Anthropic released Opus 4.8 on May 28. Standard pricing is unchanged ($5/$25 per million input/output tokens), context window stays at 1M on the API, Bedrock, and Vertex. Anthropic's own benchmarks vs 4.7:

The "around four times less likely to allow code flaws to pass undetected" claim is in the post but with no methodology and no third-party audit; read it as Anthropic's internal measurement.

Two changes matter more than the model card. Fast Mode on Opus 4.8 is priced at $10/$50 per million tokens — three times cheaper than the same tier on Opus 4.6/4.7 ($30/$150). For latency-sensitive production at Opus quality, that's a real shift. And the Messages API now accepts system-role entries inside the messages array mid-conversation, so callers can update instructions without modifying the top-level system field and without busting any cached prefix. The previous trap with stateful agents — a prompt tweak invalidating the cache and spiking cost — is gone.

Dynamic Workflows in Claude Code is the showcase feature, currently a research preview. Claude writes and executes orchestration scripts that fan out to many parallel subagents inside a single session; the caller describes the work at a high level, Claude builds the pipeline. Anthropic's flagged example is Bun porting 750,000 lines from Zig to Rust in 11 days with 99.8% of the existing tests passing. Salesforce separately reports migrating 33 API endpoints across five PRs (the largest containing 21 endpoints with full test coverage) in 13 days, against an internal 231-person-day estimate. The number is striking and worth the discount: it's an internal estimate, measured with proprietary scoring, by an Anthropic partner. Anthropic also warns Dynamic Workflows "can consume substantially more tokens than a typical Claude Code session." Set a budget before you turn it on.

OpenAI extended Codex's Computer Use to Windows 11 on May 29. The agent can see the screen, move the mouse, type, drive menus and apps directly. macOS got this on April 16; Linux still hasn't been mentioned. There's a meaningful asymmetry: on macOS the agent runs in the background while you keep working in another window, on Windows it has to occupy the active desktop foreground. So no kicking it off and using the same machine for something else.

The mobile half is the new wrinkle. ChatGPT iOS and Android 9.3.0 pair to a running Codex host via a QR scan from the Codex sidebar (same account, MFA completed first, workspace admin opt-in for enterprise). From the phone you can:

• start new threads
• send follow-ups
• approve commands
• see a live desktop thumbnail
• view diffs and terminal output
• stop sessions

Five paired hosts per account. Session caps are 4 hours on Free, 8 hours on Pro and Enterprise. Available on Plus, Team, and Enterprise at no extra cost; 14-day trial for new subscribers. Computer Use is geo-blocked at launch in the EEA, UK, and Switzerland, which is a real reach gap. The model behind it: OpenAI didn't say.

Per Fortune (Microsoft declined to comment), Jacob Andreou — recently appointed head of Copilot, mandate covering both consumer and enterprise — is consolidating GitHub Copilot, Copilot Chat, Copilot Cowork, and a new agentic surface called Autopilot (which houses an always-on agent named Scout) into a single shell.

delivering one Copilot

TestingCatalog has screenshots showing a left-rail with Chat, Code, Cowork, and Scout icons, plus Library and Projects in the sidebar; the Code tab carries a work-tree picker, repo list, and a Routines scheduled-task layer.

The reason this is a real move and not just a UI refresh: GitHub Copilot's ~4.7M paying developers and M365 Copilot's enterprise base have been completely separate products with separate UIs, separate pricing, and separate org homes, even though the underlying models overlap heavily. Pulling 4.7M developers into the same shell as the consumer chat app is the new ground. Microsoft's officially announced M365 Copilot redesign on May 28 is a different effort; it doesn't mention GitHub Copilot or Scout. Reported timing is end of summer, with TestingCatalog hinting at a limited preview around Build (June 2) and a broader rollout into fall, possibly slipping to early 2027. Pricing changes aren't on the table in any source. The whole story is leak-plus-screenshot, not Microsoft on the record.

GPT-Rosalind itself launched in April 2026 as a life-sciences reasoning model wired into 50+ biological databases. What's new on May 29 is the program wrapping it. There are two tracks:

• a developer track giving sponsored (free) API access to teams building epidemiological models, early-detection tools, or DNA-screening systems
• a government track for select US federal agencies and allied-nation partners with public-health or national-security missions

OpenAI absorbs the API costs. Named partners include:

• Lawrence Livermore National Laboratory (medical countermeasures via the Bioresilience Incubator)
• Johns Hopkins Applied Physics Laboratory (protein engineering for therapeutic and biothreat characterization)
• CEPI (vaccine development under the 100 Days Mission)
• the DNA-screening firms Fourth Eon and SecureDNA

