AI Daily Dev — Week 28

Court Unseals the Pentagon–Anthropic Emails That Ended the DoD Deal

Full exchange between Dario Amodei and Emil Michael (Pentagon undersecretary for research and engineering) filed in ND Cal on July 2 and published by Gizmodo — first look at the private terms of the fight.
Amodei's redlines: no fully autonomous weapons, no domestic surveillance. Michael's response: 'just not workable,' with 'one more chance to align on core principles' before parting ways.
Pentagon's counter-ask was that Claude cover 'all lawful uses' — Amodei noted US law does allow domestic surveillance, so the phrase would erase both redlines.
Defense Secretary Pete Hegseth designated Anthropic a supply-chain risk the next day; multiple HN threads on the exchange.

industry gizmodo.com

FT report picked up July 2: Altman floated a public-fund-style equity grant directly to Trump, Commerce Secretary Lutnick, and Treasury Secretary Bessent.
Vehicle would be modeled on the Alaska Permanent Fund; proposal envisions Anthropic, Google, and Meta ceding matching 5% stakes.
At OpenAI's current $852B mark, the government slice would be worth about $42.6B — would require an act of Congress.
Comes days after the White House delayed the wider GPT-5.6 launch; active HN thread argues the arrangement bakes in a permanent regulatory conflict of interest.

industry cnbc.com

Anthropic case study out July 6: Alberta's Ministry of Technology and Innovation ran ~50 Claude Code agents in parallel across all 27 provincial ministries.
Coverage: 1,280 applications and 3,400 code repositories — every repository Alberta owns.
Two-stage routine: rules-engine flag pass, then Claude review that cites the exact file and line so developers can verify each finding.
Alberta says the same work would have taken about 6.5 years of manual review; the province is publishing white papers so other governments can copy the setup.

industry anthropic.com

Sol/Terra/Luna move from limited preview to full public rollout today after the Trump administration cleared additional federal testing.
Sol Ultra sets a new Terminal-Bench 2.1 SOTA at 91.9% (base Sol 88.8%), beating GPT-5.5, Fable 5, Opus 4.8, and Gemini 3.1 Pro.
Pricing per 1M tokens: Sol $5/$30, Terra $2.50/$15, Luna $1/$6; Sol's context window grows to 1.5M tokens.
New 'ultra' mode spawns parallel subagents that decompose a task and synthesize their outputs — each subagent bills independently.

models openai.com

Public launch July 8 in Grok Build, the xAI console, and inside Cursor on all plans — built on xAI's 1.5T-parameter V9 foundation.
$2/M input and $6/M output puts it under half the price of GPT-5.6 Sol and a fifth of Opus 4.8.
Averages 15,954 output tokens per SWE-Bench Pro task versus 67,020 for Opus 4.8 (max) — a 4.2× token-efficiency gap.
Trained on Cursor's real developer sessions — debugging traces, multi-file diffs, and user corrections — not just static code.
HN thread already on the front page alongside a 'we made Grok 4.5, GPT-5.5, and Claude build the same apps' comparison.

models x.ai

GPT-5.6 agent takes a goal, pulls context from connected apps, and returns spreadsheets, decks, docs, and interactive sites over hours-long runs.
Codex and ChatGPT merge into one desktop app on macOS and Windows; old app becomes ChatGPT Classic, Chat/Work/Codex land on every plan including Free.
Rolling out July 9 to Pro/Enterprise/Edu on web and mobile; Plus and Business follow in days.
New Chrome extension, revamped in-app browser, and faster Computer Use also shipped alongside.

tools openai.com

Multimodal reasoning model with 1M-token context, native primary/subagent orchestration, and MCP + custom-skill support.
Meta Model API opens in public preview at $1.25/M input and $4.25/M output — ~10× cheaper than Opus 4.8 and GPT-5.5 flagships.
Marks a strategic reversal for Meta: proprietary closed weights via a paid API, not an open Llama drop; open-source Muse Spark still promised later.
Launch partners include Replit, Cline, and Box; SOTA on MedScribe and TaxEval, top row on tool-use and tool-augmented reasoning benchmarks.

models ai.meta.com

New Microsoft operating unit sends forward-deployed pods into enterprise customers to co-design, deploy, and operate AI systems on their data.
Launch clients: London Stock Exchange Group, Unilever, Land O'Lakes, Novo Nordisk, Accenture; SI partners include Accenture, Capgemini, EY, KPMG, and PwC.
Led by Rodrigo Kede Lima, previously president of Microsoft Asia; explicitly designed to route customer data across OpenAI, Anthropic, Microsoft AI, and open-source models.
Fifth big FDE-for-AI vehicle in eight weeks — after OpenAI's and Anthropic's May JVs and AWS's $1B org announced on July 2.

industry techcrunch.com

Meituan open-sourced LongCat-2.0 on June 30 under MIT — a 1.6T-parameter MoE, ~48B active per token, native 1M-token context.
Was already running as the anonymous 'Owl Alpha' on OpenRouter — first on Hermes Agent by monthly volume, second on Claude Code, third on OpenClaw; ~10.1T tokens/month at peak.
SWE-bench Pro 59.5 beats GPT-5.5's 58.6 and Gemini 3.1 Pro's 54.2, though still behind Claude Opus 4.7/4.8.
First trillion-parameter model trained end-to-end on domestic Chinese ASICs (~50,000 chips) — no Nvidia in the loop; API is $0.75/$2.95 per 1M tokens, $0.30/$1.20 during the launch promo.

open-source venturebeat.com

Native macOS/iPhone/iPad build of the 2003 RTS Command & Conquer: Generals Zero Hour — real engine on ARM64, DXVK→MoltenVK→Metal, no emulator.
Built on EA's GPL v3 source drop plus community forks (GeneralsX, Fighter19's Unix port); the AI-driven work is the iOS/iPadOS touch controls, mobile build tooling, and engine fixes.
First build took 40 minutes; two days of debugging burned Reshi's entire Claude Max quota.
Full source open on GitHub; front-paged HN with the expected argument about how much credit belongs to Claude versus the human directing it.

community github.com