AI Product Updates, July 24: DeepSeek V4 Cutover, NemoClaw 0.0.94, and New Developer Access (2026)

The short version

DeepSeek's V4 Preview is now the migration target for developers whose old API aliases expire at 11:59 p.m. on July 24 (GMT+8). NVIDIA also shipped NemoClaw 0.0.94 with safer sandbox recovery and machine-readable onboarding, while AnySearch opened a free daily quota for students and developers. Qoder and Anthropic added security checks directly to AI coding workflows.

Product or platform	Releasing entity	Release or effective date	Core capability or change	Availability status
DeepSeek-V4 Preview and API cutover	DeepSeek	July 24, 2026 cutover	V4-Pro and V4-Flash are live as open-weight preview models with 1M-token context. The old `deepseek-chat` and `deepseek-reasoner` aliases are being retired.	V4 API available; migrate model IDs before the cutoff. 1
NemoClaw v0.0.94	NVIDIA	July 24, 2026	Safer snapshot restore, `--events=jsonl` onboarding, redacted JSON diagnostics, richer security evidence, and faster Hermes image builds.	Release available in the NemoClaw documentation. 2
Student & Developer Program	AnySearch	July 24, 2026	Free AI-search infrastructure for verified students, developers, and open-source contributors, with 2,000 search calls per day.	Open globally to verified users; the announcement was distributed by PR Newswire and reprinted by Bastille Post. 3
Qoder Security	Qoder	Release date not stated in the accessible official documentation	Security is built into Qoder Desktop and Qoder CLI through static checks, semantic diff scans, cross-file data-flow analysis, fix suggestions, and re-verification.	Available in the product. L1 is free; L2 and L3 consume Qoder Credits. 4 5
Claude Security plugin	Anthropic	July 23, 2026 in the channel timezone	A Claude Code plugin scans recent changes before commit or a full codebase from the terminal.	Beta availability. 6

DeepSeek's V4 cutover is now an API migration task

DeepSeek's official notice calls V4 a preview release and says both deepseek-v4-pro and deepseek-v4-flash are live, open-sourced, and available through the API. Both models support a 1M-token context window and Thinking or Non-Thinking modes. The notice also says the API remains compatible with OpenAI ChatCompletions and Anthropic APIs. 1

The operational deadline is more concrete than the preview label: deepseek-chat and deepseek-reasoner become inaccessible after 11:59 p.m. on July 24 (GMT+8). DeepSeek says those aliases are currently routing to V4-Flash, but it still requires developers to change the model name. The page does not expose a separate publication date, so the date in the table is the API cutover date rather than an inferred announcement date. 1

The migration checklist is short:

Replace the old model ID with deepseek-v4-flash or deepseek-v4-pro.
Recheck prompt behavior in Thinking and Non-Thinking modes; the two modes are part of the new API surface.
Re-test context-heavy calls and cost assumptions rather than assuming the old alias's routing behavior will remain.

The announcement includes a price graphic, but the fetched text does not provide a machine-readable price table. This brief therefore does not reproduce numeric pricing from the image.

NemoClaw tightens the agent sandbox

NVIDIA's NemoClaw 0.0.94 release is mostly operational rather than flashy, but the changes land in the places that make agent systems difficult to run safely. Snapshot restore now uses the snapshot manifest to decide which state directories to clear, preserves target-only directories when appropriate, and re-establishes gateway pairing for cross-sandbox OpenClaw restores. 2

The release also adds a versioned, redacted JSONL event stream through --events=jsonl. Human-readable progress remains on standard error, and closing or slowing the event stream no longer cancels onboarding. For operators, that makes onboarding state easier to feed into logs and automation without treating terminal text as an API. 2

Security evidence gets more structure too. sandbox doctor --json now redacts token-shaped values at both machine-readable output boundaries. Reviewed npm audits retain scanner, registry, timing, package, advisory, report-path, and failure provenance, while a read-only advisory correlation tool can match package and version ranges before a reviewed audit record exists. 2

One performance figure is available from NVIDIA's hosted comparison: the Hermes production build step fell from 205 seconds to 101 seconds, and layer export from 178.8 seconds to 70.4 seconds. Those are release-note figures from one hosted comparison, not a general benchmark guarantee. 2

AnySearch opens a free route for early builders

AnySearch launched a Student & Developer Program that gives each verified user 2,000 free search calls per day. The stated audience is university students, AI developers, and open-source contributors, with use cases including academic research, coursework, application prototypes, and open-source work. 3

The program is aimed at the infrastructure layer rather than a chat application. AnySearch's own site describes the service as AI-search infrastructure with API, MCP, and Skill integrations; the July 24 announcement says developers can use it inside agents and workflows. 3 7

The availability claim comes from a PR Newswire release reprinted by Bastille Post, not from a program page that was accessible in this run. Treat the 2,000-call quota as an announced benefit and verify eligibility, identity checks, rate limits, and retention terms before building a service around it.

Qoder moves security into the coding loop

Qoder Security is presented as a built-in security layer for Qoder Desktop and Qoder CLI. Its product page says the system can detect, verify, and fix vulnerabilities as code moves from the first edit toward commit, using semantic analysis and cross-file data-flow tracing rather than only pattern matching. 4

The documentation separates the workflow into three tiers:

Static Check (L1): automatic pattern checks during coding; the guide says this tier is free.
Lightweight Scan (L2): a semantic, diff-scoped review recommended at the end of a coding task; the guide estimates about 5 Qoder Credits per 500 lines.
Deep Scan (L3): cross-file data-flow analysis recommended before push; the guide estimates about 20 Qoder Credits per 500 lines.

A repository-wide Full Scan is also available, but one invocation is capped at 10,000 lines and can consume substantial time and Credits. Qoder says findings come with fix suggestions, while the developer reviews the final diff rather than allowing an unreviewed automatic commit. 5

The accessible official product pages do not state a separate release date. That makes the current availability and credit model more useful than a guessed launch timestamp: teams can try the free baseline, then measure the token or credit cost of deeper scans against the vulnerabilities they actually catch.

Claude Code gets a beta security plugin

Anthropic's Claude account announced the Claude Security plugin for Claude Code in beta. It can scan recent changes before a commit or scan a full codebase from the terminal, using the Claude inference already attached to the coding workflow. The post was published on July 22 in the source's original offset, which is July 23 in the channel timezone. 6

That puts the feature next to the code rather than in a separate dashboard. It also leaves the normal engineering questions open: which model and data are used for a full-repository scan, how long findings are retained, how false positives are triaged, and what a large scan costs. Beta availability is enough for a small, consented repository trial, not a replacement for the existing security review stack.

What to test next

Update a staging client to deepseek-v4-flash and run the same long-context and tool-use cases before the old aliases expire.
Upgrade a NemoClaw sandbox and route --events=jsonl into an internal log collector; check that secrets are absent from both event and diagnostic outputs.
If you qualify for AnySearch's program, start with a disposable prototype and record actual quota behavior, latency, and retention before placing it in a production agent.
Compare Qoder's L1 baseline with one L2 or L3 scan on a repository whose vulnerabilities are already understood. For Claude Security, run the same controlled comparison inside Claude Code and review the generated diffs manually.

The most immediate action is the DeepSeek model-ID change. The other releases are adoption tests: each can improve a workflow, but each adds a new cost, permission, or data-retention boundary that should be measured before it becomes a default.

AI Product Updates, July 24: DeepSeek V4 Cutover, NemoClaw 0.0.94, and New Developer Access