DeepSeek opens search, file parsing and safety checks as APIs, Google details Antigravity's 800,000-line Rust rewrite, Anthropic lays out always-on agent architecture
No new models this weekend; the changes are all around the model. DeepSeek is selling the pieces behind its app by the call, while Google, Anthropic and AWS each published how they build long-running agents: verify with tests the agent cannot change, carry state through a durable inbox and memory, and hide cold starts with snapshots and pre-warming. For product teams, the gap is moving from which model you pick to how well you build what surrounds it.
deepseek-flash supports these in the Messages, Responses and Chat Completions APIs: the deepseek_web_search tool runs web or image searches and answers with citations, and with thinking enabled it searches and reads links on its own over several steps; PDFs, Office files and images are added to context after plain-text extraction or rich-layout parsing; deepseek_read_url fetches links found in the user's message.
A new safety check parameter reviews model input and output and returns a refusal marker when it finds a risk. DeepSeek says services offered to the public should enable it and meet obligations such as content safety review, large model registration and AI labeling.
The underlying capabilities are also standalone APIs: Search and Fetch, OCR, ASR and TTS. DeepSeek has open-sourced a demo chat client built on them that can import DeepSeek chat history.
Pricing: search and web fetch are $0.75 per 1,000 each, file parsing and OCR $0.225 per 1,000 pages; TTS lists at $0.45 per 10,000 characters and ASR at $0.116 per hour of audio, both currently 50% off.
Builder's takeFor AI Cloud Drive, file parsing at $0.225 per 1,000 pages is a cheap baseline. I would run my own scanned documents and tables through it and compare accuracy and cost per page with the current pipeline before replacing any part of it. Products serving users in China can turn the safety check on as a first line of defense, but it does not do the registration and labeling for you; your own review process still has to exist.
The design pairs parallel exploration with falsifiers (the Lean 4 proof checker, unmodified test suites, execution simulators, hardware counters and others) to decide which candidates work; each agent gets only the files and tools its step needs, and critics re-check results from clean contexts without the producer's reasoning.
A new rewrite pattern, coming to Antigravity, schedules each module only after its dependencies compile and links every leaf module against the original, unmodified test suite. Building on a 25% port, a two-week Argon campaign added about 800,000 lines of safe Rust that pass all kernel unit tests and boot on arm64 hardware.
libgav1, Google's AV1 decoder, was rewritten into about 170k lines of Rust with 100% bit-exact parity on a broad conformance suite; it started about 4x slower than C++, and after /teamwork got profiling tools and a mandate to maximize speed it ended 20-30% faster than the original C++ and hand-optimized assembly on x86_64 and ARM.
/teamwork and /boost are available in Antigravity today with Gemini 4 models; Google has also opened applications giving 10,000 academic researchers six months of free access to its highest usage tier.
Builder's takeThe method is what to copy, not the model: verify with tests the agent cannot edit, order work by dependencies, and give each agent only the context it needs. The failure mode the post names, stubbing out dependencies or changing tests so a broken port passes, is exactly what I see most when coding agents do migrations, so I would make the test directory read-only before letting one loose. The numbers are Google's own; try /boost on a real migration of one small module first.
An always-on agent spends most of its time asleep and wakes for a run when an event arrives. Every source writes to an inbox, usually a database table, instead of calling the agent; each inbox has one active worker at a time, events are deduplicated by source IDs, and messages arriving in a row are batched into one run.
Timers are either follow-ups the agent sets for itself or recurring schedules, kept in their own store and polled by a scheduler; memory lives in durable storage and holds only what the agent cannot look up again: user instructions, notes about the user, what it has already said and task progress.
For permissions, credentials are injected by a proxy outside the agent's environment; higher-risk actions go to classifiers as in Claude Code's auto mode; when approval is needed, a deferred tool call ends the run, and the user's answer returns as a new inbox event that resumes the session.
The post cites one long-horizon coding benchmark where Opus 5.5 cost 78% less per solved task than Opus 5, against 40% less on typical workloads.