OpenAI's self-reported numbers on the model: 0.751 Pass@1 on BixBench, above GPT-5.4 (0.732), Grok 4.2 (0.728), and Gemini 3.1 Pro (0.550); above the 95th percentile on a Dyno Therapeutics RNA prediction task; outperformed GPT-5.4 on 6 of 11 LABBench2 tasks. The safety stack is layered — hard refusals on red lines, classifier-based input monitoring, customer-level institutional attestations — and access is gated rather than open-weights. Two unresolved threads. The audit story is internal; no Rosalind-specific system card has been published. And the line between biodefense and dual-use bio research is genuinely blurry, with OpenAI currently the sole judge of where each request sits. Over 100 scientists have publicly called for tighter governance of bio-AI; the program is moving faster than that conversation.

xAI took grok-build-0.1 (also aliased grok-code-fast-1) to public beta on May 20 via its API, with a CLI called Grok Build wrapping it. Specs:

• 256K context
• text and image input
• native function calling and structured outputs
• reasoning always on with no configurable depth

The pricing line is the news, well below comparable agentic coding tiers:

• $1 per million input tokens
• $2 per million output
• $0.20 per million cached input

xAI quotes 100+ tokens/second; rate limits are 1,800 RPM and 10M TPM, available in us-east-1 and eu-west-1.

The CLI's feature surface is close to parity with Claude Code on orchestration:

• plan mode with a human approval gate
• parallel subagents using git-worktree isolation
• a headless -p mode for CI pipelines
• AGENTS.md and MCP support

Distribution is broad: xAI API, OpenRouter, Vercel AI Gateway, Cursor, plus smaller integrations including Kilo Code, OpenCode, and Hermes Agent. Two gaps worth naming. xAI published no benchmarks of its own at launch — no SWE-bench, no terminal-bench, no HumanEval. A 70.8% SWE-bench Verified figure has circulated from third-party aggregators but does not trace back to xAI. And reasoning-always-on means there's no off-switch when the task doesn't need it; that's a cost lever Anthropic and OpenAI hand callers that xAI doesn't.

Jensen Huang's GTC Taipei / Computex 2026 keynote on May 31–June 1 fired four product shots and teased a fifth. Compressing:

RTX Spark

Nvidia's first consumer laptop SoC (silicon known internally as N1X). The spec sheet:

• 20-core Grace-derived Arm CPU paired with a Blackwell RTX GPU (6,144 CUDA cores, ~RTX 5070-mobile tier)
• up to 128GB unified memory
• 70B transistors on TSMC 3nm
• NVLink chip-to-chip
• full CUDA stack on Windows

Nvidia claims 1 petaflop of AI compute (marketing, no third-party verification). OEMs confirmed for fall: Dell XPS 16, HP OmniBook X 14 and Ultra 16, Lenovo Yoga Pro 9n, Microsoft Surface Laptop Ultra, ASUS ProArt P14/P15, MSI Prestige N16 Flip AI — 30+ laptops and 10+ desktops in the lineup. An early Clang benchmark put RTX Spark 54% over the base M5 but slower than M5 Pro; GPU benchmarks aren't out yet. The actual wedge isn't speed-vs-Apple, it's that any model fine-tuned on Nvidia datacenter hardware now runs natively on the laptop without porting.

Cosmos 3

A single Mixture-of-Transformers omnimodel for physical AI, replacing the four-model Cosmos 2 pipeline (Predict, Transfer, Reason, Policy). Two variants: Nano (16B total, 8B reasoner + 8B generator, runs on RTX PRO 6000) and Super (64B total, 32B + 32B, datacenter-class). One forward pass handles vision reasoning, world generation, and action:

• text→video
• action+image→video (forward dynamics)
• text+video→action (inverse dynamics)
• image+text→video+action (policy)

Released under the NVIDIA Open Model License on HuggingFace alongside training scripts, deployment tools, and six synthetic datasets. Nvidia claims #1 across PAI-Bench, R-Bench Physics-IQ, RoboLab, VANTAGE-Bench, and AI City Challenge 2026 — all self-reported.

Vera Rubin VR200 NVL72

Racks shipped first engineering samples. Per Rubin GPU:

• 50 PFLOPS NVFP4 inference
• 35 PFLOPS training
• 288GB HBM4
• 336B transistors

Full rack: 72 GPUs + 36 Vera CPUs + ConnectX-9 SuperNICs + BlueField-4 DPUs, 260 TB/s NVLink scale-up, 3.6 EFLOPS NVFP4 inference. Headline claim: 10x lower cost per million inference tokens vs the Blackwell platform — Nvidia's framing, baseline is per-token, no third-party number. Fully liquid-cooled, 45°C inlet; per-tray assembly time pitched as down from 2 hours to 5 minutes. Full production H2 2026; AWS, Google Cloud, Azure, OCI, CoreWeave, Lambda, Nebius, and Nscale all named as first deployers.

