AI Brief, 19 October 2026: a search API at 75 cents per thousand, and two silent-corruption fixes

DeepSeek has opened its in-house search engine as a paid API, and the price is the story. The changelog entry, "Opening up the in-house search API", is dated 17 October and sits directly above the V4.1-Flash-1017 entry covered here yesterday, so the two shipped together and only the model was reported at the time. Attention arrived a day later: the documentation took 559 points on Hacker News after being posted at 06:50 UTC on 18 October. Two endpoints, search and fetch, cost $0.75 per 1,000 successful requests each. Brave's own pricing page lists $5 per 1,000; Exa's lists $4 for its cheapest tier and $1 per 1,000 pages for page text on top.

The billing definition matters as much as the number. A request is billable only when it returns something: HTTP 200 with a non-empty results array for search, or actual page content for fetch. Failures, timeouts and searches that match nothing are not charged, and count changes how many results come back without changing the price. For an agent loop that issues speculative queries and discards most of them, that shifts the unit of cost from attempts to hits. The constraint sits elsewhere: 5 queries per second per account, against the 50 QPS Brave lists, so this is a tenth of the throughput ceiling at a seventh of the price. Per-request billing does not start until 00:00 Beijing time on 1 November; calls through DeepSeek Harness, Claude Code and Codex are free until then.

It was otherwise a quiet weekend, and worth saying so plainly rather than padding. None of the twelve organisations that ship weights regularly created a new repository inside the window. arXiv announced no batch at all: its Friday cutoff fell at 17:59 UTC on 16 October and there is no Saturday run, so the next listing lands on 20 October. What did ship came from the serving stack, where both vLLM and SGLang cut releases carrying fixes for bugs that produced wrong output without raising an error.

• DeepSeek's search and fetch endpoints cost $0.75 per 1,000 successful requests; empty results and failures are free, and the limit is 5 QPS per account.
• vLLM v0.32.0 (18 October) fixes "silent CUTLASS NVFP4/MXFP4 MoE corruption under EP" and drops Python 3.10; max_tokens is now an upper bound rather than a context reservation.
• SGLang v0.5.22 (19 October, 02:55 UTC) fixes a case where hybrid linear models resumed from the wrong Mamba states when speculative decoding and structured output were combined.
• SGLang's new InstantTensor loader reports DeepSeek-R1 loading in 15.3 seconds instead of 160 on eight H200s from 50 GB/s storage.
• Anthropic's own employment report, published 16 October, puts the junior share of its technical hires at 9%, down from 27% in early 2025, and says agent work hours passed human work hours in March.
• No vendor in the decision-model class has published a fail-open rate, asked here daily since 10 October and now ten days.

DeepSeek prices retrieval below the retrieval vendors #

The endpoints are plain JSON over HTTPS, not a model call, and they are documented as generally available with no beta label or waitlist. POST /v1/tools/search takes a query of 1 to 128 characters and supports the operator set a search engineer would expect: site:, -site:, negation, quoted phrases, lang:, after: and before:. POST /v1/tools/fetch takes a single url and returns the page as Markdown.

curl -s https://api.deepseek.com/v1/tools/search \
  -H "Authorization: Bearer $DEEPSEEK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"query": "mamba state speculative decoding bug site:github.com",
       "count": 5,
       "need_content": false}'          # content costs nothing extra, but fetch is a separate charge

What comes back is shaped for a model rather than a results page. Each hit carries a snippet of roughly 100 to 200 characters for display and a summary of roughly 600 to 800 characters that the documentation says is written to be read by a model. Two integer fields do the ranking work: rerank_score from 0 to 3, which DeepSeek describes as combining relevance, authority and freshness, with 1.5 and above meaning high quality; and authority from 0 to 4, with -1 for unknown and 3 and above meaning a highly authoritative source. A date field is RFC 3339 UTC or null. None of this is measured by anyone outside DeepSeek, and the claim that the index covers "hundreds of billions of web pages" is the company's own.

The comparison that matters for an agent is not search alone but search followed by reading a page, because that pair is what a retrieval step actually does. At DeepSeek that is two billable requests, $1.50 per thousand. At Exa's cheapest tier it is $4 for the search plus $1 per thousand pages for the text, so $5. Brave's $5 listing is for search results without page bodies at all.

