Top Stories
The U.S. Government Is Now Deciding Who Uses GPT-5.6
OpenAI has publicly previewed GPT-5.6, shipping the three-model lineup (Sol, Terra, Luna) only to a small group of trusted partners while the U.S. government vets the first wave of users. The company’s post acknowledges the unusual arrangement in unusually direct terms, warning that it “keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them” and that this kind of access process should not become the long-term default. Preview partners’ participation is being shared with the executive branch during the gated window.
Independent reporting from the Washington Post confirms the executive branch is now the gatekeeper for the frontier model release list. Read together with yesterday’s Anthropic Mythos reversal (see below), the two events confirm that the voluntary review process behind the executive order on cybersecurity is operating as a real release-gate for both American frontier labs, not a courtesy notification.
Anthropic Mythos Returns to 100+ U.S. Trusted Partners, Fable 5 Talks Continue
Commerce Secretary Howard Lutnick has authorized Anthropic to redeploy Mythos 5 to more than 100 U.S. agencies and companies, ending a roughly two-week standoff. The June 26 letter is addressed to Anthropic’s chief compute officer Tom Brown and states Lutnick determined “appropriate safeguards are in place to permit certain trusted partners to access the Claude Mythos 5 Model.” No license is required to export, reexport, or transfer the model to entities in Annex A, including their foreign national employees and to Anthropic’s own foreign staff. The Mythos 5 review began after concerns that the model had reached partners too closely linked to China, including reportedly a South Korean telecom.
Fable 5, the more public Mythos variant available before the original directive, was not addressed by the letter and remains offline for general users. Anthropic is also reportedly negotiating around Fable 5 separately, with talks described as moving toward release but without a clear timeline. European officials have publicly expressed frustration that allied access to frontier American models is now subject to U.S. licensing terms.
A Founder Fed Claude His Scans, Bloodwork, and Journal to Fight Cancer
Keragon founder Conno Christou has published a first-person account of running Anthropic’s Claude on a continuous medical workflow through six months of chemotherapy for an aggressive non-Hodgkin’s lymphoma discovered at age 35. He fed the model blood results, scan data, wearable output (Whoop band, Oura ring), and voice-transcribed symptom journal entries, then used it to gather twelve medical opinions and weigh conflicting recommendations (60% vs. 85% success rate regimens). The harder regimen won the support of eleven of twelve doctors he consulted.
In a more striking moment, when a final PET scan came back ambiguous, Christou loaded three PET scans and an MRI into Claude. The model flagged thymus rebound (a known false-positive risk in under-40 lymphoma survivors) at roughly 90% probability. A fourth specialist confirmed the call, sparing him radiotherapy. Christou’s framing is sober: “It didn’t replace the doctors, but helped me ask the right questions.” The piece is a useful anecdote for anyone tracking how consumers are stitching frontier models into health decisions outside the clinic.
Anthropic Accuses Alibaba of “Largest Known Distillation Attack” on Claude
Anthropic has publicly accused Alibaba of running a deliberate, large-scale distillation campaign against Claude, describing it as the “largest known distillation attack” on the company. Distillation trains a weaker model on a stronger one’s outputs, and Anthropic used the word “brazen” in its characterization. The accusation follows a pattern: Anthropic has previously alleged similar techniques by other Chinese AI labs.
The allegation lands in the same week Anthropic’s chief compute officer Tom Brown, not CEO Dario Amodei, has been identified by MIT Technology Review as the company’s point of contact with the White House, framed as a more pragmatic counterpart. The combination suggests Anthropic is now fighting on three fronts simultaneously: defending its access to U.S. customers, expanding access for allied nations, and pushing back on what it views as capability extraction by Chinese rivals.
Open vs Closed Source: The Benchmark You Pick Decides the Verdict
Jamie Dborin at Doubleword has published a clean technical test of the claim that “open-source catches closed by Christmas.” A widely circulated chart, drawn from a single Artificial Analysis benchmark, suggests the gap is collapsing to zero around December 3, 2026. Dborin repeated the same exercise across 18 different Artificial Analysis benchmarks (AIME, GPQA, HLE, MMLU Pro, MATH 500, LiveCodeBench, Tau2, TerminalBench, and others). When you average across all of them, the line of best fit is almost completely flat at just under five months behind.
The visible improvement is concentrated in coding, where the gap moved from 15 months to one or two. Most other datasets show only modest movement, and some show the gap slightly growing. The headline implication: whether open weights are catching up depends almost entirely on which benchmark you draw the line through, and aggregate-by-default measures paint a much more stable picture than single-benchmark cherry-picks.
