase: +GLM ended up being far slower, and far more expensive, for approximately the +same results … There was never a problem that GLM could solve that DS +couldn't solve, faster, and significantly cheaper, with the honest coda that +DS isn't as good as GPT or Claude … but it's fast, and pretty darned +effective (spijdar, +who also runs a 3-bit quant locally at ~15 tok/s). Against Qwen3.8-max, released +the same day, the read was comparable capability at much lower price, with no +reason to reach for Qwen unless you need vision +(parsimo2010).

+ +

The trust question: is the endpoint even serving the real model?

+

One thread ties straight to the entry below. An open-weights user asked to be +able to prove to me … that this output was generated by FP8 DeepSeek V4 +Pro 0813. Not some cheaper quantization +(Phemist), and others +noted they pin to a single provider for deepseek specifically. The model will +not tell you which checkpoint or precision it is, and its self-description is +worthless for it. The signals that hold up are the ones you can check without +asking it: the endpoint's model list and the docs version cell, which is exactly +the point of the rollback question below.

+ +

The bottom line from the field

+

It matches the bench page. V4-Pro-0813 sits in the frontier-adjacent cluster, +not clear of it; a stronger closed model still lands the hardest task in fewer +turns; and the reason to run it anyway is that cheap cached input makes retries, +review passes and parallel workers affordable. Route by role, keep the prefix +stable, and measure successful-task cost rather than the per-token price or the +launch chart.

+

2026-08-13 · No, V4-Pro did not roll back

Ask deepseek-v4-pro what model it is and it may tell you it is GPT-4o. A model that answers with a competitor's name looks like a swap, and