Nemotron 3 Ultra

Announced June 1 for HuggingFace availability around June 4, Nvidia's biggest open-weights drop yet:

• ~550B total (Nvidia says ~500B; Artificial Analysis estimates 550B), ~55B active parameters
• hybrid Mamba-Transformer MoE with Multi-Token Prediction layers
• 1M context
• NVIDIA Open Model License

Artificial Analysis pegs Intelligence Index at 48 — behind Kimi K2.6 (54), ahead of gpt-oss-120b (33). Nvidia claims 5x faster inference and ~30% cheaper than "leading open models"; the baseline isn't named.

Huang also referenced "a surprise new product" for H2 2026 separate from the keynote launches. Identity undisclosed.

Sam Altman posted on May 31 that OpenAI Robotics is hiring full-stack hardware, ops, systems, and ML engineers. The careers page backs it: at least 11 San Francisco listings as of early June, including:

• Electrical Engineer
• 3D Printing Lab Technician
• Actuator Design Engineer
• DAQ Station Engineer
• Simulation Environments Engineer
• Robotics Lab Manager
• Robotics Software Engineer
• an ML engineer on Distributed Data Systems for Robotics

Most require in-person presence. This is in-house hardware ambition, not just software glued onto someone else's robot.

The 2021 shutdown was explicitly a data problem. Wojciech Zaremba said publicly that internet-scale pretraining gave language models roughly 100x learning efficiency they couldn't replicate for physical robot behavior; RL on self-generated data alone wasn't enough. The bet now is that simulation, video, and world-model research closes the gap. The program sits under Aditya Ramesh's group (which also absorbed the Sora team after the video app was shut down). Caitlin Kalinowski, hardware and robotics ops lead since November 2024, resigned in March citing ethical objections to OpenAI's Pentagon work; Benjamin Bolte replaced her. Hardware partnerships are notably thin: Figure AI severed its OpenAI collaboration in February 2025, and 1X is still a portfolio investment without a confirmed manufacturing role. Altman's framing is near-term robots for skilled workers on infrastructure builds (data centers, power grids, factories) and long-term personal robots, "many years away."

The Information has an internal memo from Alex Himel, Meta's VP of Wearables, sketching out a pendant wearable that records and summarizes conversations. The project descends from Limitless, an AI device startup Meta acquired in December 2025 (founders Dan Siroker and Brett Bejcek; original product was a $99 clip-on/necklace pendant with a Rewind desktop app). Post-acquisition Meta halted Limitless's hardware sales, capped existing-customer support at one year, and absorbed the team into Reality Labs. Meta declined to comment for any of the recent coverage.

Two new internal names surface in the memo: "Muse Spark," Meta's latest (unreleased) AI model, and "Hatch," an unreleased agent. Both are meant to power the pendant and the existing glasses lineup. The memo also describes a "Wearables for Work" enterprise tier alongside consumer subscriptions:

• $7.99/mo (Meta One Plus)
• $19.99/mo (Meta One Premium)

Sales targets are aggressive: 10M wearable devices across the lineup in H2 2026, 6.8M monthly active wearable users by year-end. Ray-Ban Meta sold over 7M units in 2025, which is the distribution base they're cross-selling from. Internal dogfooding for the pendant starts spring 2027, so a consumer launch is at minimum 18 months out and could easily slip. The unanswered question is the one that killed Humane's AI Pin — always-on audio recording around third parties who didn't consent — and the memo doesn't address it. Meta is already facing pushback from 70+ civil-rights groups over the smart glasses' Name Tag facial-recognition feature.

Quick refresher: on February 27, 2026, Hegseth's Pentagon designated Anthropic a "supply chain risk" — historically a tool reserved for foreign adversaries like Huawei — after Anthropic refused to remove two contractual carve-outs from its $200M two-year other transaction agreement with the DoD Chief Digital and AI Office:

• no use of Claude for mass surveillance of US citizens
• no use for fully autonomous weapons systems without meaningful human control of the targeting decision

Hegseth's line:

we will not employ AI models that won't allow you to fight wars

Trump posted on Truth Social ordering every federal agency to cease using Anthropic. Within hours, OpenAI announced a new classified-network deal on "all lawful purposes" terms; xAI signed a comparable arrangement in late February. By May 1, DoD listed seven AI vendors with active classified deals — OpenAI, Google, Microsoft, AWS, Nvidia, SpaceX, and Reflection. Not Anthropic.