Two New Routers Aim to Cut Local-AI Costs Without an LLM in the Loop
Two new routing tools landed this week that decide where your prompts go without making their own model call. Wayfinder Router (Apache-2.0) is a deterministic, offline scorer that picks a local-or-hosted tier by structural features of the prompt (length, code presence, lists, headings). Same prompt plus same threshold always yields the same answer, and decisions take microseconds. It supports any OpenAI-compatible endpoint, including Ollama, vLLM, LM Studio, llama.cpp, plus Claude and Gemini adapters.
workweave/router takes a different approach: it sits as a drop-in proxy that picks the best upstream per request using an on-device embedder and a cluster scorer based on the Avengers-Pro routing paper. It advertises sub-50ms routing latency and 40-70% cost reductions just by switching endpoints. Native support covers Anthropic Messages, OpenAI Chat Completions, and Gemini, with Claude Code, Codex CLI, opencode, and Cursor integrations. Together with DeepSeek’s newly released DSpark speculative-decoding paper under DeepSpec, the week reinforces a pattern: the next wave of self-hosting wins is coming from inference plumbing (routing, drafting, batching) as much as from new base models.
AI Is Already Doing PhD-Level Mathematics
IEEE Spectrum profiles three frontier demonstrations of AI doing publishable research mathematics. Math Inc.’s Gauss formalized Maryna Viazovska’s Fields Medal-winning sphere-packing proof, helped humans finish the 8-dimensional case in days, then autonomously formalized the 24-dimensional case in two weeks. Google DeepMind’s Aletheia produced publishable PhD-level research calculating structure constants in arithmetic geometry on an unsolved problem. A general-purpose OpenAI system disproved an important conjecture in combinatorial geometry, a result top mathematicians described as a milestone in independent, original, and sophisticated thinking.
These are not benchmark wins. They are peer-review-worthy contributions to research mathematics. The spectrum piece also flags the broader concern: as AI systems generate results faster than humans can verify them, the verification burden shifts, and the meaning of authorship in research mathematics is up for grabs.
AI-Designed Radio Chips That Look Like QR Codes
A Princeton team has built a two-stage reinforcement-learning plus diffusion pipeline that produces RFIC layouts in roughly six minutes, designs that look more like QR codes than the careful hand-drawn geometries of human radio engineers. The first stage (an AlphaGo Zero-style RL agent) decides system architecture and circuit topology; the second stage uses a convolutional neural network as a Maxwell’s-equation emulator and a diffusion model to invert from desired scattering parameters to physical layouts. In 2023, the pipeline produced a power amplifier spanning 30 to 100 GHz that hit the best reported combination of bandwidth, output power, and efficiency for silicon.
The piece also flags the open-data gap: the U.S. CHIPS Act R&D program and Natcast’s ML/RFIC efforts have been closed, and industry design data stays behind NDAs. Without a shared, ImageNet-equivalent dataset for radio chips, the technique won’t generalize as fast as it could.
Quick Hits
-
Post-Mythos cybersecurity: keep calm and carry on: A practitioner-side read on defensive posture after the Mythos export-control cycle and the Sakana/360 frontier cyber releases argues against reactive over-investment. Most pre-Mythos controls still hold. Cephalosec
-
Tom Brown, not Dario Amodei, leads Anthropic White House talks: Anthropic’s chief compute officer is now the company’s main point of contact with the executive branch, framed in reporting as a more pragmatic counterpart. The shift is part of a broader effort to keep Mythos-class releases flowing under the new licensing regime. MIT Technology Review
-
DeepSeek ships DSpark speculative decoding: Part of the broader DeepSpec release, the paper describes a new speculative-decoding design for faster LLM inference. Self-hosters chasing throughput per dollar will want to read it. DeepSeek (paper)
-
EU joins U.S. alignment to limit Chinese AI reliance: Brussels has aligned more directly with Washington’s China-AI decoupling posture, with the Global South remaining the contested battleground for both blocs. MIT Technology Review
Worth Watching
The voluntary-licensing regime is now operating. The Anthropic Mythos reversal and OpenAI’s GPT-5.6 customer-by-customer preview both route frontier model access through the executive branch. The next test is what the new executive-order framework on cybersecurity defines as a “repeatable process” for future releases. That document will decide whether this stays a short-term friction point or becomes the structural default for allied-nation access, open-weight releases, and the next Mythos-class model.
The distillation fight is going public. Anthropic’s “brazen” framing of the alleged Alibaba campaign is unusually direct and likely to trigger retaliatory export-control moves, renewed debate about weight-leak defenses, and copycat accusations from other labs. Watch whether Alibaba responds publicly and whether the U.S. government ties the allegation to the broader China-AI policy alignment now being negotiated with the EU.
Inference plumbing is the new battleground. DeepSeek’s DSpark paper, paired with Wayfinder Router and workweave/router, points to the next twelve months of self-hosting gains coming from routing, drafting, and batching rather than base-model improvements. The economics of local-AI are going to shift faster than the model-release calendar suggests.