The 5 QPS ceiling is the part to check before building on it. It is per account rather than per key, the search and fetch limits are counted independently, and fetch carries an additional per-domain limit. Anything over the limit gets HTTP 429.

Two serving stacks shipped fixes for bugs that did not throw #

Both of the major open-source inference servers cut releases inside the window, and in each case the most consequential line is a correctness fix rather than a feature.

vLLM v0.32.0, on 18 October, lists under quantisation fixes: "silent CUTLASS NVFP4/MXFP4 MoE corruption under EP". Expert parallelism with NVFP4 or MXFP4 mixture-of-experts kernels was returning corrupted output without erroring. The release notes do not name the affected GPUs or quantify the accuracy impact, which leaves anyone who has been serving a quantised MoE under expert parallelism unable to tell from the notes whether their outputs were affected or for how long.

SGLang v0.5.22, at 02:55 UTC on 19 October, carries the same shape of problem in a different subsystem. Listed as an important fix: hybrid linear models "could resume from wrong mamba states when speculative decoding and structured output were used together". Two features that each work alone corrupted recurrent state when combined. The notes credit the find to Baseten, which is to say it surfaced from someone running it in production rather than from the project's own tests.

Both releases also raised their Python floor to 3.11 in the same 24 hours, and vLLM's breaking list is long enough that an upgrade needs reading rather than skimming. Three changes alter behaviour rather than removing it:

# 1. max_tokens is now an UPPER BOUND, not a context-window reservation.
#    Previously this could be rejected up front; now it simply caps the completion.
{"max_tokens": 4096}

# 2. CORS no longer defaults to "*". Browser clients get nothing without this flag.
#    vllm serve ... --allowed-origins https://app.example.com

# 3. mamba_cache_mode="all" now FAILS VALIDATION rather than being accepted.

Cascade attention is removed outright, as are the torch.compile sequence-parallelism and AsyncTP passes, and transformers is pinned to >=5.16.1,<5.20.0. EAGLE drafters for Llama 4 and Mistral Large 3 are fixed, a regression that had been in place since v0.28.0.

SGLang's headline feature is a loader, and it is worth working through because the number explains the mechanism. The notes claim models "load 10x faster" with the new InstantTensor loader, with the specific case given as DeepSeek-R1 on eight H200s loading in 15.3 seconds instead of 160 from storage rated at 50 GB/s.

The DeepSeek-R1 repository holds 684,489,845,504 bytes of safetensors, about 684.5 GB, almost all of it fp8. So the two load times correspond to

$$\frac{684.5\mathrm{GB}}{15.3\mathrm{s}}=44.7\mathrm{GB}/\mathrm{s},\frac{684.5\mathrm{GB}}{160\mathrm{s}}=4.3\mathrm{GB}/\mathrm{s}$$

where the numerator is the weight volume and the denominator the wall-clock load. Against a 50 GB/s line rate, the new path sustains 89% of available bandwidth and the old one sustained 8.6%. The speedup is therefore not compression or a cleverer format; it is that the previous loader left roughly nine tenths of the storage bandwidth idle. That reframes cold-start time for large-weight serving as an I/O scheduling problem, which is checkable by anyone with the hardware, and the figure is SGLang's own.

Elsewhere in the release, DeepSeek-V4.1 now runs on AMD MI350X and MI355X with up to 1.9 times the throughput of plain decode from DSpark on four MI350X, giving the model released two days ago a second-vendor path. The remaining performance figures in both releases, including vLLM's 43% to 85% more requests per second on 32-token Qwen3-0.6B prompts and SGLang's 2.3 times throughput per GPU on H200, come from individual pull-request benchmarks rather than one harness, and none has been reproduced independently.

Anthropic published its own hiring numbers #

How employment is changing at Anthropic, an Anthropic Institute technical report by Nathan Wilmers, Regina Catipon, Maxim Massenkoff and Peter McCrory, went up on 16 October and was not covered here at the time. It is a company measuring itself, which makes every figure self-reported, and it is still the most specific public account of what agent adoption does to the shape of a workforce.

