Claude Code Scanned 466 Million Lines of Government Code in 20 Hours — Top 10 AI & Flutter News July 07, 2026
Claude / Anthropic
Alberta’s Government Used Claude Code to Scan 466 Million Lines of Code in 20 Hours
The scale: Alberta’s Ministry of Technology and Innovation maintains the systems of all 27 provincial ministries — roughly 1,280 applications and 3,400 code repositories, covering everything from social services to public safety. Using Claude Code (running both Opus and Sonnet models), the team scanned that entire estate in 20 hours, at 466 million lines of code.
How the remediation works: A “red team” agent probes each application the way an outside attacker would, mapping how a vulnerability could be exploited. A “blue team” agent then checks the application’s defenses against an international security standard and writes a remediation plan pointing to the exact files to fix. Where a fix is safe to ship, Claude generates the patch, writes tests if none existed, and builds it.
The standout example: a subsidy-program portal hand-coded in Java roughly 25 years ago — too outdated and tangled to patch efficiently — got rebuilt from scratch in a more maintainable language in just four to five days. Every patch, regardless of how it was produced, still goes through human engineer review before shipping. That’s the detail worth sitting with: this isn’t autonomous government software deployment, it’s a human-reviewed pipeline that happens to move at AI speed.
Source: https://www.anthropic.com/news/alberta-government-claude-cybersecurity
Claude Code v2.1.202: A New /config Setting Lets You Control Agent Parallelism Directly
Dynamic workflow size: A new /config setting called "Dynamic workflow size" lets you directly control how many Claude agents run in parallel, instead of leaving it to the harness's own judgment.
Better observability: New OpenTelemetry attributes — workflow.run_id and workflow.name — make it easier to trace and monitor /workflows runs.
Smaller fixes: This release also patches a crash in inline Ctrl+R history search, a bug where /rename on background sessions got reverted after a job restart, transient mTLS handshake failures during certificate rotation, and cleans up the /workflows agent list layout with wider titles and a dedicated time column.
Source: https://code.claude.com/docs/en/changelog
New Anthropic Research Finds a “Global Workspace” Structure Inside Claude
What the paper says: Anthropic’s new research paper presents evidence that Claude and other frontier language models have developed something structurally similar to the “conscious accessibility” seen in the human brain — the mechanism that determines which information becomes broadly available for further processing.
Worth being precise here: this isn’t a claim that Claude is conscious. It’s an architectural observation about how information propagates and gets shared internally within the model — closer to a finding about information routing than a philosophical statement about experience.
Where it fits: Anthropic has been running parallel tracks of research — the Economic Index series measuring real-world impact, and interpretability work trying to understand what’s actually happening inside the model. This paper sits squarely in the second bucket.
Source: https://www.anthropic.com/research/global-workspace
OpenAI / Codex
GPT-5.6’s General Access Window Opens July 7–14, and Pricing Is Finally Confirmed
The price tags: Sol, the flagship, is $5 input / $30 output per million tokens. Terra, the mid-tier option, is $2.50 / $15. Luna, the fastest and cheapest, is $1 / $6.
Where it stands today: Access remains limited to approved partners and organizations through the OpenAI API and Codex — there’s still no confirmed general-availability date. OpenAI is also rolling out Sol on Cerebras hardware this month, targeting up to 750 tokens per second.
The gap between “pricing confirmed” and “GA date confirmed” is doing a lot of work here — it signals OpenAI is comfortable committing to unit economics before it’s comfortable committing to capacity at scale.
Source: https://help.openai.com/en/articles/20001325-a-preview-of-gpt-56-sol-terra-and-luna
Codex’s New /import Command Pulls In Your Claude Code Setup and Chat History
Migration, made simple: The new /import command lets you selectively bring in your Claude Code project setup, configuration, and recent chat history directly into Codex.
Operational upgrades: This release also adds a /usage view for daily, weekly, and cumulative token consumption, permanent session deletion (via codex delete, /delete, or the app-server's thread/delete, with confirmation safeguards), and expanded Bedrock and OAuth credential support.
What got cut: The experimental /realtime voice controls were removed in this release — development effort has clearly shifted toward cloud operations, alongside the DigitalOcean plugin for remote workspace provisioning.
Source: https://developers.openai.com/codex/changelog
ChatGPT Business Gets the Same Six-Tier Model Picker Personal Accounts Already Had
The new lineup: Across web, iOS, and Android, the model picker for ChatGPT Business now reads Instant, Medium, High, Extra High, Pro Standard, and Pro Extended. Medium replaces Thinking Standard, High replaces Thinking Extended, Extra High replaces Thinking Heavy, and Thinking Light is gone entirely.