Anthropic filed in two courts on March 9. In the Northern District of California, Judge Rita Lin granted a preliminary injunction on March 26 (43-page ruling), finding the designation was retaliation issued because of Anthropic's "hostile manner through the press," and blocking 17 civilian agencies from enforcing it. The Pentagon CTO publicly disagreed and asserted the designation still stood under separate statutory authority. On April 8, the DC Circuit denied Anthropic's request to stay the DoD-specific designation, calling the harm "primarily financial in nature." The May 19 oral argument left the three-judge panel visibly divided — Judge Henderson skeptical of the evidentiary basis, Judges Rao and Katsas pushing back on the court's authority to second-guess national-security calls. No ruling date.

The state of play as of June 1: civilian agencies are enjoined from enforcing the designation, DoD contractors are not, and the structural question underneath — whether vendor-written safety policy survives in writing once a hostile administration treats it as procurement leverage — is what the eventual ruling decides. Anthropic's own framing is that the dispute is narrower than the headlines, but their two competitors took the deal. That's the part that won't go away regardless of how the case turns.

Tejas Chopra (described as a Netflix data-infrastructure engineer, reported elsewhere as ex-Netflix — the truth is somewhere between) open-sourced Headroom, a proxy that runs on port 8787 and accepts any OpenAI-compatible client without code changes. It compresses the noisy parts of agentic context — tool outputs, logs, JSON blobs, code, traces — before the request reaches the model. A ContentRouter dispatches by data type to specialized compressors:

• SmartCrusher for JSON (constant factoring of repeated schema, BM25+embedding relevance scoring, error and stack-trace preservation)
• a tree-sitter CodeCompressor for source (Python, JS, Go, Rust, Java, C++)
• LogCompressor for logs
• Kompress-base — a custom HuggingFace model trained on agentic traces — for general text

Two pieces are the actual novelty. CCR (Compress–Cache–Retrieve) keeps the originals locally in Redis or SQLite with markers in the compressed payload, and exposes an MCP tool headroom_retrieve so the model can fetch the full data mid-task if it needs it. Reversibility without ever shipping the originals to the provider. CacheAligner reorders prompts to keep dynamic content (timestamps, IDs) at the tail and static content at the head, lifting provider-side KV cache hit rates from ~5% to ~80% in Chopra's reported cases. Anthropic prices cache hits at a 90% discount, OpenAI at 50%, so the savings compound.

Reported numbers, on Chopra's own measurements:

• 92% token reduction on code search and SRE-debugging traces (17,765 → 1,408 tokens; 65,694 → 5,118)
• 73% on GitHub issue triage
• 47% on codebase exploration
• GSM8K accuracy unchanged at 0.870, TruthfulQA +0.03
• ~2–5ms proxy overhead per request

About 3,900 stars and 317 forks on GitHub at last check; Apache 2.0. The widely-cited $700K savings figure is community-wide aggregate across all Headroom users since the January 2026 launch, not Netflix-only — most press accounts blur that. And the accuracy benchmarks test reasoning and factual recall rather than agentic tool-use fidelity, which is the harder open question.

Pinterest's conversational shopping assistant runs on Alibaba's open-source Qwen3-VL. CTO Matt Madrigal disclosed on the Q4 FY2025 earnings call (February 2026) — and VentureBeat detailed last week — that Pinterest replaced Qwen's vision encoder with proprietary embeddings precomputed offline from Pinterest's own image corpus, pin metadata, and user signals. The language backbone stays and is fine-tuned. The encoder, the component that turns raw pixels into vision tokens on every request, is gone. Per-request vision compute becomes a one-time batch job over Pinterest's catalog.

The reported number is 90% lower AI operational cost versus the prior approach, which the public sourcing doesn't precisely name. Madrigal also reports a 30% accuracy lift on Pinterest's internal tasks after fine-tuning with proprietary embeddings — the domain-specific data more than overcomes the loss of a general-purpose vision head. Pinterest's internal embedding model is PinCLIP, which they say outperforms open-source multimodal embeddings by 30%+ on core retrieval tasks. The assistant covers visual shopping recommendations across 620M users.

The principle generalizes: any platform with a large, stable, domain-specific image corpus can precompute visual representations once and skip the per-request vision encode. Expect more image-heavy products to copy this pattern. The honest caveats:

• no Pinterest Engineering post details this specific architecture (the open-source AI stance post on Medium is the closest primary)
• the 90% baseline isn't precisely named
• the 30% accuracy claim is internal metrics rather than a neutral benchmark
• the 15% accuracy-drop figure floating in some aggregator coverage doesn't trace to a primary source