The hiring figures are the sharpest. The junior share of employees in technical departments has fallen from 21% to 14% since early 2025; the junior share of technical hires has fallen from 27% to 9%. In April 2026 about a third of ML managers and a fifth of software managers said they were hiring less because of AI. Mentions of a candidate's AI tool use in interview scorecards roughly tripled for technical roles between 2024 and 2026.

On the work itself: agent work hours passed human work hours for the first time in March 2026, and by August individual employees' interactive agents were logging six times the hours of their human handlers. Commits per engineer reached five times the level of six months earlier. Coding-agent sessions by non-technical staff went from 2 a month in August 2025 to 90 by May 2026. The share of interactive sessions spent on what the report calls meta-work, building agent tooling and monitoring pipelines, rose from 10% to 22%. Standing instructions for agents grew from a few hundred in July 2025 to 8,500 a year later.

Two numbers cut against a simple productivity reading. About 80% of interactive session time is idle, waiting on a human, which puts the bottleneck on review rather than generation. And in a within-person comparison of 192 engineers who adopted coding agents against 74 who did not, agent- written and human-written code showed no difference in later deletion or reversion rates: evidence that the extra code is not obviously worse, not that it is better. The authors are explicit that adoption was not random, that adopters were already on an upward trajectory, that more code does not mean more business impact, and that Anthropic is an unusual firm whose practices may not generalise.

Also notable #

• Anthropic committed 5% of its shares to a Public Benefit Fund on 18 October, rising by a further 2% per 50% increase in valuation, as reported by TechCrunch. The figures are secondhand; the company's own announcement was not traced.

• Anthropic's proposed federal AI safety legislation drew attention on 18 October over its trigger: reporting says obligations attach only above 50 deaths or $1 billion in damages, and route enforcement through the Commerce Department. That characterisation is Tom's Hardware's and the bill text itself has not been read against it. It sits opposite the Chinese draft covered here on 17 October: one jurisdiction proposing licensing at a capability threshold, the other liability at a casualty threshold.

• Terence Tao published "Will it be a golden age or a dark age for mathematics (and science)?" on 18 October, on his own blog.

• OrcaCyber-Nano-1.0, a security-focused vision-language model, has collected 72 likes against zero downloads since 16 October. Its card claims 98.60% on MATH-500 and 78.4% on SWE-bench while also stating that the "benchmark harness, settings and full result tables are being finalized", puts the parameter count at 27B in prose against 2B in the repository metadata, and gates the weights behind an access agreement. Nothing on it is checkable yet.

• Mistral's Timestral family, published 15 October and not covered here, is a set of time-series forecasters at 5M, 28M, 123M and 209M parameters under a modified MIT licence, forecasting on CPU in 140 ms for the largest. Mistral claims top-ten placings on GIFT-Eval, TIME and fev-bench, but the post carries its scores only inside chart images, with no numbers in text.

• Sphinx, a git hook that quizzes the committer on code an agent wrote before letting it ship, took 242 points on Hacker News on 18 October. It inverts the usual direction of agent code review by gating the human rather than the model, and it claims no measurements of any kind.

What to watch #

• Whether anyone measures DeepSeek's search quality or latency before billing starts. Per-request billing begins 1 November, and until then the calls are free through three harnesses, which makes this the cheapest window there will be to benchmark result relevance against Brave or Exa. Every figure attached to the index is currently DeepSeek's own.
• Whether vLLM scopes the NVFP4/MXFP4 corruption. Which GPUs, which expert-parallel configurations, and from which release. Without that, nobody serving a quantised MoE can tell whether published numbers measured on an affected build need rerunning.
• Whether anyone reproduces the 15.3-second load. The claim is a bandwidth-utilisation figure, the loader is public, and a storage rate is easy to state, so this is among the more falsifiable performance claims of the month.
• Whether Timestral's technical report puts its scores in text. The same rasterised-benchmark problem was flagged here on 13 and 16 October against two Meta launches; a third instance in a month makes it a pattern rather than an oversight.
• Whether any decision-model vendor publishes a fail-open rate. Asked here on every day since 10 October, now ten. The newest open artifact in the class, Jeva-4B, carries an Apache-2.0 licence and 1,246 downloads, and its card reports no null-detection, coverage or abstention figure at all, while the paper behind it claims "near-perfect null detection" without a number.