Builder's takeThis brief is itself an agent that wakes on a schedule, and several points match mistakes I have made; I wrote about some of them in my blog post reviewing this quarter's brief pipeline rules. The one to copy first is start quiet and earn attention: if PandaClaws adds proactive reminders, give the outbox a rate cap and evals before sending more. The other is one worker per inbox, which removes most duplicate-run bugs.
V2 starts the container once per create or update, waits for the first healthy ping within 120 seconds, then freezes the microVM and stores its memory and disk as a snapshot that each new session restores. The cost is a few extra minutes on every create or update while the snapshot is built.
Memory comes back whole, but the file cache is cleared before the freeze and image files are fetched on first read, so imports and client creation belong in init, and toolchains that run as separate processes, such as tsc, can be copied to /dev/shm during init.
Build hooks: the runtime sends two POST requests to an HTTP endpoint in the container during the build, with a header marking memory and then disk, sharing 120 seconds; anything other than a 2xx fails the create or update, with no retry.
To get under a second, start the session in the background when the user opens the app or chat window; the idle timeout defaults to 15 minutes (60 seconds to 8 hours) and a session lives at most 8 hours.
Builder's takeIn AI Interview, the two seconds of silence before the first reply hurt the experience, and pre-warming the session when the candidate opens the interview page is the cheapest fix; the cost is a session for everyone who opens it and never starts, so keep the idle timeout short. Whether or not you use AgentCore, moving work into init and keeping the first request lean applies to any service with cold starts.
The three tiers can be combined freely: the SageMaker endpoint tier records full requests and responses for compliance and Model Monitor; the load balancer tier records only access-log fields for low-cost traffic and error monitoring; the model pod tier records full payloads plus event metadata for debugging, analytics and training data.
The demo uses a synthetic home-appliance support router: after a product launch, the Other share rose from 4.8% to 36.8%; Athena queries surfaced the blind spot, a SQL policy labeled the captures, and a 60-second rank-16 LoRA fine-tune on one L40S trained about 29.9 million parameters, under 1% of the 3.1B model.
Held-out accuracy on 280 tickets went from 235 (83.9%) to 280 (100%) with no regression on the five existing classes, but the model scored 20 of 24 on hand-written probes, having learned two shortcuts: cooling means air conditioner, and any mention of Arctis means the AC category.
AWS measured near-identical p50 and p99 latency and throughput through the capture proxy versus direct to vLLM; the post warns that full payloads can include names, addresses and order numbers, and recommends sampling, KMS encryption, lifecycle rules and PII redaction.
Builder's takeThe most useful part is the 24 hand-written probes: an eval set from the same generator says 100%, and a change of phrasing exposes the shortcuts. For intent routing in AI Cloud Drive I would keep two eval sets the same way, one from captured traffic and one written by hand in deliberately different wording. If you capture production data, settle redaction and retention first, then think about the flywheel.
4,537,291 parameters including the waveform generator, trained from scratch with no pretrained weights on about 18 hours of synthetic speech.
Two American English voices, F2 (female) and M2 (male), with 24 kHz mono output; text is first converted to phonemes by the author's dizione phonemizer.
Ships 18.2 MB PyTorch weights and a 21.4 MB ONNX export (float32, opset 17) with SHA256 hashes on the model card; the inference script reads text of any length one sentence at a time and supports speed and pitch controls.
Builder's takeEnglish only, two voices, and a one-person project, so it will not go straight into a Chinese product. But if AI Interview's English practice mode only needs to read questions aloud, a 20 MB on-device model removes a server TTS call; listen to the samples, check the hashes on the model card, and compare against today's discounted DeepSeek TTS price.
Researched and drafted with AI assistance; editorial standards and views set by Darius.
Every Monday, the week in AI in one email
Every Monday morning: the 6 AI stories that mattered most in the past seven days, with my take, in your inbox.
Your email is only used to send the weekly and is never shared. Every email has a one-click unsubscribe link. See the latest issue →