Nothing changes under the hood: The underlying models and usage limits included with ChatGPT Business stay the same — this is a relabeling exercise. Pro Standard and Pro Extended both still run on GPT-5.5 Pro, and Instant can auto-escalate to Medium when a request needs more reasoning.
Rollout pattern: This same six-tier picker already shipped for personal ChatGPT accounts back in June; Business is just catching up.
Source: https://help.openai.com/en/articles/11391654-chatgpt-business-release-notes
Flutter / Dart
Flutter’s Global Tour Lands in India: I/O Connect Bengaluru on July 14
The event: Google I/O Connect Bengaluru runs July 14 at the Bangalore International Exhibition Centre, part of the Flutter team’s 2026 world tour spanning more than 20 cities.
What’s on the agenda: Four tracks anchor the day, with AI front and center — hands-on sessions on building autonomous coding agents and working with the latest Gemini models. Flutter’s generative UI work, which dynamically composes and personalizes interface elements, is also expected to make an appearance.
This is the Flutter × AI story that keeps compounding: every stop on this tour — Berlin in June, now Bengaluru — treats AI-assisted, agent-driven development as the default framing rather than a bolt-on topic, which says something about where the core team thinks the ecosystem is actually headed.
Why it matters beyond one event: Following Berlin’s stop on June 25, Bengaluru marks the tour’s serious push into Asia — a signal that Flutter’s community investment isn’t concentrated in any single region.
Source: https://developers.googleblog.com/en/the-io-journey-continues-in-bengaluru/
Content Authenticity / AI Detection
ICML 2026 Opens in Seoul — Watermark-Based Detection Catches 398 Reviewers Breaking the Rules
The scale: ICML 2026 opened July 6 in Seoul with 23,918 submissions, roughly double last year’s total. To catch reviewers using LLMs against policy, the program committee embedded machine-readable instructions in submitted PDFs — invisible to humans, readable by any LLM processing the document — telling it to include two specific phrases drawn randomly from a dictionary of over 170,000 entries.
Why it worked: The conference ran a two-track policy: Policy A banned all LLM use in review writing, while Policy B allowed limited assistance (understanding a paper, polishing prose) but not judging quality or drafting the review itself. Frontier LLMs followed the injected instructions more than 80% of the time in pre-deadline tests. The result: 398 reviewers caught violating policy, triggering 497 desk-rejections.
This is a genuinely useful data point for anyone working on AI content detection: a watermark-style approach just worked at real operational scale, in a setting — peer review — where the stakes for getting it wrong are unusually high. Whether that generalizes to messier, adversarial contexts outside academic publishing is the open question.
GitHub Copilot Tightens Up Its Usage Metrics API — Three Fixes Aimed at Cost Visibility
What improved: Copilot CLI now reports suggested lines of code, users previously identified only through server-side telemetry now have their IDE correctly attributed, and AI credit consumption is tracked with more complete attribution.
Why now: As agentic usage scales up, admins need to know exactly which client and which feature is burning through AI credits — this update answers that need directly.
Same-day context: GitHub also shipped a public preview of agent session streaming on July 2 — usage visibility is clearly a running theme in Copilot’s current release cadence.
Gemini 3.5 Pro Enters Its Second Week of July Still Missing in Action
The delay, again: Announced at Google I/O in May, Gemini 3.5 Pro has now missed both of its own self-imposed deadlines — a June GA target, then a June 30 target — and remains stuck in a limited Vertex AI enterprise preview with no confirmed GA date, benchmarks, or pricing.
Why it’s stuck: Beyond coding performance and long-task reasoning quality, token efficiency appears to be the bigger blocker. Token efficiency has become a real procurement metric in 2026, and some enterprise testers reported the Flash-tier models burning through tokens faster than expected, pushing real-world cost above what benchmark numbers suggested.
Two missed deadlines in a row is a pattern, not a fluke — and it’s happening against a market where OpenAI just locked in GPT-5.6 pricing and Anthropic is shipping weekly Claude Code releases. Standing still is its own kind of signal.
Who has access: A small group of approved enterprise customers, plus testers on Google’s Antigravity platform and LMArena.
I’m a Flutter engineer collecting AI × dev news daily using Claude Code. Follow for tomorrow’s digest.
