Claude coordinates parallel work for bigger builds
Bigger agent jobs are getting easier to delegate, while the rules around specialized access and model evaluation are getting more specific.
In today’s edition
- Claude Projects coordinates parallel cloud code sessions.
- Anthropic opens verified access for biology work.
- OpenAI launches Astra for Law.
- Epoch begins public reviews of AI benchmarks.
Claude Projects becomes a coordinator for parallel code work
Claude Code Projects now lets you hand over a larger goal in one conversation, then have Claude split it into parallel cloud sessions, review their outputs, and assemble the result.
- Each work thread is a Claude Code cloud session on its own branch and copy of the repo.
- Claude can scope work, run tests, open pull requests, and coordinate repo-level tasks such as migrating callers from a deprecated endpoint.
- Threads share project memory and a library of input files and generated artifacts, reducing repeated briefing across longer jobs.
- The beta is available to select Pro and Max subscribers using Claude Code cloud sessions. Threads cannot yet access local files, local tools, or internal networks.
- Parallel threads can hit usage limits sooner, and overlapping code changes still create ordinary merge conflicts.
What it means Use Projects for multi-repo migrations or parallel performance work when cloud-only access is enough, but plan for PR review, merge conflicts, and higher usage before delegating a large job.
Source: claudeai
Anthropic opens verified access for biology work
Anthropic’s Life Sciences Verification Program gives vetted organizations more permissive access to Mythos, Opus, and Sonnet for biology-related work that can be blocked in its generally available models.
- Teams can apply for a Standard Use grant covering most biology research and development workflows across Mythos, Opus, and Sonnet.
- A separately vetted High-risk Use grant applies to a specific research project and removes safeguards that block life sciences requests for Opus and Sonnet.
- Anthropic shifts from real-time blocking to offline monitoring for LSVP traffic and requires 30-day data retention to review flagged activity.
- The beta is available through Anthropic’s first-party API console and Claude Enterprise and Team plans, but not individual plans, third-party platforms, or BAA-enabled organizations.
What it means Life-science teams can pursue previously blocked model workflows through a verified route, but must weigh the project-specific access rules and 30-day monitoring requirement before moving sensitive work.
Source: AnthropicAI
OpenAI packages Astra into a legal AI platform
OpenAI has launched Astra for Law, pairing GPT-6 Astra with legal instructions, higher-effort settings, a U.S. legal search index, and plugins for law firms and legal technology companies.
- The Legal Search Index covers U.S. case law, statutes, regulations, court rules, and administrative decisions across more than 230 million URLs, with sources added daily.
- OpenAI reports that Astra for Law passed 54.0% of questions on a private Vals AI Legal Research Bench validation set, versus 38.7% for GPT-6 Astra with web search alone. This is OpenAI’s reported evaluation.
- The launch includes 26 partner-built plugins and 47 community plugins for legal work in ChatGPT.
- Initial access is limited to selected firms through Trusted Access in ChatGPT and Codex. API access is coming soon.
What it means Legal AI builders should watch this direct frontier-lab move into a valuable vertical, combining model access, retrieval, governance controls, and an ecosystem layer under one product.
Source: OpenAI
Epoch starts publicly grading AI benchmarks
Epoch AI has launched Benchmark Reviews, a public audit program that begins with 15 benchmark versions: four verified, nine flawed, and two lacking enough information for a verdict.
- The reviews apply to specific benchmark versions, not to every AI benchmark or every model result.
- Epoch’s SWE-bench Verified review labels the benchmark flawed, citing contamination risks from public codebases and published evidence of scoring problems.
- Its review says an OpenAI audit found that more than half of commonly unsolved questions had tests that rejected functionally correct submissions.
- Epoch says its reviews use a published rubric and invite benchmark developers to respond.
What it means Treat benchmark wins as a starting point, not a buying decision: check whether the evaluated version has contamination, scoring, or task-validity problems before using its score in product choices.
Source: EpochAIResearch